Mailing List Archive

Also technically, it's just the threat of a veto since we are not actually
in a vote thread....

On Sun, Apr 9, 2023 at 12:46?PM Gus Heck <gus.heck@gmail.com> wrote:

> What I see so far:
>
> 1. Much positive support for raising the limit
> 2. Slightly less support for removing it or making it configurable
> 3. A single veto which argues that a (as yet undefined) performance
> standard must be met before raising the limit
> 4. Hot tempers (various) making this discussion difficult
>
> As I understand it, vetoes must have technical merit. I'm not sure that
> this veto rises to "technical merit" on 2 counts:
>
> 1. No standard for the performance is given so it cannot be
> technically met. Without hard criteria it's a moving target.
> 2. It appears to encode a valuation of the user's time, and that
> valuation is really up to the user. Some users may consider 2hours useless
> and not worth it, and others might happily wait 2 hours. This is not a
> technical decision, it's a business decision regarding the relative value
> of the time invested vs the value of the result. If I can cure cancer by
> indexing for a year, that might be worth it... (hyperbole of course).
>
> Things I would consider to have technical merit that I don't hear:
>
> 1. Impact on the speed of **other** indexing operations. (devaluation
> of other functionality)
> 2. Actual scenarios that work when the limit is low and fail when the
> limit is high (new failure on the same data with the limit raised).
>
> One thing that might or might not have technical merit
>
> 1. If someone feels there is a lack of documentation of the
> costs/performance implications of using large vectors, possibly including
> reproducible benchmarks establishing the scaling behavior (there seems to
> be disagreement on O(n) vs O(n^2)).
>
> The users *should* know what they are getting into, but if the cost is
> worth it to them, they should be able to pay it without forking the
> project. If this veto causes a fork that's not good.
>
> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com> wrote:
>
>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100
>> and 300 dimensional varieties and can easily enough generate large numbers
>> of vector documents from the articles data. To go higher we could
>> concatenate vectors from that and I believe the performance numbers would
>> be plausible.
>>
>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>>
>>> Can we set up a branch in which the limit is bumped to 2048, then have
>>> a realistic, free data set (wikipedia sample or something) that has,
>>> say, 5 million docs and vectors created using public data (glove
>>> pre-trained embeddings or the like)? We then could run indexing on the
>>> same hardware with 512, 1024 and 2048 and see what the numbers, limits
>>> and behavior actually are.
>>>
>>> I can help in writing this but not until after Easter.
>>>
>>>
>>> Dawid
>>>
>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>> >
>>> > As Dawid pointed out earlier on this thread, this is the rule for
>>> > Apache projects: a single -1 vote on a code change is a veto and
>>> > cannot be overridden. Furthermore, Robert is one of the people on this
>>> > project who worked the most on debugging subtle bugs, making Lucene
>>> > more robust and improving our test framework, so I'm listening when he
>>> > voices quality concerns.
>>> >
>>> > The argument against removing/raising the limit that resonates with me
>>> > the most is that it is a one-way door. As MikeS highlighted earlier on
>>> > this thread, implementations may want to take advantage of the fact
>>> > that there is a limit at some point too. This is why I don't want to
>>> > remove the limit and would prefer a slight increase, such as 2048 as
>>> > suggested in the original issue, which would enable most of the things
>>> > that users who have been asking about raising the limit would like to
>>> > do.
>>> >
>>> > I agree that the merge-time memory usage and slow indexing rate are
>>> > not great. But it's still possible to index multi-million vector
>>> > datasets with a 4GB heap without hitting OOMEs regardless of the
>>> > number of dimensions, and the feedback I'm seeing is that many users
>>> > are still interested in indexing multi-million vector datasets despite
>>> > the slow indexing rate. I wish we could do better, and vector indexing
>>> > is certainly more expert than text indexing, but it still is usable in
>>> > my opinion. I understand how giving Lucene more information about
>>> > vectors prior to indexing (e.g. clustering information as Jim pointed
>>> > out) could help make merging faster and more memory-efficient, but I
>>> > would really like to avoid making it a requirement for indexing
>>> > vectors as it also makes this feature much harder to use.
>>> >
>>> > On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>> > <a.benedetti@sease.io> wrote:
>>> > >
>>> > > I am very attentive to listen opinions but I am un-convinced here
>>> and I an not sure that a single person opinion should be allowed to be
>>> detrimental for such an important project.
>>> > >
>>> > > The limit as far as I know is literally just raising an exception.
>>> > > Removing it won't alter in any way the current performance for users
>>> in low dimensional space.
>>> > > Removing it will just enable more users to use Lucene.
>>> > >
>>> > > If new users in certain situations will be unhappy with the
>>> performance, they may contribute improvements.
>>> > > This is how you make progress.
>>> > >
>>> > > If it's a reputation thing, trust me that not allowing users to play
>>> with high dimensional space will equally damage it.
>>> > >
>>> > > To me it's really a no brainer.
>>> > > Removing the limit and enable people to use high dimensional vectors
>>> will take minutes.
>>> > > Improving the hnsw implementation can take months.
>>> > > Pick one to begin with...
>>> > >
>>> > > And there's no-one paying me here, no company interest whatsoever,
>>> actually I pay people to contribute, I am just convinced it's a good idea.
>>> > >
>>> > >
>>> > > On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>> > >>
>>> > >> I disagree with your categorization. I put in plenty of work and
>>> > >> experienced plenty of pain myself, writing tests and fighting these
>>> > >> issues, after i saw that, two releases in a row, vector indexing
>>> fell
>>> > >> over and hit integer overflows etc on small datasets:
>>> > >>
>>> > >> https://github.com/apache/lucene/pull/11905
>>> > >>
>>> > >> Attacking me isn't helping the situation.
>>> > >>
>>> > >> PS: when i said the "one guy who wrote the code" I didn't mean it in
>>> > >> any kind of demeaning fashion really. I meant to describe the
>>> current
>>> > >> state of usability with respect to indexing a few million docs with
>>> > >> high dimensions. You can scroll up the thread and see that at least
>>> > >> one other committer on the project experienced similar pain as me.
>>> > >> Then, think about users who aren't committers trying to use the
>>> > >> functionality!
>>> > >>
>>> > >> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <msokolov@gmail.com>
>>> wrote:
>>> > >> >
>>> > >> > What you said about increasing dimensions requiring a bigger ram
>>> buffer on merge is wrong. That's the point I was trying to make. Your
>>> concerns about merge costs are not wrong, but your conclusion that we need
>>> to limit dimensions is not justified.
>>> > >> >
>>> > >> > You complain that hnsw sucks it doesn't scale, but when I show it
>>> scales linearly with dimension you just ignore that and complain about
>>> something entirely different.
>>> > >> >
>>> > >> > You demand that people run all kinds of tests to prove you wrong
>>> but when they do, you don't listen and you won't put in the work yourself
>>> or complain that it's too hard.
>>> > >> >
>>> > >> > Then you complain about people not meeting you half way. Wow
>>> > >> >
>>> > >> > On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
>>> wrote:
>>> > >> >>
>>> > >> >> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>> > >> >> <michael.wechner@wyona.com> wrote:
>>> > >> >> >
>>> > >> >> > What exactly do you consider reasonable?
>>> > >> >>
>>> > >> >> Let's begin a real discussion by being HONEST about the current
>>> > >> >> status. Please put politically correct or your own company's
>>> wishes
>>> > >> >> aside, we know it's not in a good state.
>>> > >> >>
>>> > >> >> Current status is the one guy who wrote the code can set a
>>> > >> >> multi-gigabyte ram buffer and index a small dataset with 1024
>>> > >> >> dimensions in HOURS (i didn't ask what hardware).
>>> > >> >>
>>> > >> >> My concerns are everyone else except the one guy, I want it to be
>>> > >> >> usable. Increasing dimensions just means even bigger
>>> multi-gigabyte
>>> > >> >> ram buffer and bigger heap to avoid OOM on merge.
>>> > >> >> It is also a permanent backwards compatibility decision, we have
>>> to
>>> > >> >> support it once we do this and we can't just say "oops" and flip
>>> it
>>> > >> >> back.
>>> > >> >>
>>> > >> >> It is unclear to me, if the multi-gigabyte ram buffer is really
>>> to
>>> > >> >> avoid merges because they are so slow and it would be DAYS
>>> otherwise,
>>> > >> >> or if its to avoid merges so it doesn't hit OOM.
>>> > >> >> Also from personal experience, it takes trial and error (means
>>> > >> >> experiencing OOM on merge!!!) before you get those heap values
>>> correct
>>> > >> >> for your dataset. This usually means starting over which is
>>> > >> >> frustrating and wastes more time.
>>> > >> >>
>>> > >> >> Jim mentioned some ideas about the memory usage in IndexWriter,
>>> seems
>>> > >> >> to me like its a good idea. maybe the multigigabyte ram buffer
>>> can be
>>> > >> >> avoided in this way and performance improved by writing bigger
>>> > >> >> segments with lucene's defaults. But this doesn't mean we can
>>> simply
>>> > >> >> ignore the horrors of what happens on merge. merging needs to
>>> scale so
>>> > >> >> that indexing really scales.
>>> > >> >>
>>> > >> >> At least it shouldnt spike RAM on trivial data amounts and cause
>>> OOM,
>>> > >> >> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>> > >> >> fashion when indexing.
>>> > >> >>
>>> > >> >>
>>> ---------------------------------------------------------------------
>>> > >> >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > >> >> For additional commands, e-mail: dev-help@lucene.apache.org
>>> > >> >>
>>> > >>
>>> > >>
>>> ---------------------------------------------------------------------
>>> > >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > >> For additional commands, e-mail: dev-help@lucene.apache.org
>>> > >>
>>> >
>>> >
>>> > --
>>> > Adrien
>>> >
>>> > ---------------------------------------------------------------------
>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>> >
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>
>>>
>
> --
> http://www.needhamsoftware.com (work)
> http://www.the111shift.com (play)
>

--
http://www.needhamsoftware.com (work)
http://www.the111shift.com (play)

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

vaiju1981 at gmail

Apr 9, 2023, 10:19 AM

Post #78 of 99 (942 views)

Joining this pretty late, But we do have a use case to use 2K dimensions
for embedding purposes.
We decided to break embeddings into high and low where first 1K goes into
high field and next 1K goes into low, while indexing this worked well. We
also add a few other fields like segment/cluster id to help us narrow
searches in subsequent.

While searching we then do a multiple pass to use both of these along with
other fields ( yes this might not scale for others ). It is working so far.

--Thanks and Regards
Vaijanath

On Sun, Apr 9, 2023 at 10:02?AM Gus Heck <gus.heck@gmail.com> wrote:

> Also technically, it's just the threat of a veto since we are not actually
> in a vote thread....
>
> On Sun, Apr 9, 2023 at 12:46?PM Gus Heck <gus.heck@gmail.com> wrote:
>
>> What I see so far:
>>
>> 1. Much positive support for raising the limit
>> 2. Slightly less support for removing it or making it configurable
>> 3. A single veto which argues that a (as yet undefined) performance
>> standard must be met before raising the limit
>> 4. Hot tempers (various) making this discussion difficult
>>
>> As I understand it, vetoes must have technical merit. I'm not sure that
>> this veto rises to "technical merit" on 2 counts:
>>
>> 1. No standard for the performance is given so it cannot be
>> technically met. Without hard criteria it's a moving target.
>> 2. It appears to encode a valuation of the user's time, and that
>> valuation is really up to the user. Some users may consider 2hours useless
>> and not worth it, and others might happily wait 2 hours. This is not a
>> technical decision, it's a business decision regarding the relative value
>> of the time invested vs the value of the result. If I can cure cancer by
>> indexing for a year, that might be worth it... (hyperbole of course).
>>
>> Things I would consider to have technical merit that I don't hear:
>>
>> 1. Impact on the speed of **other** indexing operations. (devaluation
>> of other functionality)
>> 2. Actual scenarios that work when the limit is low and fail when the
>> limit is high (new failure on the same data with the limit raised).
>>
>> One thing that might or might not have technical merit
>>
>> 1. If someone feels there is a lack of documentation of the
>> costs/performance implications of using large vectors, possibly including
>> reproducible benchmarks establishing the scaling behavior (there seems to
>> be disagreement on O(n) vs O(n^2)).
>>
>> The users *should* know what they are getting into, but if the cost is
>> worth it to them, they should be able to pay it without forking the
>> project. If this veto causes a fork that's not good.
>>
>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com>
>> wrote:
>>
>>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100
>>> and 300 dimensional varieties and can easily enough generate large numbers
>>> of vector documents from the articles data. To go higher we could
>>> concatenate vectors from that and I believe the performance numbers would
>>> be plausible.
>>>
>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>>>
>>>> Can we set up a branch in which the limit is bumped to 2048, then have
>>>> a realistic, free data set (wikipedia sample or something) that has,
>>>> say, 5 million docs and vectors created using public data (glove
>>>> pre-trained embeddings or the like)? We then could run indexing on the
>>>> same hardware with 512, 1024 and 2048 and see what the numbers, limits
>>>> and behavior actually are.
>>>>
>>>> I can help in writing this but not until after Easter.
>>>>
>>>>
>>>> Dawid
>>>>
>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>>> >
>>>> > As Dawid pointed out earlier on this thread, this is the rule for
>>>> > Apache projects: a single -1 vote on a code change is a veto and
>>>> > cannot be overridden. Furthermore, Robert is one of the people on this
>>>> > project who worked the most on debugging subtle bugs, making Lucene
>>>> > more robust and improving our test framework, so I'm listening when he
>>>> > voices quality concerns.
>>>> >
>>>> > The argument against removing/raising the limit that resonates with me
>>>> > the most is that it is a one-way door. As MikeS highlighted earlier on
>>>> > this thread, implementations may want to take advantage of the fact
>>>> > that there is a limit at some point too. This is why I don't want to
>>>> > remove the limit and would prefer a slight increase, such as 2048 as
>>>> > suggested in the original issue, which would enable most of the things
>>>> > that users who have been asking about raising the limit would like to
>>>> > do.
>>>> >
>>>> > I agree that the merge-time memory usage and slow indexing rate are
>>>> > not great. But it's still possible to index multi-million vector
>>>> > datasets with a 4GB heap without hitting OOMEs regardless of the
>>>> > number of dimensions, and the feedback I'm seeing is that many users
>>>> > are still interested in indexing multi-million vector datasets despite
>>>> > the slow indexing rate. I wish we could do better, and vector indexing
>>>> > is certainly more expert than text indexing, but it still is usable in
>>>> > my opinion. I understand how giving Lucene more information about
>>>> > vectors prior to indexing (e.g. clustering information as Jim pointed
>>>> > out) could help make merging faster and more memory-efficient, but I
>>>> > would really like to avoid making it a requirement for indexing
>>>> > vectors as it also makes this feature much harder to use.
>>>> >
>>>> > On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>>> > <a.benedetti@sease.io> wrote:
>>>> > >
>>>> > > I am very attentive to listen opinions but I am un-convinced here
>>>> and I an not sure that a single person opinion should be allowed to be
>>>> detrimental for such an important project.
>>>> > >
>>>> > > The limit as far as I know is literally just raising an exception.
>>>> > > Removing it won't alter in any way the current performance for
>>>> users in low dimensional space.
>>>> > > Removing it will just enable more users to use Lucene.
>>>> > >
>>>> > > If new users in certain situations will be unhappy with the
>>>> performance, they may contribute improvements.
>>>> > > This is how you make progress.
>>>> > >
>>>> > > If it's a reputation thing, trust me that not allowing users to
>>>> play with high dimensional space will equally damage it.
>>>> > >
>>>> > > To me it's really a no brainer.
>>>> > > Removing the limit and enable people to use high dimensional
>>>> vectors will take minutes.
>>>> > > Improving the hnsw implementation can take months.
>>>> > > Pick one to begin with...
>>>> > >
>>>> > > And there's no-one paying me here, no company interest whatsoever,
>>>> actually I pay people to contribute, I am just convinced it's a good idea.
>>>> > >
>>>> > >
>>>> > > On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>>> > >>
>>>> > >> I disagree with your categorization. I put in plenty of work and
>>>> > >> experienced plenty of pain myself, writing tests and fighting these
>>>> > >> issues, after i saw that, two releases in a row, vector indexing
>>>> fell
>>>> > >> over and hit integer overflows etc on small datasets:
>>>> > >>
>>>> > >> https://github.com/apache/lucene/pull/11905
>>>> > >>
>>>> > >> Attacking me isn't helping the situation.
>>>> > >>
>>>> > >> PS: when i said the "one guy who wrote the code" I didn't mean it
>>>> in
>>>> > >> any kind of demeaning fashion really. I meant to describe the
>>>> current
>>>> > >> state of usability with respect to indexing a few million docs with
>>>> > >> high dimensions. You can scroll up the thread and see that at least
>>>> > >> one other committer on the project experienced similar pain as me.
>>>> > >> Then, think about users who aren't committers trying to use the
>>>> > >> functionality!
>>>> > >>
>>>> > >> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <
>>>> msokolov@gmail.com> wrote:
>>>> > >> >
>>>> > >> > What you said about increasing dimensions requiring a bigger ram
>>>> buffer on merge is wrong. That's the point I was trying to make. Your
>>>> concerns about merge costs are not wrong, but your conclusion that we need
>>>> to limit dimensions is not justified.
>>>> > >> >
>>>> > >> > You complain that hnsw sucks it doesn't scale, but when I show
>>>> it scales linearly with dimension you just ignore that and complain about
>>>> something entirely different.
>>>> > >> >
>>>> > >> > You demand that people run all kinds of tests to prove you wrong
>>>> but when they do, you don't listen and you won't put in the work yourself
>>>> or complain that it's too hard.
>>>> > >> >
>>>> > >> > Then you complain about people not meeting you half way. Wow
>>>> > >> >
>>>> > >> > On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
>>>> wrote:
>>>> > >> >>
>>>> > >> >> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>>> > >> >> <michael.wechner@wyona.com> wrote:
>>>> > >> >> >
>>>> > >> >> > What exactly do you consider reasonable?
>>>> > >> >>
>>>> > >> >> Let's begin a real discussion by being HONEST about the current
>>>> > >> >> status. Please put politically correct or your own company's
>>>> wishes
>>>> > >> >> aside, we know it's not in a good state.
>>>> > >> >>
>>>> > >> >> Current status is the one guy who wrote the code can set a
>>>> > >> >> multi-gigabyte ram buffer and index a small dataset with 1024
>>>> > >> >> dimensions in HOURS (i didn't ask what hardware).
>>>> > >> >>
>>>> > >> >> My concerns are everyone else except the one guy, I want it to
>>>> be
>>>> > >> >> usable. Increasing dimensions just means even bigger
>>>> multi-gigabyte
>>>> > >> >> ram buffer and bigger heap to avoid OOM on merge.
>>>> > >> >> It is also a permanent backwards compatibility decision, we
>>>> have to
>>>> > >> >> support it once we do this and we can't just say "oops" and
>>>> flip it
>>>> > >> >> back.
>>>> > >> >>
>>>> > >> >> It is unclear to me, if the multi-gigabyte ram buffer is really
>>>> to
>>>> > >> >> avoid merges because they are so slow and it would be DAYS
>>>> otherwise,
>>>> > >> >> or if its to avoid merges so it doesn't hit OOM.
>>>> > >> >> Also from personal experience, it takes trial and error (means
>>>> > >> >> experiencing OOM on merge!!!) before you get those heap values
>>>> correct
>>>> > >> >> for your dataset. This usually means starting over which is
>>>> > >> >> frustrating and wastes more time.
>>>> > >> >>
>>>> > >> >> Jim mentioned some ideas about the memory usage in IndexWriter,
>>>> seems
>>>> > >> >> to me like its a good idea. maybe the multigigabyte ram buffer
>>>> can be
>>>> > >> >> avoided in this way and performance improved by writing bigger
>>>> > >> >> segments with lucene's defaults. But this doesn't mean we can
>>>> simply
>>>> > >> >> ignore the horrors of what happens on merge. merging needs to
>>>> scale so
>>>> > >> >> that indexing really scales.
>>>> > >> >>
>>>> > >> >> At least it shouldnt spike RAM on trivial data amounts and
>>>> cause OOM,
>>>> > >> >> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>>> > >> >> fashion when indexing.
>>>> > >> >>
>>>> > >> >>
>>>> ---------------------------------------------------------------------
>>>> > >> >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> > >> >> For additional commands, e-mail: dev-help@lucene.apache.org
>>>> > >> >>
>>>> > >>
>>>> > >>
>>>> ---------------------------------------------------------------------
>>>> > >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> > >> For additional commands, e-mail: dev-help@lucene.apache.org
>>>> > >>
>>>> >
>>>> >
>>>> > --
>>>> > Adrien
>>>> >
>>>> > ---------------------------------------------------------------------
>>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>>> >
>>>>
>>>> ---------------------------------------------------------------------
>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>
>>>>
>>
>> --
>> http://www.needhamsoftware.com (work)
>> http://www.the111shift.com (play)
>>
>
>
> --
> http://www.needhamsoftware.com (work)
> http://www.the111shift.com (play)
>

--
I am feeling fine, healthier and Happier, what about you

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

dawid.weiss at gmail

Apr 9, 2023, 11:43 AM

Post #79 of 99 (942 views)

> We do have a dataset built from Wikipedia in luceneutil. It comes in 100 and 300 dimensional varieties and can easily enough generate large numbers of vector documents from the articles data. To go higher we could concatenate vectors from that and I believe the performance numbers would be plausible.

Apologies - I wasn't clear - I thought of building the 1k or 2k
vectors that would be realistic. Perhaps using glove or perhaps using
some other software but something that would reflect a true 2k
dimensional space accurately with "real" data underneath. I am not
familiar enough with the field to tell whether a simple concatenation
is a good enough simulation - perhaps it is.

I would really prefer to focus on doing this kind of assessment of
feasibility/ limitations rather than arguing back and forth. I did my
experiment a while ago and I can't really tell whether there have been
improvements in the indexing/ merging part - your email contradicts my
experience Mike, so I'm a bit intrigued and would like to revisit it.
But it'd be ideal to work with real vectors rather than a simulation.

Dawid

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 9, 2023, 12:14 PM

Post #80 of 99 (942 views)

Our use case is the following:

We have a dataset of several thousand questions and answers, for which
we generate vectors with various models resp. services, for example:

SentenceBERT, all-mpnet-base-v2
Aleph Alpha, luminous-base
Cohere, multilingual-22-12
OpenAI, text-similarity-ada-001 or text-similarity-davinci-001

These models produce vectors with different vector dimensions.
Depending on the model and on the benchmarks / datasets the accuracy is
higher when the vector dimension is higher.

We index these vectors with Lucene and we use Lucene to do a similarity
search.

I think this is exactly why Lucene KNN / ANN was developed, to do
similarity search, right?

Thanks

Michael W

Am 09.04.23 um 13:47 schrieb Robert Muir:
> I don't care. you guys personally attacked me first. And then it turns
> out, you were being dishonest the entire time and hiding your true
> intent, which was not search at all but instead some chatgpt pyramid
> scheme or similar.
>
> i'm done with this thread.
>
> On Sun, Apr 9, 2023 at 7:37?AM Alessandro Benedetti
> <a.benedetti@sease.io> wrote:
>> I don't think this tone and language is appropriate for a community of volunteers and men of science.
>>
>> I personally find offensive to generalise Lucene people here to be "crazy people hyped about chatGPT".
>>
>> I personally don't give a damn about chatGPT except the fact it is a very interesting technology.
>>
>> As usual I see very little motivation and a lot of "convince me".
>> We're discussing here about a limit that raises an exception.
>>
>> Improving performance is absolutely important and no-one here is saying we won't address it, it's just a separate discussion.
>>
>>
>> On Sun, 9 Apr 2023, 12:59 Robert Muir, <rcmuir@gmail.com> wrote:
>>> Also, please let's only disucss SEARCH. lucene is a SEARCH ENGINE
>>> LIBRARY. not a vector database or whatever trash is being proposed
>>> here.
>>>
>>> i think we should table this and revisit it after chatgpt hype has dissipated.
>>>
>>> this hype is causing ppl to behave irrationally, it is why i can't
>>> converse with basically anyone on this thread because they are all
>>> stating crazy things that don't make sense.
>>>
>>> On Sun, Apr 9, 2023 at 6:25?AM Robert Muir <rcmuir@gmail.com> wrote:
>>>> Yes, its very clear that folks on this thread are ignoring reason
>>>> entirely and completely swooned by chatgpt-hype.
>>>> And what happens when they make chatgpt-8 that uses even more dimensions?
>>>> backwards compatibility decisions can't be made by garbage hype such
>>>> as cryptocurrency or chatgpt.
>>>> Trying to convince me we should bump it because of chatgpt, well, i
>>>> think it has the opposite effect.
>>>>
>>>> Please, lemme see real technical arguments why this limit needs to be
>>>> bumped. not including trash like chatgpt.
>>>>
>>>> On Sat, Apr 8, 2023 at 7:50?PM Marcus Eagan <marcuseagan@gmail.com> wrote:
>>>>> Given the massive amounts of funding going into the development and investigation of the project, I think it would be good to at least have Lucene be a part of the conversation. Simply because academics typically focus on vectors <= 784 dimensions does not mean all users will. A large swathe of very important users of the Lucene project never exceed 500k documents, though they are shifting to other search engines to try out very popular embeddings.
>>>>>
>>>>> I think giving our users the opportunity to build chat bots or LLM memory machines using Lucene is a positive development, even if some datasets won't be able to work well. We don't limit the number of fields someone can add in most cases, though we did just undeprecate that API to better support multi-tenancy. But people still add so many fields and can crash their clusters with mapping explosions when unlimited. The limit to vectors feels similar. I expect more people to dig into Lucene due to its openness and robustness as they run into problems. Today, they are forced to consider other engines that are more permissive.
>>>>>
>>>>> Not everyone important or valuable Lucene workload is in the millions of documents. Many of them only have lots of queries or computationally expensive access patterns for B-trees. We can document that it is very ill-advised to make a deployment with vectors too large. What others will do with it is on them.
>>>>>
>>>>>
>>>>> On Sat, Apr 8, 2023 at 2:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>>>>> As Dawid pointed out earlier on this thread, this is the rule for
>>>>>> Apache projects: a single -1 vote on a code change is a veto and
>>>>>> cannot be overridden. Furthermore, Robert is one of the people on this
>>>>>> project who worked the most on debugging subtle bugs, making Lucene
>>>>>> more robust and improving our test framework, so I'm listening when he
>>>>>> voices quality concerns.
>>>>>>
>>>>>> The argument against removing/raising the limit that resonates with me
>>>>>> the most is that it is a one-way door. As MikeS highlighted earlier on
>>>>>> this thread, implementations may want to take advantage of the fact
>>>>>> that there is a limit at some point too. This is why I don't want to
>>>>>> remove the limit and would prefer a slight increase, such as 2048 as
>>>>>> suggested in the original issue, which would enable most of the things
>>>>>> that users who have been asking about raising the limit would like to
>>>>>> do.
>>>>>>
>>>>>> I agree that the merge-time memory usage and slow indexing rate are
>>>>>> not great. But it's still possible to index multi-million vector
>>>>>> datasets with a 4GB heap without hitting OOMEs regardless of the
>>>>>> number of dimensions, and the feedback I'm seeing is that many users
>>>>>> are still interested in indexing multi-million vector datasets despite
>>>>>> the slow indexing rate. I wish we could do better, and vector indexing
>>>>>> is certainly more expert than text indexing, but it still is usable in
>>>>>> my opinion. I understand how giving Lucene more information about
>>>>>> vectors prior to indexing (e.g. clustering information as Jim pointed
>>>>>> out) could help make merging faster and more memory-efficient, but I
>>>>>> would really like to avoid making it a requirement for indexing
>>>>>> vectors as it also makes this feature much harder to use.
>>>>>>
>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>>>>> <a.benedetti@sease.io> wrote:
>>>>>>> I am very attentive to listen opinions but I am un-convinced here and I an not sure that a single person opinion should be allowed to be detrimental for such an important project.
>>>>>>>
>>>>>>> The limit as far as I know is literally just raising an exception.
>>>>>>> Removing it won't alter in any way the current performance for users in low dimensional space.
>>>>>>> Removing it will just enable more users to use Lucene.
>>>>>>>
>>>>>>> If new users in certain situations will be unhappy with the performance, they may contribute improvements.
>>>>>>> This is how you make progress.
>>>>>>>
>>>>>>> If it's a reputation thing, trust me that not allowing users to play with high dimensional space will equally damage it.
>>>>>>>
>>>>>>> To me it's really a no brainer.
>>>>>>> Removing the limit and enable people to use high dimensional vectors will take minutes.
>>>>>>> Improving the hnsw implementation can take months.
>>>>>>> Pick one to begin with...
>>>>>>>
>>>>>>> And there's no-one paying me here, no company interest whatsoever, actually I pay people to contribute, I am just convinced it's a good idea.
>>>>>>>
>>>>>>>
>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>>>>>>> I disagree with your categorization. I put in plenty of work and
>>>>>>>> experienced plenty of pain myself, writing tests and fighting these
>>>>>>>> issues, after i saw that, two releases in a row, vector indexing fell
>>>>>>>> over and hit integer overflows etc on small datasets:
>>>>>>>>
>>>>>>>> https://github.com/apache/lucene/pull/11905
>>>>>>>>
>>>>>>>> Attacking me isn't helping the situation.
>>>>>>>>
>>>>>>>> PS: when i said the "one guy who wrote the code" I didn't mean it in
>>>>>>>> any kind of demeaning fashion really. I meant to describe the current
>>>>>>>> state of usability with respect to indexing a few million docs with
>>>>>>>> high dimensions. You can scroll up the thread and see that at least
>>>>>>>> one other committer on the project experienced similar pain as me.
>>>>>>>> Then, think about users who aren't committers trying to use the
>>>>>>>> functionality!
>>>>>>>>
>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <msokolov@gmail.com> wrote:
>>>>>>>>> What you said about increasing dimensions requiring a bigger ram buffer on merge is wrong. That's the point I was trying to make. Your concerns about merge costs are not wrong, but your conclusion that we need to limit dimensions is not justified.
>>>>>>>>>
>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when I show it scales linearly with dimension you just ignore that and complain about something entirely different.
>>>>>>>>>
>>>>>>>>> You demand that people run all kinds of tests to prove you wrong but when they do, you don't listen and you won't put in the work yourself or complain that it's too hard.
>>>>>>>>>
>>>>>>>>> Then you complain about people not meeting you half way. Wow
>>>>>>>>>
>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com> wrote:
>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>>>>>>>>>>> What exactly do you consider reasonable?
>>>>>>>>>> Let's begin a real discussion by being HONEST about the current
>>>>>>>>>> status. Please put politically correct or your own company's wishes
>>>>>>>>>> aside, we know it's not in a good state.
>>>>>>>>>>
>>>>>>>>>> Current status is the one guy who wrote the code can set a
>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset with 1024
>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>>>>>>>>>>
>>>>>>>>>> My concerns are everyone else except the one guy, I want it to be
>>>>>>>>>> usable. Increasing dimensions just means even bigger multi-gigabyte
>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>>>>>>>>>> It is also a permanent backwards compatibility decision, we have to
>>>>>>>>>> support it once we do this and we can't just say "oops" and flip it
>>>>>>>>>> back.
>>>>>>>>>>
>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer is really to
>>>>>>>>>> avoid merges because they are so slow and it would be DAYS otherwise,
>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>>>>>>>>>> Also from personal experience, it takes trial and error (means
>>>>>>>>>> experiencing OOM on merge!!!) before you get those heap values correct
>>>>>>>>>> for your dataset. This usually means starting over which is
>>>>>>>>>> frustrating and wastes more time.
>>>>>>>>>>
>>>>>>>>>> Jim mentioned some ideas about the memory usage in IndexWriter, seems
>>>>>>>>>> to me like its a good idea. maybe the multigigabyte ram buffer can be
>>>>>>>>>> avoided in this way and performance improved by writing bigger
>>>>>>>>>> segments with lucene's defaults. But this doesn't mean we can simply
>>>>>>>>>> ignore the horrors of what happens on merge. merging needs to scale so
>>>>>>>>>> that indexing really scales.
>>>>>>>>>>
>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts and cause OOM,
>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>>>>>>>>> fashion when indexing.
>>>>>>>>>>
>>>>>>>>>> ---------------------------------------------------------------------
>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>>>>>>
>>>>>>>> ---------------------------------------------------------------------
>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>>>>
>>>>>>
>>>>>> --
>>>>>> Adrien
>>>>>>
>>>>>> ---------------------------------------------------------------------
>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>>
>>>>>
>>>>> --
>>>>> Marcus Eagan
>>>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 9, 2023, 12:23 PM

Post #81 of 99 (942 views)

I think for testing the performance and scalability one can also use
synthetic data and it does not have to be real world data in the sense
of vectors generated from real world text.

But I think the more people revisit the testing of performance and
scalability the better and any help on this would be great!

Thanks

Michael W

Am 09.04.23 um 20:43 schrieb Dawid Weiss:
>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100 and 300 dimensional varieties and can easily enough generate large numbers of vector documents from the articles data. To go higher we could concatenate vectors from that and I believe the performance numbers would be plausible.
> Apologies - I wasn't clear - I thought of building the 1k or 2k
> vectors that would be realistic. Perhaps using glove or perhaps using
> some other software but something that would reflect a true 2k
> dimensional space accurately with "real" data underneath. I am not
> familiar enough with the field to tell whether a simple concatenation
> is a good enough simulation - perhaps it is.
>
> I would really prefer to focus on doing this kind of assessment of
> feasibility/ limitations rather than arguing back and forth. I did my
> experiment a while ago and I can't really tell whether there have been
> improvements in the indexing/ merging part - your email contradicts my
> experience Mike, so I'm a bit intrigued and would like to revisit it.
> But it'd be ideal to work with real vectors rather than a simulation.
>
> Dawid
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 10, 2023, 6:54 AM

Post #82 of 99 (942 views)

I think concatenating word-embedding vectors is a reasonable thing to
do. It captures information about the sequence of tokens which is
being lost by the current approach (summing them). Random article I
found in a search
https://medium.com/@dhartidhami/understanding-bert-word-embeddings-7dc4d2ea54ca
shows higher performance with a concatenative approach. So it seems to
me we could take the 300-dim Glove vectors and produce somewhat
meaningful (say) 1200- or 1500-dim vectors by running a sliding window
over the tokens in a document and concatenating the token-vectors

On Sun, Apr 9, 2023 at 2:44?PM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>
> > We do have a dataset built from Wikipedia in luceneutil. It comes in 100 and 300 dimensional varieties and can easily enough generate large numbers of vector documents from the articles data. To go higher we could concatenate vectors from that and I believe the performance numbers would be plausible.
>
> Apologies - I wasn't clear - I thought of building the 1k or 2k
> vectors that would be realistic. Perhaps using glove or perhaps using
> some other software but something that would reflect a true 2k
> dimensional space accurately with "real" data underneath. I am not
> familiar enough with the field to tell whether a simple concatenation
> is a good enough simulation - perhaps it is.
>
> I would really prefer to focus on doing this kind of assessment of
> feasibility/ limitations rather than arguing back and forth. I did my
> experiment a while ago and I can't really tell whether there have been
> improvements in the indexing/ merging part - your email contradicts my
> experience Mike, so I'm a bit intrigued and would like to revisit it.
> But it'd be ideal to work with real vectors rather than a simulation.
>
> Dawid
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 10, 2023, 8:06 AM

Post #83 of 99 (942 views)

I poked around on huggingface looking at various models that are being
promoted there; this is the highest-performing text model they list,
which is expected to take sentences as input; it uses so-called
"attention" to capture the context of words:
https://huggingface.co/sentence-transformers/all-mpnet-base-v2 and it
is 768-dimensional. This is a list of models designed for "asymmetric
semantic search" ie short queries and long documents:
https://www.sbert.net/docs/pretrained-models/msmarco-v3.html. The
highest ranking one there also seems to be 768d
https://huggingface.co/sebastian-hofstaetter/distilbert-dot-tas_b-b256-msmarco

I did see some other larger-dimensional model, but they all seem to
involve images+text.

On Mon, Apr 10, 2023 at 9:54?AM Michael Sokolov <msokolov@gmail.com> wrote:
>
> I think concatenating word-embedding vectors is a reasonable thing to
> do. It captures information about the sequence of tokens which is
> being lost by the current approach (summing them). Random article I
> found in a search
> https://medium.com/@dhartidhami/understanding-bert-word-embeddings-7dc4d2ea54ca
> shows higher performance with a concatenative approach. So it seems to
> me we could take the 300-dim Glove vectors and produce somewhat
> meaningful (say) 1200- or 1500-dim vectors by running a sliding window
> over the tokens in a document and concatenating the token-vectors
>
> On Sun, Apr 9, 2023 at 2:44?PM Dawid Weiss <dawid.weiss@gmail.com> wrote:
> >
> > > We do have a dataset built from Wikipedia in luceneutil. It comes in 100 and 300 dimensional varieties and can easily enough generate large numbers of vector documents from the articles data. To go higher we could concatenate vectors from that and I believe the performance numbers would be plausible.
> >
> > Apologies - I wasn't clear - I thought of building the 1k or 2k
> > vectors that would be realistic. Perhaps using glove or perhaps using
> > some other software but something that would reflect a true 2k
> > dimensional space accurately with "real" data underneath. I am not
> > familiar enough with the field to tell whether a simple concatenation
> > is a good enough simulation - perhaps it is.
> >
> > I would really prefer to focus on doing this kind of assessment of
> > feasibility/ limitations rather than arguing back and forth. I did my
> > experiment a while ago and I can't really tell whether there have been
> > improvements in the indexing/ merging part - your email contradicts my
> > experience Mike, so I'm a bit intrigued and would like to revisit it.
> > But it'd be ideal to work with real vectors rather than a simulation.
> >
> > Dawid
> >
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> > For additional commands, e-mail: dev-help@lucene.apache.org
> >

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

a.benedetti at sease

Apr 10, 2023, 3:41 PM

Post #84 of 99 (942 views)

I think Gus points are on target.

I recommend we move this forward in this way:
We stop any discussion and everyone interested proposes an option with a
motivation, then we aggregate the options and we create a Vote maybe?

I am also on the same page on the fact that a veto should come with a clear
and reasonable technical merit, which also in my opinion has not come yet.

I also apologise if any of my words sounded harsh or personal attacks,
never meant to do so.

My proposed option:

1) remove the limit and potentially make it configurable,
Motivation:
The system administrator can enforce a limit its users need to respect that
it's in line with whatever the admin decided to be acceptable for them.
Default can stay the current one.

That's my favourite at the moment, but I agree that potentially in the
future this may need to change, as we may optimise the data structures for
certain dimensions. I am a big fan of Yagni (you aren't going to need it)
so I am ok we'll face a different discussion if that happens in the future.

On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:

> What I see so far:
>
> 1. Much positive support for raising the limit
> 2. Slightly less support for removing it or making it configurable
> 3. A single veto which argues that a (as yet undefined) performance
> standard must be met before raising the limit
> 4. Hot tempers (various) making this discussion difficult
>
> As I understand it, vetoes must have technical merit. I'm not sure that
> this veto rises to "technical merit" on 2 counts:
>
> 1. No standard for the performance is given so it cannot be
> technically met. Without hard criteria it's a moving target.
> 2. It appears to encode a valuation of the user's time, and that
> valuation is really up to the user. Some users may consider 2hours useless
> and not worth it, and others might happily wait 2 hours. This is not a
> technical decision, it's a business decision regarding the relative value
> of the time invested vs the value of the result. If I can cure cancer by
> indexing for a year, that might be worth it... (hyperbole of course).
>
> Things I would consider to have technical merit that I don't hear:
>
> 1. Impact on the speed of **other** indexing operations. (devaluation
> of other functionality)
> 2. Actual scenarios that work when the limit is low and fail when the
> limit is high (new failure on the same data with the limit raised).
>
> One thing that might or might not have technical merit
>
> 1. If someone feels there is a lack of documentation of the
> costs/performance implications of using large vectors, possibly including
> reproducible benchmarks establishing the scaling behavior (there seems to
> be disagreement on O(n) vs O(n^2)).
>
> The users *should* know what they are getting into, but if the cost is
> worth it to them, they should be able to pay it without forking the
> project. If this veto causes a fork that's not good.
>
> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com> wrote:
>
>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100
>> and 300 dimensional varieties and can easily enough generate large numbers
>> of vector documents from the articles data. To go higher we could
>> concatenate vectors from that and I believe the performance numbers would
>> be plausible.
>>
>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>>
>>> Can we set up a branch in which the limit is bumped to 2048, then have
>>> a realistic, free data set (wikipedia sample or something) that has,
>>> say, 5 million docs and vectors created using public data (glove
>>> pre-trained embeddings or the like)? We then could run indexing on the
>>> same hardware with 512, 1024 and 2048 and see what the numbers, limits
>>> and behavior actually are.
>>>
>>> I can help in writing this but not until after Easter.
>>>
>>>
>>> Dawid
>>>
>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>> >
>>> > As Dawid pointed out earlier on this thread, this is the rule for
>>> > Apache projects: a single -1 vote on a code change is a veto and
>>> > cannot be overridden. Furthermore, Robert is one of the people on this
>>> > project who worked the most on debugging subtle bugs, making Lucene
>>> > more robust and improving our test framework, so I'm listening when he
>>> > voices quality concerns.
>>> >
>>> > The argument against removing/raising the limit that resonates with me
>>> > the most is that it is a one-way door. As MikeS highlighted earlier on
>>> > this thread, implementations may want to take advantage of the fact
>>> > that there is a limit at some point too. This is why I don't want to
>>> > remove the limit and would prefer a slight increase, such as 2048 as
>>> > suggested in the original issue, which would enable most of the things
>>> > that users who have been asking about raising the limit would like to
>>> > do.
>>> >
>>> > I agree that the merge-time memory usage and slow indexing rate are
>>> > not great. But it's still possible to index multi-million vector
>>> > datasets with a 4GB heap without hitting OOMEs regardless of the
>>> > number of dimensions, and the feedback I'm seeing is that many users
>>> > are still interested in indexing multi-million vector datasets despite
>>> > the slow indexing rate. I wish we could do better, and vector indexing
>>> > is certainly more expert than text indexing, but it still is usable in
>>> > my opinion. I understand how giving Lucene more information about
>>> > vectors prior to indexing (e.g. clustering information as Jim pointed
>>> > out) could help make merging faster and more memory-efficient, but I
>>> > would really like to avoid making it a requirement for indexing
>>> > vectors as it also makes this feature much harder to use.
>>> >
>>> > On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>> > <a.benedetti@sease.io> wrote:
>>> > >
>>> > > I am very attentive to listen opinions but I am un-convinced here
>>> and I an not sure that a single person opinion should be allowed to be
>>> detrimental for such an important project.
>>> > >
>>> > > The limit as far as I know is literally just raising an exception.
>>> > > Removing it won't alter in any way the current performance for users
>>> in low dimensional space.
>>> > > Removing it will just enable more users to use Lucene.
>>> > >
>>> > > If new users in certain situations will be unhappy with the
>>> performance, they may contribute improvements.
>>> > > This is how you make progress.
>>> > >
>>> > > If it's a reputation thing, trust me that not allowing users to play
>>> with high dimensional space will equally damage it.
>>> > >
>>> > > To me it's really a no brainer.
>>> > > Removing the limit and enable people to use high dimensional vectors
>>> will take minutes.
>>> > > Improving the hnsw implementation can take months.
>>> > > Pick one to begin with...
>>> > >
>>> > > And there's no-one paying me here, no company interest whatsoever,
>>> actually I pay people to contribute, I am just convinced it's a good idea.
>>> > >
>>> > >
>>> > > On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>> > >>
>>> > >> I disagree with your categorization. I put in plenty of work and
>>> > >> experienced plenty of pain myself, writing tests and fighting these
>>> > >> issues, after i saw that, two releases in a row, vector indexing
>>> fell
>>> > >> over and hit integer overflows etc on small datasets:
>>> > >>
>>> > >> https://github.com/apache/lucene/pull/11905
>>> > >>
>>> > >> Attacking me isn't helping the situation.
>>> > >>
>>> > >> PS: when i said the "one guy who wrote the code" I didn't mean it in
>>> > >> any kind of demeaning fashion really. I meant to describe the
>>> current
>>> > >> state of usability with respect to indexing a few million docs with
>>> > >> high dimensions. You can scroll up the thread and see that at least
>>> > >> one other committer on the project experienced similar pain as me.
>>> > >> Then, think about users who aren't committers trying to use the
>>> > >> functionality!
>>> > >>
>>> > >> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <msokolov@gmail.com>
>>> wrote:
>>> > >> >
>>> > >> > What you said about increasing dimensions requiring a bigger ram
>>> buffer on merge is wrong. That's the point I was trying to make. Your
>>> concerns about merge costs are not wrong, but your conclusion that we need
>>> to limit dimensions is not justified.
>>> > >> >
>>> > >> > You complain that hnsw sucks it doesn't scale, but when I show it
>>> scales linearly with dimension you just ignore that and complain about
>>> something entirely different.
>>> > >> >
>>> > >> > You demand that people run all kinds of tests to prove you wrong
>>> but when they do, you don't listen and you won't put in the work yourself
>>> or complain that it's too hard.
>>> > >> >
>>> > >> > Then you complain about people not meeting you half way. Wow
>>> > >> >
>>> > >> > On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
>>> wrote:
>>> > >> >>
>>> > >> >> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>> > >> >> <michael.wechner@wyona.com> wrote:
>>> > >> >> >
>>> > >> >> > What exactly do you consider reasonable?
>>> > >> >>
>>> > >> >> Let's begin a real discussion by being HONEST about the current
>>> > >> >> status. Please put politically correct or your own company's
>>> wishes
>>> > >> >> aside, we know it's not in a good state.
>>> > >> >>
>>> > >> >> Current status is the one guy who wrote the code can set a
>>> > >> >> multi-gigabyte ram buffer and index a small dataset with 1024
>>> > >> >> dimensions in HOURS (i didn't ask what hardware).
>>> > >> >>
>>> > >> >> My concerns are everyone else except the one guy, I want it to be
>>> > >> >> usable. Increasing dimensions just means even bigger
>>> multi-gigabyte
>>> > >> >> ram buffer and bigger heap to avoid OOM on merge.
>>> > >> >> It is also a permanent backwards compatibility decision, we have
>>> to
>>> > >> >> support it once we do this and we can't just say "oops" and flip
>>> it
>>> > >> >> back.
>>> > >> >>
>>> > >> >> It is unclear to me, if the multi-gigabyte ram buffer is really
>>> to
>>> > >> >> avoid merges because they are so slow and it would be DAYS
>>> otherwise,
>>> > >> >> or if its to avoid merges so it doesn't hit OOM.
>>> > >> >> Also from personal experience, it takes trial and error (means
>>> > >> >> experiencing OOM on merge!!!) before you get those heap values
>>> correct
>>> > >> >> for your dataset. This usually means starting over which is
>>> > >> >> frustrating and wastes more time.
>>> > >> >>
>>> > >> >> Jim mentioned some ideas about the memory usage in IndexWriter,
>>> seems
>>> > >> >> to me like its a good idea. maybe the multigigabyte ram buffer
>>> can be
>>> > >> >> avoided in this way and performance improved by writing bigger
>>> > >> >> segments with lucene's defaults. But this doesn't mean we can
>>> simply
>>> > >> >> ignore the horrors of what happens on merge. merging needs to
>>> scale so
>>> > >> >> that indexing really scales.
>>> > >> >>
>>> > >> >> At least it shouldnt spike RAM on trivial data amounts and cause
>>> OOM,
>>> > >> >> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>> > >> >> fashion when indexing.
>>> > >> >>
>>> > >> >>
>>> ---------------------------------------------------------------------
>>> > >> >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > >> >> For additional commands, e-mail: dev-help@lucene.apache.org
>>> > >> >>
>>> > >>
>>> > >>
>>> ---------------------------------------------------------------------
>>> > >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > >> For additional commands, e-mail: dev-help@lucene.apache.org
>>> > >>
>>> >
>>> >
>>> > --
>>> > Adrien
>>> >
>>> > ---------------------------------------------------------------------
>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>> >
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>
>>>
>
> --
> http://www.needhamsoftware.com (work)
> http://www.the111shift.com (play)
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

lucene at mikemccandless

Apr 11, 2023, 5:29 AM

Post #85 of 99 (942 views)

+1 to test on real vector data -- if you test on synthetic data you draw
synthetic conclusions.

Can someone post the theoretical performance (CPU and RAM required) of HNSW
construction? Do we know/believe our HNSW implementation has achieved that
theoretical big-O performance? Maybe we have some silly performance bug
that's causing it not to?

As I understand it, HNSW makes the tradeoff of costly construction for
faster searching, which is typically the right tradeoff for search use
cases. We do this in other parts of the Lucene index too.

Lucene will do a logarithmic number of merges over time, i.e. each doc will
be merged O(log(N)) times in its lifetime in the index. We need to
multiply that by the cost of re-building the whole HNSW graph on each
merge. BTW, other things in Lucene, like BKD/dimensional points, also
rebuild the whole data structure on each merge, I think? But, as Rob
pointed out, stored fields merging do indeed do some sneaky tricks to avoid
excessive block decompress/recompress on each merge.

> As I understand it, vetoes must have technical merit. I'm not sure that
this veto rises to "technical merit" on 2 counts:

Actually I think Robert's veto stands on its technical merit already.
Robert's take on technical matters very much resonate with me, even if he
is sometimes prickly in how he expresses them ;)

His point is that we, as a dev community, are not paying enough attention
to the indexing performance of our KNN algo (HNSW) and implementation, and
that it is reckless to increase / remove limits in that state. It is
indeed a one-way door decision and one must confront such decisions with
caution, especially for such a widely used base infrastructure as Lucene.
We don't even advertise today in our javadocs that you need XXX heap if you
index vectors with dimension Y, fanout X, levels Z, etc.

RAM used during merging is unaffected by dimensionality, but is affected by
fanout, because the HNSW graph (not the raw vectors) is memory resident, I
think? Maybe we could move it off-heap and let the OS manage the memory
(and still document the RAM requirements)? Maybe merge RAM costs should be
accounted for in IW's RAM buffer accounting? It is not today, and there
are some other things that use non-trivial RAM, e.g. the doc mapping (to
compress docid space when deletions are reclaimed).

When we added KNN vector testing to Lucene's nightly benchmarks, the
indexing time massively increased -- see annotations DH and DP here:
https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly
benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of
course, that is using a single thread for indexing (on a box that has 128
cores!) so we produce a deterministic index every night ...

Stepping out (meta) a bit ... this discussion is precisely one of the
awesome benefits of the (informed) veto. It means risky changes to the
software, as determined by any single informed developer on the project,
can force a healthy discussion about the problem at hand. Robert is
legitimately concerned about a real issue and so we should use our creative
energies to characterize our HNSW implementation's performance, document it
clearly for users, and uncover ways to improve it.

Mike McCandless

http://blog.mikemccandless.com

On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <a.benedetti@sease.io>
wrote:

> I think Gus points are on target.
>
> I recommend we move this forward in this way:
> We stop any discussion and everyone interested proposes an option with a
> motivation, then we aggregate the options and we create a Vote maybe?
>
> I am also on the same page on the fact that a veto should come with a
> clear and reasonable technical merit, which also in my opinion has not come
> yet.
>
> I also apologise if any of my words sounded harsh or personal attacks,
> never meant to do so.
>
> My proposed option:
>
> 1) remove the limit and potentially make it configurable,
> Motivation:
> The system administrator can enforce a limit its users need to respect
> that it's in line with whatever the admin decided to be acceptable for
> them.
> Default can stay the current one.
>
> That's my favourite at the moment, but I agree that potentially in the
> future this may need to change, as we may optimise the data structures for
> certain dimensions. I am a big fan of Yagni (you aren't going to need it)
> so I am ok we'll face a different discussion if that happens in the future.
>
>
>
> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>
>> What I see so far:
>>
>> 1. Much positive support for raising the limit
>> 2. Slightly less support for removing it or making it configurable
>> 3. A single veto which argues that a (as yet undefined) performance
>> standard must be met before raising the limit
>> 4. Hot tempers (various) making this discussion difficult
>>
>> As I understand it, vetoes must have technical merit. I'm not sure that
>> this veto rises to "technical merit" on 2 counts:
>>
>> 1. No standard for the performance is given so it cannot be
>> technically met. Without hard criteria it's a moving target.
>> 2. It appears to encode a valuation of the user's time, and that
>> valuation is really up to the user. Some users may consider 2hours useless
>> and not worth it, and others might happily wait 2 hours. This is not a
>> technical decision, it's a business decision regarding the relative value
>> of the time invested vs the value of the result. If I can cure cancer by
>> indexing for a year, that might be worth it... (hyperbole of course).
>>
>> Things I would consider to have technical merit that I don't hear:
>>
>> 1. Impact on the speed of **other** indexing operations. (devaluation
>> of other functionality)
>> 2. Actual scenarios that work when the limit is low and fail when the
>> limit is high (new failure on the same data with the limit raised).
>>
>> One thing that might or might not have technical merit
>>
>> 1. If someone feels there is a lack of documentation of the
>> costs/performance implications of using large vectors, possibly including
>> reproducible benchmarks establishing the scaling behavior (there seems to
>> be disagreement on O(n) vs O(n^2)).
>>
>> The users *should* know what they are getting into, but if the cost is
>> worth it to them, they should be able to pay it without forking the
>> project. If this veto causes a fork that's not good.
>>
>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com>
>> wrote:
>>
>>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100
>>> and 300 dimensional varieties and can easily enough generate large numbers
>>> of vector documents from the articles data. To go higher we could
>>> concatenate vectors from that and I believe the performance numbers would
>>> be plausible.
>>>
>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>>>
>>>> Can we set up a branch in which the limit is bumped to 2048, then have
>>>> a realistic, free data set (wikipedia sample or something) that has,
>>>> say, 5 million docs and vectors created using public data (glove
>>>> pre-trained embeddings or the like)? We then could run indexing on the
>>>> same hardware with 512, 1024 and 2048 and see what the numbers, limits
>>>> and behavior actually are.
>>>>
>>>> I can help in writing this but not until after Easter.
>>>>
>>>>
>>>> Dawid
>>>>
>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>>> >
>>>> > As Dawid pointed out earlier on this thread, this is the rule for
>>>> > Apache projects: a single -1 vote on a code change is a veto and
>>>> > cannot be overridden. Furthermore, Robert is one of the people on this
>>>> > project who worked the most on debugging subtle bugs, making Lucene
>>>> > more robust and improving our test framework, so I'm listening when he
>>>> > voices quality concerns.
>>>> >
>>>> > The argument against removing/raising the limit that resonates with me
>>>> > the most is that it is a one-way door. As MikeS highlighted earlier on
>>>> > this thread, implementations may want to take advantage of the fact
>>>> > that there is a limit at some point too. This is why I don't want to
>>>> > remove the limit and would prefer a slight increase, such as 2048 as
>>>> > suggested in the original issue, which would enable most of the things
>>>> > that users who have been asking about raising the limit would like to
>>>> > do.
>>>> >
>>>> > I agree that the merge-time memory usage and slow indexing rate are
>>>> > not great. But it's still possible to index multi-million vector
>>>> > datasets with a 4GB heap without hitting OOMEs regardless of the
>>>> > number of dimensions, and the feedback I'm seeing is that many users
>>>> > are still interested in indexing multi-million vector datasets despite
>>>> > the slow indexing rate. I wish we could do better, and vector indexing
>>>> > is certainly more expert than text indexing, but it still is usable in
>>>> > my opinion. I understand how giving Lucene more information about
>>>> > vectors prior to indexing (e.g. clustering information as Jim pointed
>>>> > out) could help make merging faster and more memory-efficient, but I
>>>> > would really like to avoid making it a requirement for indexing
>>>> > vectors as it also makes this feature much harder to use.
>>>> >
>>>> > On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>>> > <a.benedetti@sease.io> wrote:
>>>> > >
>>>> > > I am very attentive to listen opinions but I am un-convinced here
>>>> and I an not sure that a single person opinion should be allowed to be
>>>> detrimental for such an important project.
>>>> > >
>>>> > > The limit as far as I know is literally just raising an exception.
>>>> > > Removing it won't alter in any way the current performance for
>>>> users in low dimensional space.
>>>> > > Removing it will just enable more users to use Lucene.
>>>> > >
>>>> > > If new users in certain situations will be unhappy with the
>>>> performance, they may contribute improvements.
>>>> > > This is how you make progress.
>>>> > >
>>>> > > If it's a reputation thing, trust me that not allowing users to
>>>> play with high dimensional space will equally damage it.
>>>> > >
>>>> > > To me it's really a no brainer.
>>>> > > Removing the limit and enable people to use high dimensional
>>>> vectors will take minutes.
>>>> > > Improving the hnsw implementation can take months.
>>>> > > Pick one to begin with...
>>>> > >
>>>> > > And there's no-one paying me here, no company interest whatsoever,
>>>> actually I pay people to contribute, I am just convinced it's a good idea.
>>>> > >
>>>> > >
>>>> > > On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>>> > >>
>>>> > >> I disagree with your categorization. I put in plenty of work and
>>>> > >> experienced plenty of pain myself, writing tests and fighting these
>>>> > >> issues, after i saw that, two releases in a row, vector indexing
>>>> fell
>>>> > >> over and hit integer overflows etc on small datasets:
>>>> > >>
>>>> > >> https://github.com/apache/lucene/pull/11905
>>>> > >>
>>>> > >> Attacking me isn't helping the situation.
>>>> > >>
>>>> > >> PS: when i said the "one guy who wrote the code" I didn't mean it
>>>> in
>>>> > >> any kind of demeaning fashion really. I meant to describe the
>>>> current
>>>> > >> state of usability with respect to indexing a few million docs with
>>>> > >> high dimensions. You can scroll up the thread and see that at least
>>>> > >> one other committer on the project experienced similar pain as me.
>>>> > >> Then, think about users who aren't committers trying to use the
>>>> > >> functionality!
>>>> > >>
>>>> > >> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <
>>>> msokolov@gmail.com> wrote:
>>>> > >> >
>>>> > >> > What you said about increasing dimensions requiring a bigger ram
>>>> buffer on merge is wrong. That's the point I was trying to make. Your
>>>> concerns about merge costs are not wrong, but your conclusion that we need
>>>> to limit dimensions is not justified.
>>>> > >> >
>>>> > >> > You complain that hnsw sucks it doesn't scale, but when I show
>>>> it scales linearly with dimension you just ignore that and complain about
>>>> something entirely different.
>>>> > >> >
>>>> > >> > You demand that people run all kinds of tests to prove you wrong
>>>> but when they do, you don't listen and you won't put in the work yourself
>>>> or complain that it's too hard.
>>>> > >> >
>>>> > >> > Then you complain about people not meeting you half way. Wow
>>>> > >> >
>>>> > >> > On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
>>>> wrote:
>>>> > >> >>
>>>> > >> >> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>>> > >> >> <michael.wechner@wyona.com> wrote:
>>>> > >> >> >
>>>> > >> >> > What exactly do you consider reasonable?
>>>> > >> >>
>>>> > >> >> Let's begin a real discussion by being HONEST about the current
>>>> > >> >> status. Please put politically correct or your own company's
>>>> wishes
>>>> > >> >> aside, we know it's not in a good state.
>>>> > >> >>
>>>> > >> >> Current status is the one guy who wrote the code can set a
>>>> > >> >> multi-gigabyte ram buffer and index a small dataset with 1024
>>>> > >> >> dimensions in HOURS (i didn't ask what hardware).
>>>> > >> >>
>>>> > >> >> My concerns are everyone else except the one guy, I want it to
>>>> be
>>>> > >> >> usable. Increasing dimensions just means even bigger
>>>> multi-gigabyte
>>>> > >> >> ram buffer and bigger heap to avoid OOM on merge.
>>>> > >> >> It is also a permanent backwards compatibility decision, we
>>>> have to
>>>> > >> >> support it once we do this and we can't just say "oops" and
>>>> flip it
>>>> > >> >> back.
>>>> > >> >>
>>>> > >> >> It is unclear to me, if the multi-gigabyte ram buffer is really
>>>> to
>>>> > >> >> avoid merges because they are so slow and it would be DAYS
>>>> otherwise,
>>>> > >> >> or if its to avoid merges so it doesn't hit OOM.
>>>> > >> >> Also from personal experience, it takes trial and error (means
>>>> > >> >> experiencing OOM on merge!!!) before you get those heap values
>>>> correct
>>>> > >> >> for your dataset. This usually means starting over which is
>>>> > >> >> frustrating and wastes more time.
>>>> > >> >>
>>>> > >> >> Jim mentioned some ideas about the memory usage in IndexWriter,
>>>> seems
>>>> > >> >> to me like its a good idea. maybe the multigigabyte ram buffer
>>>> can be
>>>> > >> >> avoided in this way and performance improved by writing bigger
>>>> > >> >> segments with lucene's defaults. But this doesn't mean we can
>>>> simply
>>>> > >> >> ignore the horrors of what happens on merge. merging needs to
>>>> scale so
>>>> > >> >> that indexing really scales.
>>>> > >> >>
>>>> > >> >> At least it shouldnt spike RAM on trivial data amounts and
>>>> cause OOM,
>>>> > >> >> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>>> > >> >> fashion when indexing.
>>>> > >> >>
>>>> > >> >>
>>>> ---------------------------------------------------------------------
>>>> > >> >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> > >> >> For additional commands, e-mail: dev-help@lucene.apache.org
>>>> > >> >>
>>>> > >>
>>>> > >>
>>>> ---------------------------------------------------------------------
>>>> > >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> > >> For additional commands, e-mail: dev-help@lucene.apache.org
>>>> > >>
>>>> >
>>>> >
>>>> > --
>>>> > Adrien
>>>> >
>>>> > ---------------------------------------------------------------------
>>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>>> >
>>>>
>>>> ---------------------------------------------------------------------
>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>
>>>>
>>
>> --
>> http://www.needhamsoftware.com (work)
>> http://www.the111shift.com (play)
>>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 11, 2023, 6:30 AM

Post #86 of 99 (942 views)

What exactly do you consider real vector data? Vector data which is
based on texts written by humans?

I am asking, because I recently attended the following presentation by
Anastassia Shaitarova (UZH Institute for Computational Linguistics,
https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)

*----
*

*Can we Identify Machine-Generated Text? An Overview of Current Approaches*
by Anastassia Shaitarova (UZH Institute for Computational Linguistics)

/The detection of machine-generated text has become increasingly
important due to the prevalence of automated content generation and its
potential for misuse. In this talk, we will discuss the motivation for
automatic detection of generated text. We will present the currently
available methods, including feature-based classification as a “first
line-of-defense.” We will provide an overview of the detection tools
that have been made available so far and discuss their limitations.
Finally, we will reflect on some open problems associated with the
automatic discrimination of generated texts./

/----/

and her conclusion was that it has become basically impossible to
differentiate between text generated by humans and text generated by for
example ChatGPT.

Whereas others have a slightly different opinion, see for example

https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/

But I would argue that real world and synthetic have become close enough
that testing performance and scalability of indexing should be possible
with synthetic data.

I completely agree that we have to base our discussions and decisions on
scientific methods and that we have to make sure that Lucene performs
and scales well and that we understand the limits and what is going on
under the hood.

Thanks

Michael W

Am 11.04.23 um 14:29 schrieb Michael McCandless:
> +1 to test on real vector data -- if you test on synthetic data you
> draw synthetic conclusions.
>
> Can someone post the theoretical performance (CPU and RAM required) of
> HNSW construction? Do we know/believe our HNSW implementation has
> achieved that theoretical big-O performance? Maybe we have some silly
> performance bug that's causing it not to?
>
> As I understand it, HNSW makes the tradeoff of costly construction for
> faster searching, which is typically the right tradeoff for search use
> cases. We do this in other parts of the Lucene index too.
>
> Lucene will do a logarithmic number of merges over time, i.e. each doc
> will be merged O(log(N)) times in its lifetime in the index. We need
> to multiply that by the cost of re-building the whole HNSW graph on
> each merge. BTW, other things in Lucene, like BKD/dimensional points,
> also rebuild the whole data structure on each merge, I think? But, as
> Rob pointed out, stored fields merging do indeed do some sneaky tricks
> to avoid excessive block decompress/recompress on each merge.
>
> > As I understand it, vetoes must have technical merit. I'm not sure
> that this veto rises to "technical merit" on 2 counts:
>
> Actually I think Robert's veto stands on its technical merit already.
> Robert's take on technical matters very much resonate with me, even if
> he is sometimes prickly in how he expresses them ;)
>
> His point is that we, as a dev community, are not paying enough
> attention to the indexing performance of our KNN algo (HNSW) and
> implementation, and that it is reckless to increase / remove limits in
> that state. It is indeed a one-way door decision and one must
> confront such decisions with caution, especially for such a widely
> used base infrastructure as Lucene. We don't even advertise today in
> our javadocs that you need XXX heap if you index vectors with
> dimension Y, fanout X, levels Z, etc.
>
> RAM used during merging is unaffected by dimensionality, but is
> affected by fanout, because the HNSW graph (not the raw vectors) is
> memory resident, I think? Maybe we could move it off-heap and let the
> OS manage the memory (and still document the RAM requirements)? Maybe
> merge RAM costs should be accounted for in IW's RAM buffer
> accounting? It is not today, and there are some other things that use
> non-trivial RAM, e.g. the doc mapping (to compress docid space when
> deletions are reclaimed).
>
> When we added KNN vector testing to Lucene's nightly benchmarks, the
> indexing time massively increased -- see annotations DH and DP here:
> https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly
> benchmarks now start at 6 PM and don't finish until ~14.5 hours
> later. Of course, that is using a single thread for indexing (on a
> box that has 128 cores!) so we produce a deterministic index every
> night ...
>
> Stepping out (meta) a bit ... this discussion is precisely one of the
> awesome benefits of the (informed) veto. It means risky changes to
> the software, as determined by any single informed developer on the
> project, can force a healthy discussion about the problem at hand.
> Robert is legitimately concerned about a real issue and so we should
> use our creative energies to characterize our HNSW implementation's
> performance, document it clearly for users, and uncover ways to
> improve it.
>
> Mike McCandless
>
> http://blog.mikemccandless.com
>
>
> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti
> <a.benedetti@sease.io> wrote:
>
> I think Gus points are on target.
>
> I recommend we move this forward in this way:
> We stop any discussion and everyone interested proposes an option
> with a motivation, then we aggregate the options and we create a
> Vote maybe?
>
> I am also on the same page on the fact that a veto should come
> with a clear and reasonable technical merit, which also in my
> opinion has not come yet.
>
> I also apologise if any of my words sounded harsh or personal
> attacks, never meant to do so.
>
> My proposed option:
>
> 1) remove the limit and potentially make it configurable,
> Motivation:
> The system administrator can enforce a limit its users need to
> respect that it's in line with whatever the admin decided to be
> acceptable for them.
> Default can stay the current one.
>
> That's my favourite at the moment, but I agree that potentially in
> the future this may need to change, as we may optimise the data
> structures for certain dimensions. I am a big fan of Yagni (you
> aren't going to need it) so I am ok we'll face a different
> discussion if that happens in the future.
>
>
>
> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>
> What I see so far:
>
> 1. Much positive support for raising the limit
> 2. Slightly less support for removing it or making it
> configurable
> 3. A single veto which argues that a (as yet undefined)
> performance standard must be met before raising the limit
> 4. Hot tempers (various) making this discussion difficult
>
> As I understand it, vetoes must have technical merit. I'm not
> sure that this veto rises to "technical merit" on 2 counts:
>
> 1. No standard for the performance is given so it cannot be
> technically met. Without hard criteria it's a moving target.
> 2. It appears to encode a valuation of the user's time, and
> that valuation is really up to the user. Some users may
> consider 2hours useless and not worth it, and others might
> happily wait 2 hours. This is not a technical decision,
> it's a business decision regarding the relative value of
> the time invested vs the value of the result. If I can
> cure cancer by indexing for a year, that might be worth
> it... (hyperbole of course).
>
> Things I would consider to have technical merit that I don't hear:
>
> 1. Impact on the speed of **other** indexing operations.
> (devaluation of other functionality)
> 2. Actual scenarios that work when the limit is low and fail
> when the limit is high (new failure on the same data with
> the limit raised).
>
> One thing that might or might not have technical merit
>
> 1. If someone feels there is a lack of documentation of the
> costs/performance implications of using large vectors,
> possibly including reproducible benchmarks establishing
> the scaling behavior (there seems to be disagreement on
> O(n) vs O(n^2)).
>
> The users *should* know what they are getting into, but if the
> cost is worth it to them, they should be able to pay it
> without forking the project. If this veto causes a fork that's
> not good.
>
> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov
> <msokolov@gmail.com> wrote:
>
> We do have a dataset built from Wikipedia in luceneutil.
> It comes in 100 and 300 dimensional varieties and can
> easily enough generate large numbers of vector documents
> from the articles data. To go higher we could concatenate
> vectors from that and I believe the performance numbers
> would be plausible.
>
> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss
> <dawid.weiss@gmail.com> wrote:
>
> Can we set up a branch in which the limit is bumped to
> 2048, then have
> a realistic, free data set (wikipedia sample or
> something) that has,
> say, 5 million docs and vectors created using public
> data (glove
> pre-trained embeddings or the like)? We then could run
> indexing on the
> same hardware with 512, 1024 and 2048 and see what the
> numbers, limits
> and behavior actually are.
>
> I can help in writing this but not until after Easter.
>
>
> Dawid
>
> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand
> <jpountz@gmail.com> wrote:
> >
> > As Dawid pointed out earlier on this thread, this is
> the rule for
> > Apache projects: a single -1 vote on a code change
> is a veto and
> > cannot be overridden. Furthermore, Robert is one of
> the people on this
> > project who worked the most on debugging subtle
> bugs, making Lucene
> > more robust and improving our test framework, so I'm
> listening when he
> > voices quality concerns.
> >
> > The argument against removing/raising the limit that
> resonates with me
> > the most is that it is a one-way door. As MikeS
> highlighted earlier on
> > this thread, implementations may want to take
> advantage of the fact
> > that there is a limit at some point too. This is why
> I don't want to
> > remove the limit and would prefer a slight increase,
> such as 2048 as
> > suggested in the original issue, which would enable
> most of the things
> > that users who have been asking about raising the
> limit would like to
> > do.
> >
> > I agree that the merge-time memory usage and slow
> indexing rate are
> > not great. But it's still possible to index
> multi-million vector
> > datasets with a 4GB heap without hitting OOMEs
> regardless of the
> > number of dimensions, and the feedback I'm seeing is
> that many users
> > are still interested in indexing multi-million
> vector datasets despite
> > the slow indexing rate. I wish we could do better,
> and vector indexing
> > is certainly more expert than text indexing, but it
> still is usable in
> > my opinion. I understand how giving Lucene more
> information about
> > vectors prior to indexing (e.g. clustering
> information as Jim pointed
> > out) could help make merging faster and more
> memory-efficient, but I
> > would really like to avoid making it a requirement
> for indexing
> > vectors as it also makes this feature much harder to
> use.
> >
> > On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
> > <a.benedetti@sease.io> wrote:
> > >
> > > I am very attentive to listen opinions but I am
> un-convinced here and I an not sure that a single
> person opinion should be allowed to be detrimental for
> such an important project.
> > >
> > > The limit as far as I know is literally just
> raising an exception.
> > > Removing it won't alter in any way the current
> performance for users in low dimensional space.
> > > Removing it will just enable more users to use Lucene.
> > >
> > > If new users in certain situations will be unhappy
> with the performance, they may contribute improvements.
> > > This is how you make progress.
> > >
> > > If it's a reputation thing, trust me that not
> allowing users to play with high dimensional space
> will equally damage it.
> > >
> > > To me it's really a no brainer.
> > > Removing the limit and enable people to use high
> dimensional vectors will take minutes.
> > > Improving the hnsw implementation can take months.
> > > Pick one to begin with...
> > >
> > > And there's no-one paying me here, no company
> interest whatsoever, actually I pay people to
> contribute, I am just convinced it's a good idea.
> > >
> > >
> > > On Sat, 8 Apr 2023, 18:57 Robert Muir,
> <rcmuir@gmail.com> wrote:
> > >>
> > >> I disagree with your categorization. I put in
> plenty of work and
> > >> experienced plenty of pain myself, writing tests
> and fighting these
> > >> issues, after i saw that, two releases in a row,
> vector indexing fell
> > >> over and hit integer overflows etc on small datasets:
> > >>
> > >> https://github.com/apache/lucene/pull/11905
> > >>
> > >> Attacking me isn't helping the situation.
> > >>
> > >> PS: when i said the "one guy who wrote the code"
> I didn't mean it in
> > >> any kind of demeaning fashion really. I meant to
> describe the current
> > >> state of usability with respect to indexing a few
> million docs with
> > >> high dimensions. You can scroll up the thread and
> see that at least
> > >> one other committer on the project experienced
> similar pain as me.
> > >> Then, think about users who aren't committers
> trying to use the
> > >> functionality!
> > >>
> > >> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov
> <msokolov@gmail.com> wrote:
> > >> >
> > >> > What you said about increasing dimensions
> requiring a bigger ram buffer on merge is wrong.
> That's the point I was trying to make. Your concerns
> about merge costs are not wrong, but your conclusion
> that we need to limit dimensions is not justified.
> > >> >
> > >> > You complain that hnsw sucks it doesn't scale,
> but when I show it scales linearly with dimension you
> just ignore that and complain about something entirely
> different.
> > >> >
> > >> > You demand that people run all kinds of tests
> to prove you wrong but when they do, you don't listen
> and you won't put in the work yourself or complain
> that it's too hard.
> > >> >
> > >> > Then you complain about people not meeting you
> half way. Wow
> > >> >
> > >> > On Sat, Apr 8, 2023, 12:40 PM Robert Muir
> <rcmuir@gmail.com> wrote:
> > >> >>
> > >> >> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
> > >> >> <michael.wechner@wyona.com> wrote:
> > >> >> >
> > >> >> > What exactly do you consider reasonable?
> > >> >>
> > >> >> Let's begin a real discussion by being HONEST
> about the current
> > >> >> status. Please put politically correct or your
> own company's wishes
> > >> >> aside, we know it's not in a good state.
> > >> >>
> > >> >> Current status is the one guy who wrote the
> code can set a
> > >> >> multi-gigabyte ram buffer and index a small
> dataset with 1024
> > >> >> dimensions in HOURS (i didn't ask what hardware).
> > >> >>
> > >> >> My concerns are everyone else except the one
> guy, I want it to be
> > >> >> usable. Increasing dimensions just means even
> bigger multi-gigabyte
> > >> >> ram buffer and bigger heap to avoid OOM on merge.
> > >> >> It is also a permanent backwards compatibility
> decision, we have to
> > >> >> support it once we do this and we can't just
> say "oops" and flip it
> > >> >> back.
> > >> >>
> > >> >> It is unclear to me, if the multi-gigabyte ram
> buffer is really to
> > >> >> avoid merges because they are so slow and it
> would be DAYS otherwise,
> > >> >> or if its to avoid merges so it doesn't hit OOM.
> > >> >> Also from personal experience, it takes trial
> and error (means
> > >> >> experiencing OOM on merge!!!) before you get
> those heap values correct
> > >> >> for your dataset. This usually means starting
> over which is
> > >> >> frustrating and wastes more time.
> > >> >>
> > >> >> Jim mentioned some ideas about the memory
> usage in IndexWriter, seems
> > >> >> to me like its a good idea. maybe the
> multigigabyte ram buffer can be
> > >> >> avoided in this way and performance improved
> by writing bigger
> > >> >> segments with lucene's defaults. But this
> doesn't mean we can simply
> > >> >> ignore the horrors of what happens on merge.
> merging needs to scale so
> > >> >> that indexing really scales.
> > >> >>
> > >> >> At least it shouldnt spike RAM on trivial data
> amounts and cause OOM,
> > >> >> and definitely it shouldnt burn hours and
> hours of CPU in O(n^2)
> > >> >> fashion when indexing.
> > >> >>
> > >> >>
> ---------------------------------------------------------------------
> > >> >> To unsubscribe, e-mail:
> dev-unsubscribe@lucene.apache.org
> > >> >> For additional commands, e-mail:
> dev-help@lucene.apache.org
> > >> >>
> > >>
> > >>
> ---------------------------------------------------------------------
> > >> To unsubscribe, e-mail:
> dev-unsubscribe@lucene.apache.org
> > >> For additional commands, e-mail:
> dev-help@lucene.apache.org
> > >>
> >
> >
> > --
> > Adrien
> >
> >
> ---------------------------------------------------------------------
> > To unsubscribe, e-mail:
> dev-unsubscribe@lucene.apache.org
> > For additional commands, e-mail:
> dev-help@lucene.apache.org
> >
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail:
> dev-help@lucene.apache.org
>
>
>
> --
> http://www.needhamsoftware.com (work)
> http://www.the111shift.com (play)
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 11, 2023, 10:05 AM

Post #87 of 99 (942 views)

> What exactly do you consider real vector data? Vector data which is based on texts written by humans?

We have plenty of text; the problem is coming up with a realistic
vector model that requires as many dimensions as people seem to be
demanding. As I said above, after surveying huggingface I couldn't
find any text-based model using more than 768 dimensions. So far we
have some ideas of generating higher-dimensional data by dithering or
concatenating existing data, but it seems artificial.

On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
<michael.wechner@wyona.com> wrote:
>
> What exactly do you consider real vector data? Vector data which is based on texts written by humans?
>
> I am asking, because I recently attended the following presentation by Anastassia Shaitarova (UZH Institute for Computational Linguistics, https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>
> ----
>
> Can we Identify Machine-Generated Text? An Overview of Current Approaches
> by Anastassia Shaitarova (UZH Institute for Computational Linguistics)
>
> The detection of machine-generated text has become increasingly important due to the prevalence of automated content generation and its potential for misuse. In this talk, we will discuss the motivation for automatic detection of generated text. We will present the currently available methods, including feature-based classification as a “first line-of-defense.” We will provide an overview of the detection tools that have been made available so far and discuss their limitations. Finally, we will reflect on some open problems associated with the automatic discrimination of generated texts.
>
> ----
>
> and her conclusion was that it has become basically impossible to differentiate between text generated by humans and text generated by for example ChatGPT.
>
> Whereas others have a slightly different opinion, see for example
>
> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>
> But I would argue that real world and synthetic have become close enough that testing performance and scalability of indexing should be possible with synthetic data.
>
> I completely agree that we have to base our discussions and decisions on scientific methods and that we have to make sure that Lucene performs and scales well and that we understand the limits and what is going on under the hood.
>
> Thanks
>
> Michael W
>
>
>
>
>
> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>
> +1 to test on real vector data -- if you test on synthetic data you draw synthetic conclusions.
>
> Can someone post the theoretical performance (CPU and RAM required) of HNSW construction? Do we know/believe our HNSW implementation has achieved that theoretical big-O performance? Maybe we have some silly performance bug that's causing it not to?
>
> As I understand it, HNSW makes the tradeoff of costly construction for faster searching, which is typically the right tradeoff for search use cases. We do this in other parts of the Lucene index too.
>
> Lucene will do a logarithmic number of merges over time, i.e. each doc will be merged O(log(N)) times in its lifetime in the index. We need to multiply that by the cost of re-building the whole HNSW graph on each merge. BTW, other things in Lucene, like BKD/dimensional points, also rebuild the whole data structure on each merge, I think? But, as Rob pointed out, stored fields merging do indeed do some sneaky tricks to avoid excessive block decompress/recompress on each merge.
>
> > As I understand it, vetoes must have technical merit. I'm not sure that this veto rises to "technical merit" on 2 counts:
>
> Actually I think Robert's veto stands on its technical merit already. Robert's take on technical matters very much resonate with me, even if he is sometimes prickly in how he expresses them ;)
>
> His point is that we, as a dev community, are not paying enough attention to the indexing performance of our KNN algo (HNSW) and implementation, and that it is reckless to increase / remove limits in that state. It is indeed a one-way door decision and one must confront such decisions with caution, especially for such a widely used base infrastructure as Lucene. We don't even advertise today in our javadocs that you need XXX heap if you index vectors with dimension Y, fanout X, levels Z, etc.
>
> RAM used during merging is unaffected by dimensionality, but is affected by fanout, because the HNSW graph (not the raw vectors) is memory resident, I think? Maybe we could move it off-heap and let the OS manage the memory (and still document the RAM requirements)? Maybe merge RAM costs should be accounted for in IW's RAM buffer accounting? It is not today, and there are some other things that use non-trivial RAM, e.g. the doc mapping (to compress docid space when deletions are reclaimed).
>
> When we added KNN vector testing to Lucene's nightly benchmarks, the indexing time massively increased -- see annotations DH and DP here: https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of course, that is using a single thread for indexing (on a box that has 128 cores!) so we produce a deterministic index every night ...
>
> Stepping out (meta) a bit ... this discussion is precisely one of the awesome benefits of the (informed) veto. It means risky changes to the software, as determined by any single informed developer on the project, can force a healthy discussion about the problem at hand. Robert is legitimately concerned about a real issue and so we should use our creative energies to characterize our HNSW implementation's performance, document it clearly for users, and uncover ways to improve it.
>
> Mike McCandless
>
> http://blog.mikemccandless.com
>
>
> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <a.benedetti@sease.io> wrote:
>>
>> I think Gus points are on target.
>>
>> I recommend we move this forward in this way:
>> We stop any discussion and everyone interested proposes an option with a motivation, then we aggregate the options and we create a Vote maybe?
>>
>> I am also on the same page on the fact that a veto should come with a clear and reasonable technical merit, which also in my opinion has not come yet.
>>
>> I also apologise if any of my words sounded harsh or personal attacks, never meant to do so.
>>
>> My proposed option:
>>
>> 1) remove the limit and potentially make it configurable,
>> Motivation:
>> The system administrator can enforce a limit its users need to respect that it's in line with whatever the admin decided to be acceptable for them.
>> Default can stay the current one.
>>
>> That's my favourite at the moment, but I agree that potentially in the future this may need to change, as we may optimise the data structures for certain dimensions. I am a big fan of Yagni (you aren't going to need it) so I am ok we'll face a different discussion if that happens in the future.
>>
>>
>>
>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>>>
>>> What I see so far:
>>>
>>> Much positive support for raising the limit
>>> Slightly less support for removing it or making it configurable
>>> A single veto which argues that a (as yet undefined) performance standard must be met before raising the limit
>>> Hot tempers (various) making this discussion difficult
>>>
>>> As I understand it, vetoes must have technical merit. I'm not sure that this veto rises to "technical merit" on 2 counts:
>>>
>>> No standard for the performance is given so it cannot be technically met. Without hard criteria it's a moving target.
>>> It appears to encode a valuation of the user's time, and that valuation is really up to the user. Some users may consider 2hours useless and not worth it, and others might happily wait 2 hours. This is not a technical decision, it's a business decision regarding the relative value of the time invested vs the value of the result. If I can cure cancer by indexing for a year, that might be worth it... (hyperbole of course).
>>>
>>> Things I would consider to have technical merit that I don't hear:
>>>
>>> Impact on the speed of **other** indexing operations. (devaluation of other functionality)
>>> Actual scenarios that work when the limit is low and fail when the limit is high (new failure on the same data with the limit raised).
>>>
>>> One thing that might or might not have technical merit
>>>
>>> If someone feels there is a lack of documentation of the costs/performance implications of using large vectors, possibly including reproducible benchmarks establishing the scaling behavior (there seems to be disagreement on O(n) vs O(n^2)).
>>>
>>> The users *should* know what they are getting into, but if the cost is worth it to them, they should be able to pay it without forking the project. If this veto causes a fork that's not good.
>>>
>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com> wrote:
>>>>
>>>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100 and 300 dimensional varieties and can easily enough generate large numbers of vector documents from the articles data. To go higher we could concatenate vectors from that and I believe the performance numbers would be plausible.
>>>>
>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>>>>>
>>>>> Can we set up a branch in which the limit is bumped to 2048, then have
>>>>> a realistic, free data set (wikipedia sample or something) that has,
>>>>> say, 5 million docs and vectors created using public data (glove
>>>>> pre-trained embeddings or the like)? We then could run indexing on the
>>>>> same hardware with 512, 1024 and 2048 and see what the numbers, limits
>>>>> and behavior actually are.
>>>>>
>>>>> I can help in writing this but not until after Easter.
>>>>>
>>>>>
>>>>> Dawid
>>>>>
>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>>>> >
>>>>> > As Dawid pointed out earlier on this thread, this is the rule for
>>>>> > Apache projects: a single -1 vote on a code change is a veto and
>>>>> > cannot be overridden. Furthermore, Robert is one of the people on this
>>>>> > project who worked the most on debugging subtle bugs, making Lucene
>>>>> > more robust and improving our test framework, so I'm listening when he
>>>>> > voices quality concerns.
>>>>> >
>>>>> > The argument against removing/raising the limit that resonates with me
>>>>> > the most is that it is a one-way door. As MikeS highlighted earlier on
>>>>> > this thread, implementations may want to take advantage of the fact
>>>>> > that there is a limit at some point too. This is why I don't want to
>>>>> > remove the limit and would prefer a slight increase, such as 2048 as
>>>>> > suggested in the original issue, which would enable most of the things
>>>>> > that users who have been asking about raising the limit would like to
>>>>> > do.
>>>>> >
>>>>> > I agree that the merge-time memory usage and slow indexing rate are
>>>>> > not great. But it's still possible to index multi-million vector
>>>>> > datasets with a 4GB heap without hitting OOMEs regardless of the
>>>>> > number of dimensions, and the feedback I'm seeing is that many users
>>>>> > are still interested in indexing multi-million vector datasets despite
>>>>> > the slow indexing rate. I wish we could do better, and vector indexing
>>>>> > is certainly more expert than text indexing, but it still is usable in
>>>>> > my opinion. I understand how giving Lucene more information about
>>>>> > vectors prior to indexing (e.g. clustering information as Jim pointed
>>>>> > out) could help make merging faster and more memory-efficient, but I
>>>>> > would really like to avoid making it a requirement for indexing
>>>>> > vectors as it also makes this feature much harder to use.
>>>>> >
>>>>> > On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>>>> > <a.benedetti@sease.io> wrote:
>>>>> > >
>>>>> > > I am very attentive to listen opinions but I am un-convinced here and I an not sure that a single person opinion should be allowed to be detrimental for such an important project.
>>>>> > >
>>>>> > > The limit as far as I know is literally just raising an exception.
>>>>> > > Removing it won't alter in any way the current performance for users in low dimensional space.
>>>>> > > Removing it will just enable more users to use Lucene.
>>>>> > >
>>>>> > > If new users in certain situations will be unhappy with the performance, they may contribute improvements.
>>>>> > > This is how you make progress.
>>>>> > >
>>>>> > > If it's a reputation thing, trust me that not allowing users to play with high dimensional space will equally damage it.
>>>>> > >
>>>>> > > To me it's really a no brainer.
>>>>> > > Removing the limit and enable people to use high dimensional vectors will take minutes.
>>>>> > > Improving the hnsw implementation can take months.
>>>>> > > Pick one to begin with...
>>>>> > >
>>>>> > > And there's no-one paying me here, no company interest whatsoever, actually I pay people to contribute, I am just convinced it's a good idea.
>>>>> > >
>>>>> > >
>>>>> > > On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>>>> > >>
>>>>> > >> I disagree with your categorization. I put in plenty of work and
>>>>> > >> experienced plenty of pain myself, writing tests and fighting these
>>>>> > >> issues, after i saw that, two releases in a row, vector indexing fell
>>>>> > >> over and hit integer overflows etc on small datasets:
>>>>> > >>
>>>>> > >> https://github.com/apache/lucene/pull/11905
>>>>> > >>
>>>>> > >> Attacking me isn't helping the situation.
>>>>> > >>
>>>>> > >> PS: when i said the "one guy who wrote the code" I didn't mean it in
>>>>> > >> any kind of demeaning fashion really. I meant to describe the current
>>>>> > >> state of usability with respect to indexing a few million docs with
>>>>> > >> high dimensions. You can scroll up the thread and see that at least
>>>>> > >> one other committer on the project experienced similar pain as me.
>>>>> > >> Then, think about users who aren't committers trying to use the
>>>>> > >> functionality!
>>>>> > >>
>>>>> > >> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <msokolov@gmail.com> wrote:
>>>>> > >> >
>>>>> > >> > What you said about increasing dimensions requiring a bigger ram buffer on merge is wrong. That's the point I was trying to make. Your concerns about merge costs are not wrong, but your conclusion that we need to limit dimensions is not justified.
>>>>> > >> >
>>>>> > >> > You complain that hnsw sucks it doesn't scale, but when I show it scales linearly with dimension you just ignore that and complain about something entirely different.
>>>>> > >> >
>>>>> > >> > You demand that people run all kinds of tests to prove you wrong but when they do, you don't listen and you won't put in the work yourself or complain that it's too hard.
>>>>> > >> >
>>>>> > >> > Then you complain about people not meeting you half way. Wow
>>>>> > >> >
>>>>> > >> > On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com> wrote:
>>>>> > >> >>
>>>>> > >> >> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>>>> > >> >> <michael.wechner@wyona.com> wrote:
>>>>> > >> >> >
>>>>> > >> >> > What exactly do you consider reasonable?
>>>>> > >> >>
>>>>> > >> >> Let's begin a real discussion by being HONEST about the current
>>>>> > >> >> status. Please put politically correct or your own company's wishes
>>>>> > >> >> aside, we know it's not in a good state.
>>>>> > >> >>
>>>>> > >> >> Current status is the one guy who wrote the code can set a
>>>>> > >> >> multi-gigabyte ram buffer and index a small dataset with 1024
>>>>> > >> >> dimensions in HOURS (i didn't ask what hardware).
>>>>> > >> >>
>>>>> > >> >> My concerns are everyone else except the one guy, I want it to be
>>>>> > >> >> usable. Increasing dimensions just means even bigger multi-gigabyte
>>>>> > >> >> ram buffer and bigger heap to avoid OOM on merge.
>>>>> > >> >> It is also a permanent backwards compatibility decision, we have to
>>>>> > >> >> support it once we do this and we can't just say "oops" and flip it
>>>>> > >> >> back.
>>>>> > >> >>
>>>>> > >> >> It is unclear to me, if the multi-gigabyte ram buffer is really to
>>>>> > >> >> avoid merges because they are so slow and it would be DAYS otherwise,
>>>>> > >> >> or if its to avoid merges so it doesn't hit OOM.
>>>>> > >> >> Also from personal experience, it takes trial and error (means
>>>>> > >> >> experiencing OOM on merge!!!) before you get those heap values correct
>>>>> > >> >> for your dataset. This usually means starting over which is
>>>>> > >> >> frustrating and wastes more time.
>>>>> > >> >>
>>>>> > >> >> Jim mentioned some ideas about the memory usage in IndexWriter, seems
>>>>> > >> >> to me like its a good idea. maybe the multigigabyte ram buffer can be
>>>>> > >> >> avoided in this way and performance improved by writing bigger
>>>>> > >> >> segments with lucene's defaults. But this doesn't mean we can simply
>>>>> > >> >> ignore the horrors of what happens on merge. merging needs to scale so
>>>>> > >> >> that indexing really scales.
>>>>> > >> >>
>>>>> > >> >> At least it shouldnt spike RAM on trivial data amounts and cause OOM,
>>>>> > >> >> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>>>> > >> >> fashion when indexing.
>>>>> > >> >>
>>>>> > >> >> ---------------------------------------------------------------------
>>>>> > >> >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>> > >> >> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>> > >> >>
>>>>> > >>
>>>>> > >> ---------------------------------------------------------------------
>>>>> > >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>> > >> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>> > >>
>>>>> >
>>>>> >
>>>>> > --
>>>>> > Adrien
>>>>> >
>>>>> > ---------------------------------------------------------------------
>>>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>>>> >
>>>>>
>>>>> ---------------------------------------------------------------------
>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>
>>>
>>>
>>> --
>>> http://www.needhamsoftware.com (work)
>>> http://www.the111shift.com (play)
>
>

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 11, 2023, 12:07 PM

Post #88 of 99 (942 views)

I understand what you mean that it seems to be artificial, but I don't
understand why this matters to test performance and scalability of the
indexing?

Let's assume the limit of Lucene would be 4 instead of 1024 and there
are only open source models generating vectors with 4 dimensions, for
example

0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814

0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844

-0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106

-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114

and now I concatenate them to vectors with 8 dimensions

0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844

-0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114

and normalize them to length 1.

Why should this be any different to a model which is acting like a black
box generating vectors with 8 dimensions?

Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>> What exactly do you consider real vector data? Vector data which is based on texts written by humans?
> We have plenty of text; the problem is coming up with a realistic
> vector model that requires as many dimensions as people seem to be
> demanding. As I said above, after surveying huggingface I couldn't
> find any text-based model using more than 768 dimensions. So far we
> have some ideas of generating higher-dimensional data by dithering or
> concatenating existing data, but it seems artificial.
>
> On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
> <michael.wechner@wyona.com> wrote:
>> What exactly do you consider real vector data? Vector data which is based on texts written by humans?
>>
>> I am asking, because I recently attended the following presentation by Anastassia Shaitarova (UZH Institute for Computational Linguistics, https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>>
>> ----
>>
>> Can we Identify Machine-Generated Text? An Overview of Current Approaches
>> by Anastassia Shaitarova (UZH Institute for Computational Linguistics)
>>
>> The detection of machine-generated text has become increasingly important due to the prevalence of automated content generation and its potential for misuse. In this talk, we will discuss the motivation for automatic detection of generated text. We will present the currently available methods, including feature-based classification as a “first line-of-defense.” We will provide an overview of the detection tools that have been made available so far and discuss their limitations. Finally, we will reflect on some open problems associated with the automatic discrimination of generated texts.
>>
>> ----
>>
>> and her conclusion was that it has become basically impossible to differentiate between text generated by humans and text generated by for example ChatGPT.
>>
>> Whereas others have a slightly different opinion, see for example
>>
>> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>>
>> But I would argue that real world and synthetic have become close enough that testing performance and scalability of indexing should be possible with synthetic data.
>>
>> I completely agree that we have to base our discussions and decisions on scientific methods and that we have to make sure that Lucene performs and scales well and that we understand the limits and what is going on under the hood.
>>
>> Thanks
>>
>> Michael W
>>
>>
>>
>>
>>
>> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>>
>> +1 to test on real vector data -- if you test on synthetic data you draw synthetic conclusions.
>>
>> Can someone post the theoretical performance (CPU and RAM required) of HNSW construction? Do we know/believe our HNSW implementation has achieved that theoretical big-O performance? Maybe we have some silly performance bug that's causing it not to?
>>
>> As I understand it, HNSW makes the tradeoff of costly construction for faster searching, which is typically the right tradeoff for search use cases. We do this in other parts of the Lucene index too.
>>
>> Lucene will do a logarithmic number of merges over time, i.e. each doc will be merged O(log(N)) times in its lifetime in the index. We need to multiply that by the cost of re-building the whole HNSW graph on each merge. BTW, other things in Lucene, like BKD/dimensional points, also rebuild the whole data structure on each merge, I think? But, as Rob pointed out, stored fields merging do indeed do some sneaky tricks to avoid excessive block decompress/recompress on each merge.
>>
>>> As I understand it, vetoes must have technical merit. I'm not sure that this veto rises to "technical merit" on 2 counts:
>> Actually I think Robert's veto stands on its technical merit already. Robert's take on technical matters very much resonate with me, even if he is sometimes prickly in how he expresses them ;)
>>
>> His point is that we, as a dev community, are not paying enough attention to the indexing performance of our KNN algo (HNSW) and implementation, and that it is reckless to increase / remove limits in that state. It is indeed a one-way door decision and one must confront such decisions with caution, especially for such a widely used base infrastructure as Lucene. We don't even advertise today in our javadocs that you need XXX heap if you index vectors with dimension Y, fanout X, levels Z, etc.
>>
>> RAM used during merging is unaffected by dimensionality, but is affected by fanout, because the HNSW graph (not the raw vectors) is memory resident, I think? Maybe we could move it off-heap and let the OS manage the memory (and still document the RAM requirements)? Maybe merge RAM costs should be accounted for in IW's RAM buffer accounting? It is not today, and there are some other things that use non-trivial RAM, e.g. the doc mapping (to compress docid space when deletions are reclaimed).
>>
>> When we added KNN vector testing to Lucene's nightly benchmarks, the indexing time massively increased -- see annotations DH and DP here: https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of course, that is using a single thread for indexing (on a box that has 128 cores!) so we produce a deterministic index every night ...
>>
>> Stepping out (meta) a bit ... this discussion is precisely one of the awesome benefits of the (informed) veto. It means risky changes to the software, as determined by any single informed developer on the project, can force a healthy discussion about the problem at hand. Robert is legitimately concerned about a real issue and so we should use our creative energies to characterize our HNSW implementation's performance, document it clearly for users, and uncover ways to improve it.
>>
>> Mike McCandless
>>
>> http://blog.mikemccandless.com
>>
>>
>> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <a.benedetti@sease.io> wrote:
>>> I think Gus points are on target.
>>>
>>> I recommend we move this forward in this way:
>>> We stop any discussion and everyone interested proposes an option with a motivation, then we aggregate the options and we create a Vote maybe?
>>>
>>> I am also on the same page on the fact that a veto should come with a clear and reasonable technical merit, which also in my opinion has not come yet.
>>>
>>> I also apologise if any of my words sounded harsh or personal attacks, never meant to do so.
>>>
>>> My proposed option:
>>>
>>> 1) remove the limit and potentially make it configurable,
>>> Motivation:
>>> The system administrator can enforce a limit its users need to respect that it's in line with whatever the admin decided to be acceptable for them.
>>> Default can stay the current one.
>>>
>>> That's my favourite at the moment, but I agree that potentially in the future this may need to change, as we may optimise the data structures for certain dimensions. I am a big fan of Yagni (you aren't going to need it) so I am ok we'll face a different discussion if that happens in the future.
>>>
>>>
>>>
>>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>>>> What I see so far:
>>>>
>>>> Much positive support for raising the limit
>>>> Slightly less support for removing it or making it configurable
>>>> A single veto which argues that a (as yet undefined) performance standard must be met before raising the limit
>>>> Hot tempers (various) making this discussion difficult
>>>>
>>>> As I understand it, vetoes must have technical merit. I'm not sure that this veto rises to "technical merit" on 2 counts:
>>>>
>>>> No standard for the performance is given so it cannot be technically met. Without hard criteria it's a moving target.
>>>> It appears to encode a valuation of the user's time, and that valuation is really up to the user. Some users may consider 2hours useless and not worth it, and others might happily wait 2 hours. This is not a technical decision, it's a business decision regarding the relative value of the time invested vs the value of the result. If I can cure cancer by indexing for a year, that might be worth it... (hyperbole of course).
>>>>
>>>> Things I would consider to have technical merit that I don't hear:
>>>>
>>>> Impact on the speed of **other** indexing operations. (devaluation of other functionality)
>>>> Actual scenarios that work when the limit is low and fail when the limit is high (new failure on the same data with the limit raised).
>>>>
>>>> One thing that might or might not have technical merit
>>>>
>>>> If someone feels there is a lack of documentation of the costs/performance implications of using large vectors, possibly including reproducible benchmarks establishing the scaling behavior (there seems to be disagreement on O(n) vs O(n^2)).
>>>>
>>>> The users *should* know what they are getting into, but if the cost is worth it to them, they should be able to pay it without forking the project. If this veto causes a fork that's not good.
>>>>
>>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com> wrote:
>>>>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100 and 300 dimensional varieties and can easily enough generate large numbers of vector documents from the articles data. To go higher we could concatenate vectors from that and I believe the performance numbers would be plausible.
>>>>>
>>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>>>>>> Can we set up a branch in which the limit is bumped to 2048, then have
>>>>>> a realistic, free data set (wikipedia sample or something) that has,
>>>>>> say, 5 million docs and vectors created using public data (glove
>>>>>> pre-trained embeddings or the like)? We then could run indexing on the
>>>>>> same hardware with 512, 1024 and 2048 and see what the numbers, limits
>>>>>> and behavior actually are.
>>>>>>
>>>>>> I can help in writing this but not until after Easter.
>>>>>>
>>>>>>
>>>>>> Dawid
>>>>>>
>>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>>>>>> As Dawid pointed out earlier on this thread, this is the rule for
>>>>>>> Apache projects: a single -1 vote on a code change is a veto and
>>>>>>> cannot be overridden. Furthermore, Robert is one of the people on this
>>>>>>> project who worked the most on debugging subtle bugs, making Lucene
>>>>>>> more robust and improving our test framework, so I'm listening when he
>>>>>>> voices quality concerns.
>>>>>>>
>>>>>>> The argument against removing/raising the limit that resonates with me
>>>>>>> the most is that it is a one-way door. As MikeS highlighted earlier on
>>>>>>> this thread, implementations may want to take advantage of the fact
>>>>>>> that there is a limit at some point too. This is why I don't want to
>>>>>>> remove the limit and would prefer a slight increase, such as 2048 as
>>>>>>> suggested in the original issue, which would enable most of the things
>>>>>>> that users who have been asking about raising the limit would like to
>>>>>>> do.
>>>>>>>
>>>>>>> I agree that the merge-time memory usage and slow indexing rate are
>>>>>>> not great. But it's still possible to index multi-million vector
>>>>>>> datasets with a 4GB heap without hitting OOMEs regardless of the
>>>>>>> number of dimensions, and the feedback I'm seeing is that many users
>>>>>>> are still interested in indexing multi-million vector datasets despite
>>>>>>> the slow indexing rate. I wish we could do better, and vector indexing
>>>>>>> is certainly more expert than text indexing, but it still is usable in
>>>>>>> my opinion. I understand how giving Lucene more information about
>>>>>>> vectors prior to indexing (e.g. clustering information as Jim pointed
>>>>>>> out) could help make merging faster and more memory-efficient, but I
>>>>>>> would really like to avoid making it a requirement for indexing
>>>>>>> vectors as it also makes this feature much harder to use.
>>>>>>>
>>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>>>>>> <a.benedetti@sease.io> wrote:
>>>>>>>> I am very attentive to listen opinions but I am un-convinced here and I an not sure that a single person opinion should be allowed to be detrimental for such an important project.
>>>>>>>>
>>>>>>>> The limit as far as I know is literally just raising an exception.
>>>>>>>> Removing it won't alter in any way the current performance for users in low dimensional space.
>>>>>>>> Removing it will just enable more users to use Lucene.
>>>>>>>>
>>>>>>>> If new users in certain situations will be unhappy with the performance, they may contribute improvements.
>>>>>>>> This is how you make progress.
>>>>>>>>
>>>>>>>> If it's a reputation thing, trust me that not allowing users to play with high dimensional space will equally damage it.
>>>>>>>>
>>>>>>>> To me it's really a no brainer.
>>>>>>>> Removing the limit and enable people to use high dimensional vectors will take minutes.
>>>>>>>> Improving the hnsw implementation can take months.
>>>>>>>> Pick one to begin with...
>>>>>>>>
>>>>>>>> And there's no-one paying me here, no company interest whatsoever, actually I pay people to contribute, I am just convinced it's a good idea.
>>>>>>>>
>>>>>>>>
>>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>>>>>>>> I disagree with your categorization. I put in plenty of work and
>>>>>>>>> experienced plenty of pain myself, writing tests and fighting these
>>>>>>>>> issues, after i saw that, two releases in a row, vector indexing fell
>>>>>>>>> over and hit integer overflows etc on small datasets:
>>>>>>>>>
>>>>>>>>> https://github.com/apache/lucene/pull/11905
>>>>>>>>>
>>>>>>>>> Attacking me isn't helping the situation.
>>>>>>>>>
>>>>>>>>> PS: when i said the "one guy who wrote the code" I didn't mean it in
>>>>>>>>> any kind of demeaning fashion really. I meant to describe the current
>>>>>>>>> state of usability with respect to indexing a few million docs with
>>>>>>>>> high dimensions. You can scroll up the thread and see that at least
>>>>>>>>> one other committer on the project experienced similar pain as me.
>>>>>>>>> Then, think about users who aren't committers trying to use the
>>>>>>>>> functionality!
>>>>>>>>>
>>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <msokolov@gmail.com> wrote:
>>>>>>>>>> What you said about increasing dimensions requiring a bigger ram buffer on merge is wrong. That's the point I was trying to make. Your concerns about merge costs are not wrong, but your conclusion that we need to limit dimensions is not justified.
>>>>>>>>>>
>>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when I show it scales linearly with dimension you just ignore that and complain about something entirely different.
>>>>>>>>>>
>>>>>>>>>> You demand that people run all kinds of tests to prove you wrong but when they do, you don't listen and you won't put in the work yourself or complain that it's too hard.
>>>>>>>>>>
>>>>>>>>>> Then you complain about people not meeting you half way. Wow
>>>>>>>>>>
>>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com> wrote:
>>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>>>>>>>>>>>> What exactly do you consider reasonable?
>>>>>>>>>>> Let's begin a real discussion by being HONEST about the current
>>>>>>>>>>> status. Please put politically correct or your own company's wishes
>>>>>>>>>>> aside, we know it's not in a good state.
>>>>>>>>>>>
>>>>>>>>>>> Current status is the one guy who wrote the code can set a
>>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset with 1024
>>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>>>>>>>>>>>
>>>>>>>>>>> My concerns are everyone else except the one guy, I want it to be
>>>>>>>>>>> usable. Increasing dimensions just means even bigger multi-gigabyte
>>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>>>>>>>>>>> It is also a permanent backwards compatibility decision, we have to
>>>>>>>>>>> support it once we do this and we can't just say "oops" and flip it
>>>>>>>>>>> back.
>>>>>>>>>>>
>>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer is really to
>>>>>>>>>>> avoid merges because they are so slow and it would be DAYS otherwise,
>>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>>>>>>>>>>> Also from personal experience, it takes trial and error (means
>>>>>>>>>>> experiencing OOM on merge!!!) before you get those heap values correct
>>>>>>>>>>> for your dataset. This usually means starting over which is
>>>>>>>>>>> frustrating and wastes more time.
>>>>>>>>>>>
>>>>>>>>>>> Jim mentioned some ideas about the memory usage in IndexWriter, seems
>>>>>>>>>>> to me like its a good idea. maybe the multigigabyte ram buffer can be
>>>>>>>>>>> avoided in this way and performance improved by writing bigger
>>>>>>>>>>> segments with lucene's defaults. But this doesn't mean we can simply
>>>>>>>>>>> ignore the horrors of what happens on merge. merging needs to scale so
>>>>>>>>>>> that indexing really scales.
>>>>>>>>>>>
>>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts and cause OOM,
>>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>>>>>>>>>> fashion when indexing.
>>>>>>>>>>>
>>>>>>>>>>> ---------------------------------------------------------------------
>>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>>>>>>>
>>>>>>>>> ---------------------------------------------------------------------
>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>> Adrien
>>>>>>>
>>>>>>> ---------------------------------------------------------------------
>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>>>
>>>>>> ---------------------------------------------------------------------
>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>>>
>>>>
>>>> --
>>>> http://www.needhamsoftware.com (work)
>>>> http://www.the111shift.com (play)
>>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

kent.fitch at gmail

Apr 11, 2023, 7:51 PM

Post #89 of 99 (942 views)

I only know some characteristics of the openAI ada-002 vectors, although
they are a very popular as embeddings/text-characterisations as they allow
more accurate/"human meaningful" semantic search results with fewer
dimensions than their predecessors - I've evaluated a few different
embedding models, including some BERT variants, CLIP ViT-L-14 (with 768
dims, which was quite good), openAI's ada-001 (1024 dims) and babbage-001
(2048 dims), and ada-002 are qualitatively the best, although that will
certainly change!

In any case, ada-002 vectors have interesting characteristics that I think
mean you could confidently create synthetic vectors which would be hard to
distinguish from "real" vectors. I found this from looking at 47K ada-002
vectors generated across a full year (1994) of newspaper articles from the
Canberra Times and 200K wikipedia articles:
- there is no discernible/significant correlation between values in any
pair of dimensions
- all but 5 of the 1536 dimensions have an almost identical distribution of
values shown in the central blob on these graphs (that just show a few of
these 1531 dimensions with clumped values and the 5 "outlier" dimensions,
but all 1531 non-outlier dims are in there, which makes for some easy
quantisation from float to byte if you dont want to go the full
kmeans/clustering/Lloyds-algorithm approach):
https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
- the variance of the value of each dimension is characteristic:
https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228

This probably represents something significant about how the ada-002
embeddings are created, but I think it also means creating "realistic"
values is possible. I did not use this information when testing recall &
performance on Lucene's HNSW implementation on 192m documents, as I
slightly dithered the values of a "real" set on 47K docs and stored other
fields in the doc that referenced the "base" document that the dithers were
made from, and used different dithering magnitudes so that I could test
recall with different neighbour sizes ("M"), construction-beamwidth and
search-beamwidths.

best regards

Kent Fitch

On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner <michael.wechner@wyona.com>
wrote:

> I understand what you mean that it seems to be artificial, but I don't
> understand why this matters to test performance and scalability of the
> indexing?
>
> Let's assume the limit of Lucene would be 4 instead of 1024 and there
> are only open source models generating vectors with 4 dimensions, for
> example
>
>
> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>
>
> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>
>
> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>
>
> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>
> and now I concatenate them to vectors with 8 dimensions
>
>
>
> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>
>
> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>
> and normalize them to length 1.
>
> Why should this be any different to a model which is acting like a black
> box generating vectors with 8 dimensions?
>
>
>
>
> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
> >> What exactly do you consider real vector data? Vector data which is
> based on texts written by humans?
> > We have plenty of text; the problem is coming up with a realistic
> > vector model that requires as many dimensions as people seem to be
> > demanding. As I said above, after surveying huggingface I couldn't
> > find any text-based model using more than 768 dimensions. So far we
> > have some ideas of generating higher-dimensional data by dithering or
> > concatenating existing data, but it seems artificial.
> >
> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
> > <michael.wechner@wyona.com> wrote:
> >> What exactly do you consider real vector data? Vector data which is
> based on texts written by humans?
> >>
> >> I am asking, because I recently attended the following presentation by
> Anastassia Shaitarova (UZH Institute for Computational Linguistics,
> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
> >>
> >> ----
> >>
> >> Can we Identify Machine-Generated Text? An Overview of Current
> Approaches
> >> by Anastassia Shaitarova (UZH Institute for Computational Linguistics)
> >>
> >> The detection of machine-generated text has become increasingly
> important due to the prevalence of automated content generation and its
> potential for misuse. In this talk, we will discuss the motivation for
> automatic detection of generated text. We will present the currently
> available methods, including feature-based classification as a “first
> line-of-defense.” We will provide an overview of the detection tools that
> have been made available so far and discuss their limitations. Finally, we
> will reflect on some open problems associated with the automatic
> discrimination of generated texts.
> >>
> >> ----
> >>
> >> and her conclusion was that it has become basically impossible to
> differentiate between text generated by humans and text generated by for
> example ChatGPT.
> >>
> >> Whereas others have a slightly different opinion, see for example
> >>
> >> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
> >>
> >> But I would argue that real world and synthetic have become close
> enough that testing performance and scalability of indexing should be
> possible with synthetic data.
> >>
> >> I completely agree that we have to base our discussions and decisions
> on scientific methods and that we have to make sure that Lucene performs
> and scales well and that we understand the limits and what is going on
> under the hood.
> >>
> >> Thanks
> >>
> >> Michael W
> >>
> >>
> >>
> >>
> >>
> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
> >>
> >> +1 to test on real vector data -- if you test on synthetic data you
> draw synthetic conclusions.
> >>
> >> Can someone post the theoretical performance (CPU and RAM required) of
> HNSW construction? Do we know/believe our HNSW implementation has achieved
> that theoretical big-O performance? Maybe we have some silly performance
> bug that's causing it not to?
> >>
> >> As I understand it, HNSW makes the tradeoff of costly construction for
> faster searching, which is typically the right tradeoff for search use
> cases. We do this in other parts of the Lucene index too.
> >>
> >> Lucene will do a logarithmic number of merges over time, i.e. each doc
> will be merged O(log(N)) times in its lifetime in the index. We need to
> multiply that by the cost of re-building the whole HNSW graph on each
> merge. BTW, other things in Lucene, like BKD/dimensional points, also
> rebuild the whole data structure on each merge, I think? But, as Rob
> pointed out, stored fields merging do indeed do some sneaky tricks to avoid
> excessive block decompress/recompress on each merge.
> >>
> >>> As I understand it, vetoes must have technical merit. I'm not sure
> that this veto rises to "technical merit" on 2 counts:
> >> Actually I think Robert's veto stands on its technical merit already.
> Robert's take on technical matters very much resonate with me, even if he
> is sometimes prickly in how he expresses them ;)
> >>
> >> His point is that we, as a dev community, are not paying enough
> attention to the indexing performance of our KNN algo (HNSW) and
> implementation, and that it is reckless to increase / remove limits in that
> state. It is indeed a one-way door decision and one must confront such
> decisions with caution, especially for such a widely used base
> infrastructure as Lucene. We don't even advertise today in our javadocs
> that you need XXX heap if you index vectors with dimension Y, fanout X,
> levels Z, etc.
> >>
> >> RAM used during merging is unaffected by dimensionality, but is
> affected by fanout, because the HNSW graph (not the raw vectors) is memory
> resident, I think? Maybe we could move it off-heap and let the OS manage
> the memory (and still document the RAM requirements)? Maybe merge RAM
> costs should be accounted for in IW's RAM buffer accounting? It is not
> today, and there are some other things that use non-trivial RAM, e.g. the
> doc mapping (to compress docid space when deletions are reclaimed).
> >>
> >> When we added KNN vector testing to Lucene's nightly benchmarks, the
> indexing time massively increased -- see annotations DH and DP here:
> https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly
> benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of
> course, that is using a single thread for indexing (on a box that has 128
> cores!) so we produce a deterministic index every night ...
> >>
> >> Stepping out (meta) a bit ... this discussion is precisely one of the
> awesome benefits of the (informed) veto. It means risky changes to the
> software, as determined by any single informed developer on the project,
> can force a healthy discussion about the problem at hand. Robert is
> legitimately concerned about a real issue and so we should use our creative
> energies to characterize our HNSW implementation's performance, document it
> clearly for users, and uncover ways to improve it.
> >>
> >> Mike McCandless
> >>
> >> http://blog.mikemccandless.com
> >>
> >>
> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <
> a.benedetti@sease.io> wrote:
> >>> I think Gus points are on target.
> >>>
> >>> I recommend we move this forward in this way:
> >>> We stop any discussion and everyone interested proposes an option with
> a motivation, then we aggregate the options and we create a Vote maybe?
> >>>
> >>> I am also on the same page on the fact that a veto should come with a
> clear and reasonable technical merit, which also in my opinion has not come
> yet.
> >>>
> >>> I also apologise if any of my words sounded harsh or personal attacks,
> never meant to do so.
> >>>
> >>> My proposed option:
> >>>
> >>> 1) remove the limit and potentially make it configurable,
> >>> Motivation:
> >>> The system administrator can enforce a limit its users need to respect
> that it's in line with whatever the admin decided to be acceptable for them.
> >>> Default can stay the current one.
> >>>
> >>> That's my favourite at the moment, but I agree that potentially in the
> future this may need to change, as we may optimise the data structures for
> certain dimensions. I am a big fan of Yagni (you aren't going to need it)
> so I am ok we'll face a different discussion if that happens in the future.
> >>>
> >>>
> >>>
> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
> >>>> What I see so far:
> >>>>
> >>>> Much positive support for raising the limit
> >>>> Slightly less support for removing it or making it configurable
> >>>> A single veto which argues that a (as yet undefined) performance
> standard must be met before raising the limit
> >>>> Hot tempers (various) making this discussion difficult
> >>>>
> >>>> As I understand it, vetoes must have technical merit. I'm not sure
> that this veto rises to "technical merit" on 2 counts:
> >>>>
> >>>> No standard for the performance is given so it cannot be technically
> met. Without hard criteria it's a moving target.
> >>>> It appears to encode a valuation of the user's time, and that
> valuation is really up to the user. Some users may consider 2hours useless
> and not worth it, and others might happily wait 2 hours. This is not a
> technical decision, it's a business decision regarding the relative value
> of the time invested vs the value of the result. If I can cure cancer by
> indexing for a year, that might be worth it... (hyperbole of course).
> >>>>
> >>>> Things I would consider to have technical merit that I don't hear:
> >>>>
> >>>> Impact on the speed of **other** indexing operations. (devaluation of
> other functionality)
> >>>> Actual scenarios that work when the limit is low and fail when the
> limit is high (new failure on the same data with the limit raised).
> >>>>
> >>>> One thing that might or might not have technical merit
> >>>>
> >>>> If someone feels there is a lack of documentation of the
> costs/performance implications of using large vectors, possibly including
> reproducible benchmarks establishing the scaling behavior (there seems to
> be disagreement on O(n) vs O(n^2)).
> >>>>
> >>>> The users *should* know what they are getting into, but if the cost
> is worth it to them, they should be able to pay it without forking the
> project. If this veto causes a fork that's not good.
> >>>>
> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com>
> wrote:
> >>>>> We do have a dataset built from Wikipedia in luceneutil. It comes in
> 100 and 300 dimensional varieties and can easily enough generate large
> numbers of vector documents from the articles data. To go higher we could
> concatenate vectors from that and I believe the performance numbers would
> be plausible.
> >>>>>
> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com>
> wrote:
> >>>>>> Can we set up a branch in which the limit is bumped to 2048, then
> have
> >>>>>> a realistic, free data set (wikipedia sample or something) that has,
> >>>>>> say, 5 million docs and vectors created using public data (glove
> >>>>>> pre-trained embeddings or the like)? We then could run indexing on
> the
> >>>>>> same hardware with 512, 1024 and 2048 and see what the numbers,
> limits
> >>>>>> and behavior actually are.
> >>>>>>
> >>>>>> I can help in writing this but not until after Easter.
> >>>>>>
> >>>>>>
> >>>>>> Dawid
> >>>>>>
> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com>
> wrote:
> >>>>>>> As Dawid pointed out earlier on this thread, this is the rule for
> >>>>>>> Apache projects: a single -1 vote on a code change is a veto and
> >>>>>>> cannot be overridden. Furthermore, Robert is one of the people on
> this
> >>>>>>> project who worked the most on debugging subtle bugs, making Lucene
> >>>>>>> more robust and improving our test framework, so I'm listening
> when he
> >>>>>>> voices quality concerns.
> >>>>>>>
> >>>>>>> The argument against removing/raising the limit that resonates
> with me
> >>>>>>> the most is that it is a one-way door. As MikeS highlighted
> earlier on
> >>>>>>> this thread, implementations may want to take advantage of the fact
> >>>>>>> that there is a limit at some point too. This is why I don't want
> to
> >>>>>>> remove the limit and would prefer a slight increase, such as 2048
> as
> >>>>>>> suggested in the original issue, which would enable most of the
> things
> >>>>>>> that users who have been asking about raising the limit would like
> to
> >>>>>>> do.
> >>>>>>>
> >>>>>>> I agree that the merge-time memory usage and slow indexing rate are
> >>>>>>> not great. But it's still possible to index multi-million vector
> >>>>>>> datasets with a 4GB heap without hitting OOMEs regardless of the
> >>>>>>> number of dimensions, and the feedback I'm seeing is that many
> users
> >>>>>>> are still interested in indexing multi-million vector datasets
> despite
> >>>>>>> the slow indexing rate. I wish we could do better, and vector
> indexing
> >>>>>>> is certainly more expert than text indexing, but it still is
> usable in
> >>>>>>> my opinion. I understand how giving Lucene more information about
> >>>>>>> vectors prior to indexing (e.g. clustering information as Jim
> pointed
> >>>>>>> out) could help make merging faster and more memory-efficient, but
> I
> >>>>>>> would really like to avoid making it a requirement for indexing
> >>>>>>> vectors as it also makes this feature much harder to use.
> >>>>>>>
> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
> >>>>>>> <a.benedetti@sease.io> wrote:
> >>>>>>>> I am very attentive to listen opinions but I am un-convinced here
> and I an not sure that a single person opinion should be allowed to be
> detrimental for such an important project.
> >>>>>>>>
> >>>>>>>> The limit as far as I know is literally just raising an exception.
> >>>>>>>> Removing it won't alter in any way the current performance for
> users in low dimensional space.
> >>>>>>>> Removing it will just enable more users to use Lucene.
> >>>>>>>>
> >>>>>>>> If new users in certain situations will be unhappy with the
> performance, they may contribute improvements.
> >>>>>>>> This is how you make progress.
> >>>>>>>>
> >>>>>>>> If it's a reputation thing, trust me that not allowing users to
> play with high dimensional space will equally damage it.
> >>>>>>>>
> >>>>>>>> To me it's really a no brainer.
> >>>>>>>> Removing the limit and enable people to use high dimensional
> vectors will take minutes.
> >>>>>>>> Improving the hnsw implementation can take months.
> >>>>>>>> Pick one to begin with...
> >>>>>>>>
> >>>>>>>> And there's no-one paying me here, no company interest
> whatsoever, actually I pay people to contribute, I am just convinced it's a
> good idea.
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
> >>>>>>>>> I disagree with your categorization. I put in plenty of work and
> >>>>>>>>> experienced plenty of pain myself, writing tests and fighting
> these
> >>>>>>>>> issues, after i saw that, two releases in a row, vector indexing
> fell
> >>>>>>>>> over and hit integer overflows etc on small datasets:
> >>>>>>>>>
> >>>>>>>>> https://github.com/apache/lucene/pull/11905
> >>>>>>>>>
> >>>>>>>>> Attacking me isn't helping the situation.
> >>>>>>>>>
> >>>>>>>>> PS: when i said the "one guy who wrote the code" I didn't mean
> it in
> >>>>>>>>> any kind of demeaning fashion really. I meant to describe the
> current
> >>>>>>>>> state of usability with respect to indexing a few million docs
> with
> >>>>>>>>> high dimensions. You can scroll up the thread and see that at
> least
> >>>>>>>>> one other committer on the project experienced similar pain as
> me.
> >>>>>>>>> Then, think about users who aren't committers trying to use the
> >>>>>>>>> functionality!
> >>>>>>>>>
> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <
> msokolov@gmail.com> wrote:
> >>>>>>>>>> What you said about increasing dimensions requiring a bigger
> ram buffer on merge is wrong. That's the point I was trying to make. Your
> concerns about merge costs are not wrong, but your conclusion that we need
> to limit dimensions is not justified.
> >>>>>>>>>>
> >>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when I show
> it scales linearly with dimension you just ignore that and complain about
> something entirely different.
> >>>>>>>>>>
> >>>>>>>>>> You demand that people run all kinds of tests to prove you
> wrong but when they do, you don't listen and you won't put in the work
> yourself or complain that it's too hard.
> >>>>>>>>>>
> >>>>>>>>>> Then you complain about people not meeting you half way. Wow
> >>>>>>>>>>
> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
> wrote:
> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
> >>>>>>>>>>>> What exactly do you consider reasonable?
> >>>>>>>>>>> Let's begin a real discussion by being HONEST about the current
> >>>>>>>>>>> status. Please put politically correct or your own company's
> wishes
> >>>>>>>>>>> aside, we know it's not in a good state.
> >>>>>>>>>>>
> >>>>>>>>>>> Current status is the one guy who wrote the code can set a
> >>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset with 1024
> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
> >>>>>>>>>>>
> >>>>>>>>>>> My concerns are everyone else except the one guy, I want it to
> be
> >>>>>>>>>>> usable. Increasing dimensions just means even bigger
> multi-gigabyte
> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
> >>>>>>>>>>> It is also a permanent backwards compatibility decision, we
> have to
> >>>>>>>>>>> support it once we do this and we can't just say "oops" and
> flip it
> >>>>>>>>>>> back.
> >>>>>>>>>>>
> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer is
> really to
> >>>>>>>>>>> avoid merges because they are so slow and it would be DAYS
> otherwise,
> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
> >>>>>>>>>>> Also from personal experience, it takes trial and error (means
> >>>>>>>>>>> experiencing OOM on merge!!!) before you get those heap values
> correct
> >>>>>>>>>>> for your dataset. This usually means starting over which is
> >>>>>>>>>>> frustrating and wastes more time.
> >>>>>>>>>>>
> >>>>>>>>>>> Jim mentioned some ideas about the memory usage in
> IndexWriter, seems
> >>>>>>>>>>> to me like its a good idea. maybe the multigigabyte ram buffer
> can be
> >>>>>>>>>>> avoided in this way and performance improved by writing bigger
> >>>>>>>>>>> segments with lucene's defaults. But this doesn't mean we can
> simply
> >>>>>>>>>>> ignore the horrors of what happens on merge. merging needs to
> scale so
> >>>>>>>>>>> that indexing really scales.
> >>>>>>>>>>>
> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts and
> cause OOM,
> >>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU in
> O(n^2)
> >>>>>>>>>>> fashion when indexing.
> >>>>>>>>>>>
> >>>>>>>>>>>
> ---------------------------------------------------------------------
> >>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
> >>>>>>>>>>>
> >>>>>>>>>
> ---------------------------------------------------------------------
> >>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
> >>>>>>>>>
> >>>>>>>
> >>>>>>> --
> >>>>>>> Adrien
> >>>>>>>
> >>>>>>>
> ---------------------------------------------------------------------
> >>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
> >>>>>>>
> >>>>>>
> ---------------------------------------------------------------------
> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
> >>>>>>
> >>>>
> >>>> --
> >>>> http://www.needhamsoftware.com (work)
> >>>> http://www.the111shift.com (play)
> >>
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> > For additional commands, e-mail: dev-help@lucene.apache.org
> >
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

gus.heck at gmail

Apr 11, 2023, 10:41 PM

Post #90 of 99 (942 views)

His point is that we, as a dev community, are not paying enough attention
> to the indexing performance of our KNN algo (HNSW) and implementation, and
> that it is reckless to increase / remove limits in that state.
>

If the argument were... "Please hold off while I'm actively improving this,
it will be ready soon and then we can adjust the limit" that might have
technical merit. As it was presented it came across more like "I'm going to
hold this feature lots of folk want hostage until *someone else* does
something I think should be done"... I doubt that was actually what he
consciously thought (I don't think anyone on this project would have that
specific intention), but the context and manner have made it seem that way,
and the net effect seems to be trending in that direction.

If there's a way that raising the limit *prevents* working on performance
that of course would be a key thing to understand. It seems to me that the
exact person who's going to go on a performance crusade is the person who
has a technique that they can prove works, but it's just too darn slow....
Maybe not the first person, maybe not the fifth, but it's going to be
*someone* who needs it...

100% the user should know that they are "off the edge of the map" and "here
there be monsters." Document it well, issue a warning, whatever. Once
they've been told, and they set sail for the unknown, let them develop an
itch so that they can scratch it.

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 11, 2023, 11:36 PM

Post #91 of 99 (942 views)

thank you very much for your feedback!

In a previous post (April 7) you wrote you could make availlable the 47K
ada-002 vectors, which would be great!

Would it make sense to setup a public gitub repo, such that others could
use or also contribute vectors?

Thanks

Michael Wechner

Am 12.04.23 um 04:51 schrieb Kent Fitch:
> I only know some characteristics of the openAI ada-002 vectors,
> although they are a very popular as embeddings/text-characterisations
> as they allow more accurate/"human meaningful" semantic search results
> with fewer dimensions than their predecessors - I've evaluated a few
> different embedding models, including some BERT variants, CLIP
> ViT-L-14 (with 768 dims, which was quite good), openAI's ada-001 (1024
> dims) and babbage-001 (2048 dims), and ada-002 are qualitatively the
> best, although that will certainly change!
>
> In any case, ada-002 vectors have interesting characteristics that I
> think mean you could confidently create synthetic vectors which
> would be hard to distinguish from "real" vectors. I found this from
> looking at 47K ada-002 vectors generated across a full year (1994) of
> newspaper articles from the Canberra Times and 200K wikipedia articles:
> - there is no discernible/significant correlation between values in
> any pair of dimensions
> - all but 5 of the 1536 dimensions have an almost identical
> distribution of values shown in the central blob on these graphs (that
> just show a few of these 1531 dimensions with clumped values and the 5
> "outlier" dimensions, but all 1531 non-outlier dims are in there,
> which makes for some easy quantisation from float to byte if you dont
> want to go the full kmeans/clustering/Lloyds-algorithm approach):
> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
> - the variance of the value of each dimension is characteristic:
> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>
> This probably represents something significant about how the ada-002
> embeddings are created, but I think it also means creating "realistic"
> values is possible. I did not use this information when testing
> recall & performance on Lucene's HNSW implementation on 192m
> documents, as I slightly dithered the values of a "real" set on 47K
> docs and stored other fields in the doc that referenced the "base"
> document that the dithers were made from, and used different dithering
> magnitudes so that I could test recall with different neighbour sizes
> ("M"), construction-beamwidth and search-beamwidths.
>
> best regards
>
> Kent Fitch
>
>
>
>
> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner
> <michael.wechner@wyona.com> wrote:
>
> I understand what you mean that it seems to be artificial, but I
> don't
> understand why this matters to test performance and scalability of
> the
> indexing?
>
> Let's assume the limit of Lucene would be 4 instead of 1024 and there
> are only open source models generating vectors with 4 dimensions, for
> example
>
> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>
> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>
> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>
> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>
> and now I concatenate them to vectors with 8 dimensions
>
>
> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>
> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>
> and normalize them to length 1.
>
> Why should this be any different to a model which is acting like a
> black
> box generating vectors with 8 dimensions?
>
>
>
>
> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
> >> What exactly do you consider real vector data? Vector data
> which is based on texts written by humans?
> > We have plenty of text; the problem is coming up with a realistic
> > vector model that requires as many dimensions as people seem to be
> > demanding. As I said above, after surveying huggingface I couldn't
> > find any text-based model using more than 768 dimensions. So far we
> > have some ideas of generating higher-dimensional data by
> dithering or
> > concatenating existing data, but it seems artificial.
> >
> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
> > <michael.wechner@wyona.com> wrote:
> >> What exactly do you consider real vector data? Vector data
> which is based on texts written by humans?
> >>
> >> I am asking, because I recently attended the following
> presentation by Anastassia Shaitarova (UZH Institute for
> Computational Linguistics,
> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
> >>
> >> ----
> >>
> >> Can we Identify Machine-Generated Text? An Overview of Current
> Approaches
> >> by Anastassia Shaitarova (UZH Institute for Computational
> Linguistics)
> >>
> >> The detection of machine-generated text has become increasingly
> important due to the prevalence of automated content generation
> and its potential for misuse. In this talk, we will discuss the
> motivation for automatic detection of generated text. We will
> present the currently available methods, including feature-based
> classification as a “first line-of-defense.” We will provide an
> overview of the detection tools that have been made available so
> far and discuss their limitations. Finally, we will reflect on
> some open problems associated with the automatic discrimination of
> generated texts.
> >>
> >> ----
> >>
> >> and her conclusion was that it has become basically impossible
> to differentiate between text generated by humans and text
> generated by for example ChatGPT.
> >>
> >> Whereas others have a slightly different opinion, see for example
> >>
> >> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
> >>
> >> But I would argue that real world and synthetic have become
> close enough that testing performance and scalability of indexing
> should be possible with synthetic data.
> >>
> >> I completely agree that we have to base our discussions and
> decisions on scientific methods and that we have to make sure that
> Lucene performs and scales well and that we understand the limits
> and what is going on under the hood.
> >>
> >> Thanks
> >>
> >> Michael W
> >>
> >>
> >>
> >>
> >>
> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
> >>
> >> +1 to test on real vector data -- if you test on synthetic data
> you draw synthetic conclusions.
> >>
> >> Can someone post the theoretical performance (CPU and RAM
> required) of HNSW construction? Do we know/believe our HNSW
> implementation has achieved that theoretical big-O performance?
> Maybe we have some silly performance bug that's causing it not to?
> >>
> >> As I understand it, HNSW makes the tradeoff of costly
> construction for faster searching, which is typically the right
> tradeoff for search use cases. We do this in other parts of the
> Lucene index too.
> >>
> >> Lucene will do a logarithmic number of merges over time, i.e.
> each doc will be merged O(log(N)) times in its lifetime in the
> index. We need to multiply that by the cost of re-building the
> whole HNSW graph on each merge. BTW, other things in Lucene, like
> BKD/dimensional points, also rebuild the whole data structure on
> each merge, I think? But, as Rob pointed out, stored fields
> merging do indeed do some sneaky tricks to avoid excessive block
> decompress/recompress on each merge.
> >>
> >>> As I understand it, vetoes must have technical merit. I'm not
> sure that this veto rises to "technical merit" on 2 counts:
> >> Actually I think Robert's veto stands on its technical merit
> already. Robert's take on technical matters very much resonate
> with me, even if he is sometimes prickly in how he expresses them ;)
> >>
> >> His point is that we, as a dev community, are not paying enough
> attention to the indexing performance of our KNN algo (HNSW) and
> implementation, and that it is reckless to increase / remove
> limits in that state. It is indeed a one-way door decision and
> one must confront such decisions with caution, especially for such
> a widely used base infrastructure as Lucene. We don't even
> advertise today in our javadocs that you need XXX heap if you
> index vectors with dimension Y, fanout X, levels Z, etc.
> >>
> >> RAM used during merging is unaffected by dimensionality, but is
> affected by fanout, because the HNSW graph (not the raw vectors)
> is memory resident, I think? Maybe we could move it off-heap and
> let the OS manage the memory (and still document the RAM
> requirements)? Maybe merge RAM costs should be accounted for in
> IW's RAM buffer accounting? It is not today, and there are some
> other things that use non-trivial RAM, e.g. the doc mapping (to
> compress docid space when deletions are reclaimed).
> >>
> >> When we added KNN vector testing to Lucene's nightly
> benchmarks, the indexing time massively increased -- see
> annotations DH and DP here:
> https://home.apache.org/~mikemccand/lucenebench/indexing.html.
> Nightly benchmarks now start at 6 PM and don't finish until ~14.5
> hours later. Of course, that is using a single thread for
> indexing (on a box that has 128 cores!) so we produce a
> deterministic index every night ...
> >>
> >> Stepping out (meta) a bit ... this discussion is precisely one
> of the awesome benefits of the (informed) veto. It means risky
> changes to the software, as determined by any single informed
> developer on the project, can force a healthy discussion about the
> problem at hand. Robert is legitimately concerned about a real
> issue and so we should use our creative energies to characterize
> our HNSW implementation's performance, document it clearly for
> users, and uncover ways to improve it.
> >>
> >> Mike McCandless
> >>
> >> http://blog.mikemccandless.com
> >>
> >>
> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti
> <a.benedetti@sease.io> wrote:
> >>> I think Gus points are on target.
> >>>
> >>> I recommend we move this forward in this way:
> >>> We stop any discussion and everyone interested proposes an
> option with a motivation, then we aggregate the options and we
> create a Vote maybe?
> >>>
> >>> I am also on the same page on the fact that a veto should come
> with a clear and reasonable technical merit, which also in my
> opinion has not come yet.
> >>>
> >>> I also apologise if any of my words sounded harsh or personal
> attacks, never meant to do so.
> >>>
> >>> My proposed option:
> >>>
> >>> 1) remove the limit and potentially make it configurable,
> >>> Motivation:
> >>> The system administrator can enforce a limit its users need to
> respect that it's in line with whatever the admin decided to be
> acceptable for them.
> >>> Default can stay the current one.
> >>>
> >>> That's my favourite at the moment, but I agree that
> potentially in the future this may need to change, as we may
> optimise the data structures for certain dimensions. I am a big
> fan of Yagni (you aren't going to need it) so I am ok we'll face a
> different discussion if that happens in the future.
> >>>
> >>>
> >>>
> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
> >>>> What I see so far:
> >>>>
> >>>> Much positive support for raising the limit
> >>>> Slightly less support for removing it or making it configurable
> >>>> A single veto which argues that a (as yet undefined)
> performance standard must be met before raising the limit
> >>>> Hot tempers (various) making this discussion difficult
> >>>>
> >>>> As I understand it, vetoes must have technical merit. I'm not
> sure that this veto rises to "technical merit" on 2 counts:
> >>>>
> >>>> No standard for the performance is given so it cannot be
> technically met. Without hard criteria it's a moving target.
> >>>> It appears to encode a valuation of the user's time, and that
> valuation is really up to the user. Some users may consider 2hours
> useless and not worth it, and others might happily wait 2 hours.
> This is not a technical decision, it's a business decision
> regarding the relative value of the time invested vs the value of
> the result. If I can cure cancer by indexing for a year, that
> might be worth it... (hyperbole of course).
> >>>>
> >>>> Things I would consider to have technical merit that I don't
> hear:
> >>>>
> >>>> Impact on the speed of **other** indexing operations.
> (devaluation of other functionality)
> >>>> Actual scenarios that work when the limit is low and fail
> when the limit is high (new failure on the same data with the
> limit raised).
> >>>>
> >>>> One thing that might or might not have technical merit
> >>>>
> >>>> If someone feels there is a lack of documentation of the
> costs/performance implications of using large vectors, possibly
> including reproducible benchmarks establishing the scaling
> behavior (there seems to be disagreement on O(n) vs O(n^2)).
> >>>>
> >>>> The users *should* know what they are getting into, but if
> the cost is worth it to them, they should be able to pay it
> without forking the project. If this veto causes a fork that's not
> good.
> >>>>
> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov
> <msokolov@gmail.com> wrote:
> >>>>> We do have a dataset built from Wikipedia in luceneutil. It
> comes in 100 and 300 dimensional varieties and can easily enough
> generate large numbers of vector documents from the articles data.
> To go higher we could concatenate vectors from that and I believe
> the performance numbers would be plausible.
> >>>>>
> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss
> <dawid.weiss@gmail.com> wrote:
> >>>>>> Can we set up a branch in which the limit is bumped to
> 2048, then have
> >>>>>> a realistic, free data set (wikipedia sample or something)
> that has,
> >>>>>> say, 5 million docs and vectors created using public data
> (glove
> >>>>>> pre-trained embeddings or the like)? We then could run
> indexing on the
> >>>>>> same hardware with 512, 1024 and 2048 and see what the
> numbers, limits
> >>>>>> and behavior actually are.
> >>>>>>
> >>>>>> I can help in writing this but not until after Easter.
> >>>>>>
> >>>>>>
> >>>>>> Dawid
> >>>>>>
> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand
> <jpountz@gmail.com> wrote:
> >>>>>>> As Dawid pointed out earlier on this thread, this is the
> rule for
> >>>>>>> Apache projects: a single -1 vote on a code change is a
> veto and
> >>>>>>> cannot be overridden. Furthermore, Robert is one of the
> people on this
> >>>>>>> project who worked the most on debugging subtle bugs,
> making Lucene
> >>>>>>> more robust and improving our test framework, so I'm
> listening when he
> >>>>>>> voices quality concerns.
> >>>>>>>
> >>>>>>> The argument against removing/raising the limit that
> resonates with me
> >>>>>>> the most is that it is a one-way door. As MikeS
> highlighted earlier on
> >>>>>>> this thread, implementations may want to take advantage of
> the fact
> >>>>>>> that there is a limit at some point too. This is why I
> don't want to
> >>>>>>> remove the limit and would prefer a slight increase, such
> as 2048 as
> >>>>>>> suggested in the original issue, which would enable most
> of the things
> >>>>>>> that users who have been asking about raising the limit
> would like to
> >>>>>>> do.
> >>>>>>>
> >>>>>>> I agree that the merge-time memory usage and slow indexing
> rate are
> >>>>>>> not great. But it's still possible to index multi-million
> vector
> >>>>>>> datasets with a 4GB heap without hitting OOMEs regardless
> of the
> >>>>>>> number of dimensions, and the feedback I'm seeing is that
> many users
> >>>>>>> are still interested in indexing multi-million vector
> datasets despite
> >>>>>>> the slow indexing rate. I wish we could do better, and
> vector indexing
> >>>>>>> is certainly more expert than text indexing, but it still
> is usable in
> >>>>>>> my opinion. I understand how giving Lucene more
> information about
> >>>>>>> vectors prior to indexing (e.g. clustering information as
> Jim pointed
> >>>>>>> out) could help make merging faster and more
> memory-efficient, but I
> >>>>>>> would really like to avoid making it a requirement for
> indexing
> >>>>>>> vectors as it also makes this feature much harder to use.
> >>>>>>>
> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
> >>>>>>> <a.benedetti@sease.io> wrote:
> >>>>>>>> I am very attentive to listen opinions but I am
> un-convinced here and I an not sure that a single person opinion
> should be allowed to be detrimental for such an important project.
> >>>>>>>>
> >>>>>>>> The limit as far as I know is literally just raising an
> exception.
> >>>>>>>> Removing it won't alter in any way the current
> performance for users in low dimensional space.
> >>>>>>>> Removing it will just enable more users to use Lucene.
> >>>>>>>>
> >>>>>>>> If new users in certain situations will be unhappy with
> the performance, they may contribute improvements.
> >>>>>>>> This is how you make progress.
> >>>>>>>>
> >>>>>>>> If it's a reputation thing, trust me that not allowing
> users to play with high dimensional space will equally damage it.
> >>>>>>>>
> >>>>>>>> To me it's really a no brainer.
> >>>>>>>> Removing the limit and enable people to use high
> dimensional vectors will take minutes.
> >>>>>>>> Improving the hnsw implementation can take months.
> >>>>>>>> Pick one to begin with...
> >>>>>>>>
> >>>>>>>> And there's no-one paying me here, no company interest
> whatsoever, actually I pay people to contribute, I am just
> convinced it's a good idea.
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com>
> wrote:
> >>>>>>>>> I disagree with your categorization. I put in plenty of
> work and
> >>>>>>>>> experienced plenty of pain myself, writing tests and
> fighting these
> >>>>>>>>> issues, after i saw that, two releases in a row, vector
> indexing fell
> >>>>>>>>> over and hit integer overflows etc on small datasets:
> >>>>>>>>>
> >>>>>>>>> https://github.com/apache/lucene/pull/11905
> >>>>>>>>>
> >>>>>>>>> Attacking me isn't helping the situation.
> >>>>>>>>>
> >>>>>>>>> PS: when i said the "one guy who wrote the code" I
> didn't mean it in
> >>>>>>>>> any kind of demeaning fashion really. I meant to
> describe the current
> >>>>>>>>> state of usability with respect to indexing a few
> million docs with
> >>>>>>>>> high dimensions. You can scroll up the thread and see
> that at least
> >>>>>>>>> one other committer on the project experienced similar
> pain as me.
> >>>>>>>>> Then, think about users who aren't committers trying to
> use the
> >>>>>>>>> functionality!
> >>>>>>>>>
> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov
> <msokolov@gmail.com> wrote:
> >>>>>>>>>> What you said about increasing dimensions requiring a
> bigger ram buffer on merge is wrong. That's the point I was trying
> to make. Your concerns about merge costs are not wrong, but your
> conclusion that we need to limit dimensions is not justified.
> >>>>>>>>>>
> >>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when
> I show it scales linearly with dimension you just ignore that and
> complain about something entirely different.
> >>>>>>>>>>
> >>>>>>>>>> You demand that people run all kinds of tests to prove
> you wrong but when they do, you don't listen and you won't put in
> the work yourself or complain that it's too hard.
> >>>>>>>>>>
> >>>>>>>>>> Then you complain about people not meeting you half
> way. Wow
> >>>>>>>>>>
> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir
> <rcmuir@gmail.com> wrote:
> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
> >>>>>>>>>>>> What exactly do you consider reasonable?
> >>>>>>>>>>> Let's begin a real discussion by being HONEST about
> the current
> >>>>>>>>>>> status. Please put politically correct or your own
> company's wishes
> >>>>>>>>>>> aside, we know it's not in a good state.
> >>>>>>>>>>>
> >>>>>>>>>>> Current status is the one guy who wrote the code can set a
> >>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset
> with 1024
> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
> >>>>>>>>>>>
> >>>>>>>>>>> My concerns are everyone else except the one guy, I
> want it to be
> >>>>>>>>>>> usable. Increasing dimensions just means even bigger
> multi-gigabyte
> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
> >>>>>>>>>>> It is also a permanent backwards compatibility
> decision, we have to
> >>>>>>>>>>> support it once we do this and we can't just say
> "oops" and flip it
> >>>>>>>>>>> back.
> >>>>>>>>>>>
> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer
> is really to
> >>>>>>>>>>> avoid merges because they are so slow and it would be
> DAYS otherwise,
> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
> >>>>>>>>>>> Also from personal experience, it takes trial and
> error (means
> >>>>>>>>>>> experiencing OOM on merge!!!) before you get those
> heap values correct
> >>>>>>>>>>> for your dataset. This usually means starting over
> which is
> >>>>>>>>>>> frustrating and wastes more time.
> >>>>>>>>>>>
> >>>>>>>>>>> Jim mentioned some ideas about the memory usage in
> IndexWriter, seems
> >>>>>>>>>>> to me like its a good idea. maybe the multigigabyte
> ram buffer can be
> >>>>>>>>>>> avoided in this way and performance improved by
> writing bigger
> >>>>>>>>>>> segments with lucene's defaults. But this doesn't mean
> we can simply
> >>>>>>>>>>> ignore the horrors of what happens on merge. merging
> needs to scale so
> >>>>>>>>>>> that indexing really scales.
> >>>>>>>>>>>
> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts
> and cause OOM,
> >>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU
> in O(n^2)
> >>>>>>>>>>> fashion when indexing.
> >>>>>>>>>>>
> >>>>>>>>>>>
> ---------------------------------------------------------------------
> >>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>>>>>>> For additional commands, e-mail:
> dev-help@lucene.apache.org
> >>>>>>>>>>>
> >>>>>>>>>
> ---------------------------------------------------------------------
> >>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
> >>>>>>>>>
> >>>>>>>
> >>>>>>> --
> >>>>>>> Adrien
> >>>>>>>
> >>>>>>>
> ---------------------------------------------------------------------
> >>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
> >>>>>>>
> >>>>>>
> ---------------------------------------------------------------------
> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> >>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
> >>>>>>
> >>>>
> >>>> --
> >>>> http://www.needhamsoftware.com (work)
> >>>> http://www.the111shift.com (play)
> >>
> >
> ---------------------------------------------------------------------
> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> > For additional commands, e-mail: dev-help@lucene.apache.org
> >
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

a.benedetti at sease

Apr 12, 2023, 5:09 AM

Post #92 of 99 (942 views)

My tentative of listing here only a set of proposals to then vote, has
unfortunately failed.

I appreciate the discussion on better benchmarking hnsw but my feeling is
that this discussion is orthogonal to the limit discussion itself, should
we create a separate mail thread/github jira issue for that?

At the moment I see at least three lines of activities as an outcome from
this (maybe too long) discussion:

1) [small task] there's a need from a good amount of people of
increasing/removing the max limit, as an enabler, to get more users to
Lucene and ease adoption for systems Lucene based (Apache Solr,
Elasticsearch, OpenSearch)

2) [medium task] we all want more benchmarks for Lucene vector-based
search, with a good variety of vector dimensions and encodings

3) [big task? ] some people would like to improve vector based search
peformance because currently not acceptable, it's not clear when and how

A question I have for point 1, does it really need to be a one way door?
Can't we reduce the max limit in the future if the implementation becomes
coupled with certain dimension sizes?
It's not ideal I agree, but is back-compatibility more important than
pragmatic benefits?
I. E.
Right now there's no implementation coupled with the max limit - > we
remove/increase the limit and get more Users

With Lucene X.Y a clever committer introduces a super nice implementation
improvements that unfortunately limit the max size to K.
Can't we just document it as a breaking change for such release? So at that
point we won't support >K vectors but for a reason?

Do we have similar precedents in Lucene?

On Wed, 12 Apr 2023, 08:36 Michael Wechner, <michael.wechner@wyona.com>
wrote:

> thank you very much for your feedback!
>
> In a previous post (April 7) you wrote you could make availlable the 47K
> ada-002 vectors, which would be great!
>
> Would it make sense to setup a public gitub repo, such that others could
> use or also contribute vectors?
>
> Thanks
>
> Michael Wechner
>
>
> Am 12.04.23 um 04:51 schrieb Kent Fitch:
>
> I only know some characteristics of the openAI ada-002 vectors, although
> they are a very popular as embeddings/text-characterisations as they allow
> more accurate/"human meaningful" semantic search results with fewer
> dimensions than their predecessors - I've evaluated a few different
> embedding models, including some BERT variants, CLIP ViT-L-14 (with 768
> dims, which was quite good), openAI's ada-001 (1024 dims) and babbage-001
> (2048 dims), and ada-002 are qualitatively the best, although that will
> certainly change!
>
> In any case, ada-002 vectors have interesting characteristics that I think
> mean you could confidently create synthetic vectors which would be hard to
> distinguish from "real" vectors. I found this from looking at 47K ada-002
> vectors generated across a full year (1994) of newspaper articles from the
> Canberra Times and 200K wikipedia articles:
> - there is no discernible/significant correlation between values in any
> pair of dimensions
> - all but 5 of the 1536 dimensions have an almost identical distribution
> of values shown in the central blob on these graphs (that just show a few
> of these 1531 dimensions with clumped values and the 5 "outlier"
> dimensions, but all 1531 non-outlier dims are in there, which makes for
> some easy quantisation from float to byte if you dont want to go the full
> kmeans/clustering/Lloyds-algorithm approach):
>
> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
>
> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
>
> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
> - the variance of the value of each dimension is characteristic:
>
> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>
> This probably represents something significant about how the ada-002
> embeddings are created, but I think it also means creating "realistic"
> values is possible. I did not use this information when testing recall &
> performance on Lucene's HNSW implementation on 192m documents, as I
> slightly dithered the values of a "real" set on 47K docs and stored other
> fields in the doc that referenced the "base" document that the dithers were
> made from, and used different dithering magnitudes so that I could test
> recall with different neighbour sizes ("M"), construction-beamwidth and
> search-beamwidths.
>
> best regards
>
> Kent Fitch
>
>
>
>
> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner <michael.wechner@wyona.com>
> wrote:
>
>> I understand what you mean that it seems to be artificial, but I don't
>> understand why this matters to test performance and scalability of the
>> indexing?
>>
>> Let's assume the limit of Lucene would be 4 instead of 1024 and there
>> are only open source models generating vectors with 4 dimensions, for
>> example
>>
>>
>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>>
>>
>> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>
>>
>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>>
>>
>> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>
>> and now I concatenate them to vectors with 8 dimensions
>>
>>
>>
>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>
>>
>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>
>> and normalize them to length 1.
>>
>> Why should this be any different to a model which is acting like a black
>> box generating vectors with 8 dimensions?
>>
>>
>>
>>
>> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>> >> What exactly do you consider real vector data? Vector data which is
>> based on texts written by humans?
>> > We have plenty of text; the problem is coming up with a realistic
>> > vector model that requires as many dimensions as people seem to be
>> > demanding. As I said above, after surveying huggingface I couldn't
>> > find any text-based model using more than 768 dimensions. So far we
>> > have some ideas of generating higher-dimensional data by dithering or
>> > concatenating existing data, but it seems artificial.
>> >
>> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
>> > <michael.wechner@wyona.com> wrote:
>> >> What exactly do you consider real vector data? Vector data which is
>> based on texts written by humans?
>> >>
>> >> I am asking, because I recently attended the following presentation by
>> Anastassia Shaitarova (UZH Institute for Computational Linguistics,
>> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>> >>
>> >> ----
>> >>
>> >> Can we Identify Machine-Generated Text? An Overview of Current
>> Approaches
>> >> by Anastassia Shaitarova (UZH Institute for Computational Linguistics)
>> >>
>> >> The detection of machine-generated text has become increasingly
>> important due to the prevalence of automated content generation and its
>> potential for misuse. In this talk, we will discuss the motivation for
>> automatic detection of generated text. We will present the currently
>> available methods, including feature-based classification as a “first
>> line-of-defense.” We will provide an overview of the detection tools that
>> have been made available so far and discuss their limitations. Finally, we
>> will reflect on some open problems associated with the automatic
>> discrimination of generated texts.
>> >>
>> >> ----
>> >>
>> >> and her conclusion was that it has become basically impossible to
>> differentiate between text generated by humans and text generated by for
>> example ChatGPT.
>> >>
>> >> Whereas others have a slightly different opinion, see for example
>> >>
>> >> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>> >>
>> >> But I would argue that real world and synthetic have become close
>> enough that testing performance and scalability of indexing should be
>> possible with synthetic data.
>> >>
>> >> I completely agree that we have to base our discussions and decisions
>> on scientific methods and that we have to make sure that Lucene performs
>> and scales well and that we understand the limits and what is going on
>> under the hood.
>> >>
>> >> Thanks
>> >>
>> >> Michael W
>> >>
>> >>
>> >>
>> >>
>> >>
>> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>> >>
>> >> +1 to test on real vector data -- if you test on synthetic data you
>> draw synthetic conclusions.
>> >>
>> >> Can someone post the theoretical performance (CPU and RAM required) of
>> HNSW construction? Do we know/believe our HNSW implementation has achieved
>> that theoretical big-O performance? Maybe we have some silly performance
>> bug that's causing it not to?
>> >>
>> >> As I understand it, HNSW makes the tradeoff of costly construction for
>> faster searching, which is typically the right tradeoff for search use
>> cases. We do this in other parts of the Lucene index too.
>> >>
>> >> Lucene will do a logarithmic number of merges over time, i.e. each doc
>> will be merged O(log(N)) times in its lifetime in the index. We need to
>> multiply that by the cost of re-building the whole HNSW graph on each
>> merge. BTW, other things in Lucene, like BKD/dimensional points, also
>> rebuild the whole data structure on each merge, I think? But, as Rob
>> pointed out, stored fields merging do indeed do some sneaky tricks to avoid
>> excessive block decompress/recompress on each merge.
>> >>
>> >>> As I understand it, vetoes must have technical merit. I'm not sure
>> that this veto rises to "technical merit" on 2 counts:
>> >> Actually I think Robert's veto stands on its technical merit already.
>> Robert's take on technical matters very much resonate with me, even if he
>> is sometimes prickly in how he expresses them ;)
>> >>
>> >> His point is that we, as a dev community, are not paying enough
>> attention to the indexing performance of our KNN algo (HNSW) and
>> implementation, and that it is reckless to increase / remove limits in that
>> state. It is indeed a one-way door decision and one must confront such
>> decisions with caution, especially for such a widely used base
>> infrastructure as Lucene. We don't even advertise today in our javadocs
>> that you need XXX heap if you index vectors with dimension Y, fanout X,
>> levels Z, etc.
>> >>
>> >> RAM used during merging is unaffected by dimensionality, but is
>> affected by fanout, because the HNSW graph (not the raw vectors) is memory
>> resident, I think? Maybe we could move it off-heap and let the OS manage
>> the memory (and still document the RAM requirements)? Maybe merge RAM
>> costs should be accounted for in IW's RAM buffer accounting? It is not
>> today, and there are some other things that use non-trivial RAM, e.g. the
>> doc mapping (to compress docid space when deletions are reclaimed).
>> >>
>> >> When we added KNN vector testing to Lucene's nightly benchmarks, the
>> indexing time massively increased -- see annotations DH and DP here:
>> https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly
>> benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of
>> course, that is using a single thread for indexing (on a box that has 128
>> cores!) so we produce a deterministic index every night ...
>> >>
>> >> Stepping out (meta) a bit ... this discussion is precisely one of the
>> awesome benefits of the (informed) veto. It means risky changes to the
>> software, as determined by any single informed developer on the project,
>> can force a healthy discussion about the problem at hand. Robert is
>> legitimately concerned about a real issue and so we should use our creative
>> energies to characterize our HNSW implementation's performance, document it
>> clearly for users, and uncover ways to improve it.
>> >>
>> >> Mike McCandless
>> >>
>> >> http://blog.mikemccandless.com
>> >>
>> >>
>> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <
>> a.benedetti@sease.io> wrote:
>> >>> I think Gus points are on target.
>> >>>
>> >>> I recommend we move this forward in this way:
>> >>> We stop any discussion and everyone interested proposes an option
>> with a motivation, then we aggregate the options and we create a Vote maybe?
>> >>>
>> >>> I am also on the same page on the fact that a veto should come with a
>> clear and reasonable technical merit, which also in my opinion has not come
>> yet.
>> >>>
>> >>> I also apologise if any of my words sounded harsh or personal
>> attacks, never meant to do so.
>> >>>
>> >>> My proposed option:
>> >>>
>> >>> 1) remove the limit and potentially make it configurable,
>> >>> Motivation:
>> >>> The system administrator can enforce a limit its users need to
>> respect that it's in line with whatever the admin decided to be acceptable
>> for them.
>> >>> Default can stay the current one.
>> >>>
>> >>> That's my favourite at the moment, but I agree that potentially in
>> the future this may need to change, as we may optimise the data structures
>> for certain dimensions. I am a big fan of Yagni (you aren't going to need
>> it) so I am ok we'll face a different discussion if that happens in the
>> future.
>> >>>
>> >>>
>> >>>
>> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>> >>>> What I see so far:
>> >>>>
>> >>>> Much positive support for raising the limit
>> >>>> Slightly less support for removing it or making it configurable
>> >>>> A single veto which argues that a (as yet undefined) performance
>> standard must be met before raising the limit
>> >>>> Hot tempers (various) making this discussion difficult
>> >>>>
>> >>>> As I understand it, vetoes must have technical merit. I'm not sure
>> that this veto rises to "technical merit" on 2 counts:
>> >>>>
>> >>>> No standard for the performance is given so it cannot be technically
>> met. Without hard criteria it's a moving target.
>> >>>> It appears to encode a valuation of the user's time, and that
>> valuation is really up to the user. Some users may consider 2hours useless
>> and not worth it, and others might happily wait 2 hours. This is not a
>> technical decision, it's a business decision regarding the relative value
>> of the time invested vs the value of the result. If I can cure cancer by
>> indexing for a year, that might be worth it... (hyperbole of course).
>> >>>>
>> >>>> Things I would consider to have technical merit that I don't hear:
>> >>>>
>> >>>> Impact on the speed of **other** indexing operations. (devaluation
>> of other functionality)
>> >>>> Actual scenarios that work when the limit is low and fail when the
>> limit is high (new failure on the same data with the limit raised).
>> >>>>
>> >>>> One thing that might or might not have technical merit
>> >>>>
>> >>>> If someone feels there is a lack of documentation of the
>> costs/performance implications of using large vectors, possibly including
>> reproducible benchmarks establishing the scaling behavior (there seems to
>> be disagreement on O(n) vs O(n^2)).
>> >>>>
>> >>>> The users *should* know what they are getting into, but if the cost
>> is worth it to them, they should be able to pay it without forking the
>> project. If this veto causes a fork that's not good.
>> >>>>
>> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com>
>> wrote:
>> >>>>> We do have a dataset built from Wikipedia in luceneutil. It comes
>> in 100 and 300 dimensional varieties and can easily enough generate large
>> numbers of vector documents from the articles data. To go higher we could
>> concatenate vectors from that and I believe the performance numbers would
>> be plausible.
>> >>>>>
>> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com>
>> wrote:
>> >>>>>> Can we set up a branch in which the limit is bumped to 2048, then
>> have
>> >>>>>> a realistic, free data set (wikipedia sample or something) that
>> has,
>> >>>>>> say, 5 million docs and vectors created using public data (glove
>> >>>>>> pre-trained embeddings or the like)? We then could run indexing on
>> the
>> >>>>>> same hardware with 512, 1024 and 2048 and see what the numbers,
>> limits
>> >>>>>> and behavior actually are.
>> >>>>>>
>> >>>>>> I can help in writing this but not until after Easter.
>> >>>>>>
>> >>>>>>
>> >>>>>> Dawid
>> >>>>>>
>> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com>
>> wrote:
>> >>>>>>> As Dawid pointed out earlier on this thread, this is the rule for
>> >>>>>>> Apache projects: a single -1 vote on a code change is a veto and
>> >>>>>>> cannot be overridden. Furthermore, Robert is one of the people on
>> this
>> >>>>>>> project who worked the most on debugging subtle bugs, making
>> Lucene
>> >>>>>>> more robust and improving our test framework, so I'm listening
>> when he
>> >>>>>>> voices quality concerns.
>> >>>>>>>
>> >>>>>>> The argument against removing/raising the limit that resonates
>> with me
>> >>>>>>> the most is that it is a one-way door. As MikeS highlighted
>> earlier on
>> >>>>>>> this thread, implementations may want to take advantage of the
>> fact
>> >>>>>>> that there is a limit at some point too. This is why I don't want
>> to
>> >>>>>>> remove the limit and would prefer a slight increase, such as 2048
>> as
>> >>>>>>> suggested in the original issue, which would enable most of the
>> things
>> >>>>>>> that users who have been asking about raising the limit would
>> like to
>> >>>>>>> do.
>> >>>>>>>
>> >>>>>>> I agree that the merge-time memory usage and slow indexing rate
>> are
>> >>>>>>> not great. But it's still possible to index multi-million vector
>> >>>>>>> datasets with a 4GB heap without hitting OOMEs regardless of the
>> >>>>>>> number of dimensions, and the feedback I'm seeing is that many
>> users
>> >>>>>>> are still interested in indexing multi-million vector datasets
>> despite
>> >>>>>>> the slow indexing rate. I wish we could do better, and vector
>> indexing
>> >>>>>>> is certainly more expert than text indexing, but it still is
>> usable in
>> >>>>>>> my opinion. I understand how giving Lucene more information about
>> >>>>>>> vectors prior to indexing (e.g. clustering information as Jim
>> pointed
>> >>>>>>> out) could help make merging faster and more memory-efficient,
>> but I
>> >>>>>>> would really like to avoid making it a requirement for indexing
>> >>>>>>> vectors as it also makes this feature much harder to use.
>> >>>>>>>
>> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>> >>>>>>> <a.benedetti@sease.io> wrote:
>> >>>>>>>> I am very attentive to listen opinions but I am un-convinced
>> here and I an not sure that a single person opinion should be allowed to be
>> detrimental for such an important project.
>> >>>>>>>>
>> >>>>>>>> The limit as far as I know is literally just raising an
>> exception.
>> >>>>>>>> Removing it won't alter in any way the current performance for
>> users in low dimensional space.
>> >>>>>>>> Removing it will just enable more users to use Lucene.
>> >>>>>>>>
>> >>>>>>>> If new users in certain situations will be unhappy with the
>> performance, they may contribute improvements.
>> >>>>>>>> This is how you make progress.
>> >>>>>>>>
>> >>>>>>>> If it's a reputation thing, trust me that not allowing users to
>> play with high dimensional space will equally damage it.
>> >>>>>>>>
>> >>>>>>>> To me it's really a no brainer.
>> >>>>>>>> Removing the limit and enable people to use high dimensional
>> vectors will take minutes.
>> >>>>>>>> Improving the hnsw implementation can take months.
>> >>>>>>>> Pick one to begin with...
>> >>>>>>>>
>> >>>>>>>> And there's no-one paying me here, no company interest
>> whatsoever, actually I pay people to contribute, I am just convinced it's a
>> good idea.
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>> >>>>>>>>> I disagree with your categorization. I put in plenty of work and
>> >>>>>>>>> experienced plenty of pain myself, writing tests and fighting
>> these
>> >>>>>>>>> issues, after i saw that, two releases in a row, vector
>> indexing fell
>> >>>>>>>>> over and hit integer overflows etc on small datasets:
>> >>>>>>>>>
>> >>>>>>>>> https://github.com/apache/lucene/pull/11905
>> >>>>>>>>>
>> >>>>>>>>> Attacking me isn't helping the situation.
>> >>>>>>>>>
>> >>>>>>>>> PS: when i said the "one guy who wrote the code" I didn't mean
>> it in
>> >>>>>>>>> any kind of demeaning fashion really. I meant to describe the
>> current
>> >>>>>>>>> state of usability with respect to indexing a few million docs
>> with
>> >>>>>>>>> high dimensions. You can scroll up the thread and see that at
>> least
>> >>>>>>>>> one other committer on the project experienced similar pain as
>> me.
>> >>>>>>>>> Then, think about users who aren't committers trying to use the
>> >>>>>>>>> functionality!
>> >>>>>>>>>
>> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <
>> msokolov@gmail.com> wrote:
>> >>>>>>>>>> What you said about increasing dimensions requiring a bigger
>> ram buffer on merge is wrong. That's the point I was trying to make. Your
>> concerns about merge costs are not wrong, but your conclusion that we need
>> to limit dimensions is not justified.
>> >>>>>>>>>>
>> >>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when I show
>> it scales linearly with dimension you just ignore that and complain about
>> something entirely different.
>> >>>>>>>>>>
>> >>>>>>>>>> You demand that people run all kinds of tests to prove you
>> wrong but when they do, you don't listen and you won't put in the work
>> yourself or complain that it's too hard.
>> >>>>>>>>>>
>> >>>>>>>>>> Then you complain about people not meeting you half way. Wow
>> >>>>>>>>>>
>> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
>> wrote:
>> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>> >>>>>>>>>>>> What exactly do you consider reasonable?
>> >>>>>>>>>>> Let's begin a real discussion by being HONEST about the
>> current
>> >>>>>>>>>>> status. Please put politically correct or your own company's
>> wishes
>> >>>>>>>>>>> aside, we know it's not in a good state.
>> >>>>>>>>>>>
>> >>>>>>>>>>> Current status is the one guy who wrote the code can set a
>> >>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset with 1024
>> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>> >>>>>>>>>>>
>> >>>>>>>>>>> My concerns are everyone else except the one guy, I want it
>> to be
>> >>>>>>>>>>> usable. Increasing dimensions just means even bigger
>> multi-gigabyte
>> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>> >>>>>>>>>>> It is also a permanent backwards compatibility decision, we
>> have to
>> >>>>>>>>>>> support it once we do this and we can't just say "oops" and
>> flip it
>> >>>>>>>>>>> back.
>> >>>>>>>>>>>
>> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer is
>> really to
>> >>>>>>>>>>> avoid merges because they are so slow and it would be DAYS
>> otherwise,
>> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>> >>>>>>>>>>> Also from personal experience, it takes trial and error (means
>> >>>>>>>>>>> experiencing OOM on merge!!!) before you get those heap
>> values correct
>> >>>>>>>>>>> for your dataset. This usually means starting over which is
>> >>>>>>>>>>> frustrating and wastes more time.
>> >>>>>>>>>>>
>> >>>>>>>>>>> Jim mentioned some ideas about the memory usage in
>> IndexWriter, seems
>> >>>>>>>>>>> to me like its a good idea. maybe the multigigabyte ram
>> buffer can be
>> >>>>>>>>>>> avoided in this way and performance improved by writing bigger
>> >>>>>>>>>>> segments with lucene's defaults. But this doesn't mean we can
>> simply
>> >>>>>>>>>>> ignore the horrors of what happens on merge. merging needs to
>> scale so
>> >>>>>>>>>>> that indexing really scales.
>> >>>>>>>>>>>
>> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts and
>> cause OOM,
>> >>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU in
>> O(n^2)
>> >>>>>>>>>>> fashion when indexing.
>> >>>>>>>>>>>
>> >>>>>>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>>>>>>
>> >>>>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>>>>
>> >>>>>>>
>> >>>>>>> --
>> >>>>>>> Adrien
>> >>>>>>>
>> >>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>>
>> >>>>>>
>> ---------------------------------------------------------------------
>> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>
>> >>>>
>> >>>> --
>> >>>> http://www.needhamsoftware.com (work)
>> >>>> http://www.the111shift.com (play)
>> >>
>> > ---------------------------------------------------------------------
>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> > For additional commands, e-mail: dev-help@lucene.apache.org
>> >
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> For additional commands, e-mail: dev-help@lucene.apache.org
>>
>>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 12, 2023, 5:53 AM

Post #93 of 99 (942 views)

Just addressing [1] I believe there is a simple workaround. Here's a
unit test demonstrating:

public void testExcessivelyLargeVector() throws Exception {
IndexableFieldType vector2048 = new FieldType() {
@Override
public int vectorDimension() {
return 2048;
}

@Override
public VectorEncoding vectorEncoding() {
return VectorEncoding.FLOAT32;
}

@Override
public VectorSimilarityFunction vectorSimilarityFunction() {
return VectorSimilarityFunction.EUCLIDEAN;
}
};
try (Directory dir = newDirectory();
IndexWriter iw = new IndexWriter(dir,
newIndexWriterConfig(null).setCodec(codec))) {
Document doc = new Document();
FieldType type = new FieldType(vector2048);
doc.add(new KnnVectorField("vector2048", new float[2048], type));
iw.addDocument(doc);
}
}

On Wed, Apr 12, 2023 at 8:10?AM Alessandro Benedetti
<a.benedetti@sease.io> wrote:
>
> My tentative of listing here only a set of proposals to then vote, has unfortunately failed.
>
> I appreciate the discussion on better benchmarking hnsw but my feeling is that this discussion is orthogonal to the limit discussion itself, should we create a separate mail thread/github jira issue for that?
>
> At the moment I see at least three lines of activities as an outcome from this (maybe too long) discussion:
>
> 1) [small task] there's a need from a good amount of people of increasing/removing the max limit, as an enabler, to get more users to Lucene and ease adoption for systems Lucene based (Apache Solr, Elasticsearch, OpenSearch)
>
> 2) [medium task] we all want more benchmarks for Lucene vector-based search, with a good variety of vector dimensions and encodings
>
> 3) [big task? ] some people would like to improve vector based search peformance because currently not acceptable, it's not clear when and how
>
> A question I have for point 1, does it really need to be a one way door?
> Can't we reduce the max limit in the future if the implementation becomes coupled with certain dimension sizes?
> It's not ideal I agree, but is back-compatibility more important than pragmatic benefits?
> I. E.
> Right now there's no implementation coupled with the max limit - > we remove/increase the limit and get more Users
>
> With Lucene X.Y a clever committer introduces a super nice implementation improvements that unfortunately limit the max size to K.
> Can't we just document it as a breaking change for such release? So at that point we won't support >K vectors but for a reason?
>
> Do we have similar precedents in Lucene?
>
>
>
> On Wed, 12 Apr 2023, 08:36 Michael Wechner, <michael.wechner@wyona.com> wrote:
>>
>> thank you very much for your feedback!
>>
>> In a previous post (April 7) you wrote you could make availlable the 47K ada-002 vectors, which would be great!
>>
>> Would it make sense to setup a public gitub repo, such that others could use or also contribute vectors?
>>
>> Thanks
>>
>> Michael Wechner
>>
>>
>> Am 12.04.23 um 04:51 schrieb Kent Fitch:
>>
>> I only know some characteristics of the openAI ada-002 vectors, although they are a very popular as embeddings/text-characterisations as they allow more accurate/"human meaningful" semantic search results with fewer dimensions than their predecessors - I've evaluated a few different embedding models, including some BERT variants, CLIP ViT-L-14 (with 768 dims, which was quite good), openAI's ada-001 (1024 dims) and babbage-001 (2048 dims), and ada-002 are qualitatively the best, although that will certainly change!
>>
>> In any case, ada-002 vectors have interesting characteristics that I think mean you could confidently create synthetic vectors which would be hard to distinguish from "real" vectors. I found this from looking at 47K ada-002 vectors generated across a full year (1994) of newspaper articles from the Canberra Times and 200K wikipedia articles:
>> - there is no discernible/significant correlation between values in any pair of dimensions
>> - all but 5 of the 1536 dimensions have an almost identical distribution of values shown in the central blob on these graphs (that just show a few of these 1531 dimensions with clumped values and the 5 "outlier" dimensions, but all 1531 non-outlier dims are in there, which makes for some easy quantisation from float to byte if you dont want to go the full kmeans/clustering/Lloyds-algorithm approach):
>> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
>> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
>> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
>> - the variance of the value of each dimension is characteristic:
>> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>>
>> This probably represents something significant about how the ada-002 embeddings are created, but I think it also means creating "realistic" values is possible. I did not use this information when testing recall & performance on Lucene's HNSW implementation on 192m documents, as I slightly dithered the values of a "real" set on 47K docs and stored other fields in the doc that referenced the "base" document that the dithers were made from, and used different dithering magnitudes so that I could test recall with different neighbour sizes ("M"), construction-beamwidth and search-beamwidths.
>>
>> best regards
>>
>> Kent Fitch
>>
>>
>>
>>
>> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner <michael.wechner@wyona.com> wrote:
>>>
>>> I understand what you mean that it seems to be artificial, but I don't
>>> understand why this matters to test performance and scalability of the
>>> indexing?
>>>
>>> Let's assume the limit of Lucene would be 4 instead of 1024 and there
>>> are only open source models generating vectors with 4 dimensions, for
>>> example
>>>
>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>>>
>>> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>
>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>>>
>>> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>
>>> and now I concatenate them to vectors with 8 dimensions
>>>
>>>
>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>
>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>
>>> and normalize them to length 1.
>>>
>>> Why should this be any different to a model which is acting like a black
>>> box generating vectors with 8 dimensions?
>>>
>>>
>>>
>>>
>>> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>>> >> What exactly do you consider real vector data? Vector data which is based on texts written by humans?
>>> > We have plenty of text; the problem is coming up with a realistic
>>> > vector model that requires as many dimensions as people seem to be
>>> > demanding. As I said above, after surveying huggingface I couldn't
>>> > find any text-based model using more than 768 dimensions. So far we
>>> > have some ideas of generating higher-dimensional data by dithering or
>>> > concatenating existing data, but it seems artificial.
>>> >
>>> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
>>> > <michael.wechner@wyona.com> wrote:
>>> >> What exactly do you consider real vector data? Vector data which is based on texts written by humans?
>>> >>
>>> >> I am asking, because I recently attended the following presentation by Anastassia Shaitarova (UZH Institute for Computational Linguistics, https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>>> >>
>>> >> ----
>>> >>
>>> >> Can we Identify Machine-Generated Text? An Overview of Current Approaches
>>> >> by Anastassia Shaitarova (UZH Institute for Computational Linguistics)
>>> >>
>>> >> The detection of machine-generated text has become increasingly important due to the prevalence of automated content generation and its potential for misuse. In this talk, we will discuss the motivation for automatic detection of generated text. We will present the currently available methods, including feature-based classification as a “first line-of-defense.” We will provide an overview of the detection tools that have been made available so far and discuss their limitations. Finally, we will reflect on some open problems associated with the automatic discrimination of generated texts.
>>> >>
>>> >> ----
>>> >>
>>> >> and her conclusion was that it has become basically impossible to differentiate between text generated by humans and text generated by for example ChatGPT.
>>> >>
>>> >> Whereas others have a slightly different opinion, see for example
>>> >>
>>> >> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>>> >>
>>> >> But I would argue that real world and synthetic have become close enough that testing performance and scalability of indexing should be possible with synthetic data.
>>> >>
>>> >> I completely agree that we have to base our discussions and decisions on scientific methods and that we have to make sure that Lucene performs and scales well and that we understand the limits and what is going on under the hood.
>>> >>
>>> >> Thanks
>>> >>
>>> >> Michael W
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>>> >>
>>> >> +1 to test on real vector data -- if you test on synthetic data you draw synthetic conclusions.
>>> >>
>>> >> Can someone post the theoretical performance (CPU and RAM required) of HNSW construction? Do we know/believe our HNSW implementation has achieved that theoretical big-O performance? Maybe we have some silly performance bug that's causing it not to?
>>> >>
>>> >> As I understand it, HNSW makes the tradeoff of costly construction for faster searching, which is typically the right tradeoff for search use cases. We do this in other parts of the Lucene index too.
>>> >>
>>> >> Lucene will do a logarithmic number of merges over time, i.e. each doc will be merged O(log(N)) times in its lifetime in the index. We need to multiply that by the cost of re-building the whole HNSW graph on each merge. BTW, other things in Lucene, like BKD/dimensional points, also rebuild the whole data structure on each merge, I think? But, as Rob pointed out, stored fields merging do indeed do some sneaky tricks to avoid excessive block decompress/recompress on each merge.
>>> >>
>>> >>> As I understand it, vetoes must have technical merit. I'm not sure that this veto rises to "technical merit" on 2 counts:
>>> >> Actually I think Robert's veto stands on its technical merit already. Robert's take on technical matters very much resonate with me, even if he is sometimes prickly in how he expresses them ;)
>>> >>
>>> >> His point is that we, as a dev community, are not paying enough attention to the indexing performance of our KNN algo (HNSW) and implementation, and that it is reckless to increase / remove limits in that state. It is indeed a one-way door decision and one must confront such decisions with caution, especially for such a widely used base infrastructure as Lucene. We don't even advertise today in our javadocs that you need XXX heap if you index vectors with dimension Y, fanout X, levels Z, etc.
>>> >>
>>> >> RAM used during merging is unaffected by dimensionality, but is affected by fanout, because the HNSW graph (not the raw vectors) is memory resident, I think? Maybe we could move it off-heap and let the OS manage the memory (and still document the RAM requirements)? Maybe merge RAM costs should be accounted for in IW's RAM buffer accounting? It is not today, and there are some other things that use non-trivial RAM, e.g. the doc mapping (to compress docid space when deletions are reclaimed).
>>> >>
>>> >> When we added KNN vector testing to Lucene's nightly benchmarks, the indexing time massively increased -- see annotations DH and DP here: https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of course, that is using a single thread for indexing (on a box that has 128 cores!) so we produce a deterministic index every night ...
>>> >>
>>> >> Stepping out (meta) a bit ... this discussion is precisely one of the awesome benefits of the (informed) veto. It means risky changes to the software, as determined by any single informed developer on the project, can force a healthy discussion about the problem at hand. Robert is legitimately concerned about a real issue and so we should use our creative energies to characterize our HNSW implementation's performance, document it clearly for users, and uncover ways to improve it.
>>> >>
>>> >> Mike McCandless
>>> >>
>>> >> http://blog.mikemccandless.com
>>> >>
>>> >>
>>> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <a.benedetti@sease.io> wrote:
>>> >>> I think Gus points are on target.
>>> >>>
>>> >>> I recommend we move this forward in this way:
>>> >>> We stop any discussion and everyone interested proposes an option with a motivation, then we aggregate the options and we create a Vote maybe?
>>> >>>
>>> >>> I am also on the same page on the fact that a veto should come with a clear and reasonable technical merit, which also in my opinion has not come yet.
>>> >>>
>>> >>> I also apologise if any of my words sounded harsh or personal attacks, never meant to do so.
>>> >>>
>>> >>> My proposed option:
>>> >>>
>>> >>> 1) remove the limit and potentially make it configurable,
>>> >>> Motivation:
>>> >>> The system administrator can enforce a limit its users need to respect that it's in line with whatever the admin decided to be acceptable for them.
>>> >>> Default can stay the current one.
>>> >>>
>>> >>> That's my favourite at the moment, but I agree that potentially in the future this may need to change, as we may optimise the data structures for certain dimensions. I am a big fan of Yagni (you aren't going to need it) so I am ok we'll face a different discussion if that happens in the future.
>>> >>>
>>> >>>
>>> >>>
>>> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>>> >>>> What I see so far:
>>> >>>>
>>> >>>> Much positive support for raising the limit
>>> >>>> Slightly less support for removing it or making it configurable
>>> >>>> A single veto which argues that a (as yet undefined) performance standard must be met before raising the limit
>>> >>>> Hot tempers (various) making this discussion difficult
>>> >>>>
>>> >>>> As I understand it, vetoes must have technical merit. I'm not sure that this veto rises to "technical merit" on 2 counts:
>>> >>>>
>>> >>>> No standard for the performance is given so it cannot be technically met. Without hard criteria it's a moving target.
>>> >>>> It appears to encode a valuation of the user's time, and that valuation is really up to the user. Some users may consider 2hours useless and not worth it, and others might happily wait 2 hours. This is not a technical decision, it's a business decision regarding the relative value of the time invested vs the value of the result. If I can cure cancer by indexing for a year, that might be worth it... (hyperbole of course).
>>> >>>>
>>> >>>> Things I would consider to have technical merit that I don't hear:
>>> >>>>
>>> >>>> Impact on the speed of **other** indexing operations. (devaluation of other functionality)
>>> >>>> Actual scenarios that work when the limit is low and fail when the limit is high (new failure on the same data with the limit raised).
>>> >>>>
>>> >>>> One thing that might or might not have technical merit
>>> >>>>
>>> >>>> If someone feels there is a lack of documentation of the costs/performance implications of using large vectors, possibly including reproducible benchmarks establishing the scaling behavior (there seems to be disagreement on O(n) vs O(n^2)).
>>> >>>>
>>> >>>> The users *should* know what they are getting into, but if the cost is worth it to them, they should be able to pay it without forking the project. If this veto causes a fork that's not good.
>>> >>>>
>>> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com> wrote:
>>> >>>>> We do have a dataset built from Wikipedia in luceneutil. It comes in 100 and 300 dimensional varieties and can easily enough generate large numbers of vector documents from the articles data. To go higher we could concatenate vectors from that and I believe the performance numbers would be plausible.
>>> >>>>>
>>> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com> wrote:
>>> >>>>>> Can we set up a branch in which the limit is bumped to 2048, then have
>>> >>>>>> a realistic, free data set (wikipedia sample or something) that has,
>>> >>>>>> say, 5 million docs and vectors created using public data (glove
>>> >>>>>> pre-trained embeddings or the like)? We then could run indexing on the
>>> >>>>>> same hardware with 512, 1024 and 2048 and see what the numbers, limits
>>> >>>>>> and behavior actually are.
>>> >>>>>>
>>> >>>>>> I can help in writing this but not until after Easter.
>>> >>>>>>
>>> >>>>>>
>>> >>>>>> Dawid
>>> >>>>>>
>>> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com> wrote:
>>> >>>>>>> As Dawid pointed out earlier on this thread, this is the rule for
>>> >>>>>>> Apache projects: a single -1 vote on a code change is a veto and
>>> >>>>>>> cannot be overridden. Furthermore, Robert is one of the people on this
>>> >>>>>>> project who worked the most on debugging subtle bugs, making Lucene
>>> >>>>>>> more robust and improving our test framework, so I'm listening when he
>>> >>>>>>> voices quality concerns.
>>> >>>>>>>
>>> >>>>>>> The argument against removing/raising the limit that resonates with me
>>> >>>>>>> the most is that it is a one-way door. As MikeS highlighted earlier on
>>> >>>>>>> this thread, implementations may want to take advantage of the fact
>>> >>>>>>> that there is a limit at some point too. This is why I don't want to
>>> >>>>>>> remove the limit and would prefer a slight increase, such as 2048 as
>>> >>>>>>> suggested in the original issue, which would enable most of the things
>>> >>>>>>> that users who have been asking about raising the limit would like to
>>> >>>>>>> do.
>>> >>>>>>>
>>> >>>>>>> I agree that the merge-time memory usage and slow indexing rate are
>>> >>>>>>> not great. But it's still possible to index multi-million vector
>>> >>>>>>> datasets with a 4GB heap without hitting OOMEs regardless of the
>>> >>>>>>> number of dimensions, and the feedback I'm seeing is that many users
>>> >>>>>>> are still interested in indexing multi-million vector datasets despite
>>> >>>>>>> the slow indexing rate. I wish we could do better, and vector indexing
>>> >>>>>>> is certainly more expert than text indexing, but it still is usable in
>>> >>>>>>> my opinion. I understand how giving Lucene more information about
>>> >>>>>>> vectors prior to indexing (e.g. clustering information as Jim pointed
>>> >>>>>>> out) could help make merging faster and more memory-efficient, but I
>>> >>>>>>> would really like to avoid making it a requirement for indexing
>>> >>>>>>> vectors as it also makes this feature much harder to use.
>>> >>>>>>>
>>> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>> >>>>>>> <a.benedetti@sease.io> wrote:
>>> >>>>>>>> I am very attentive to listen opinions but I am un-convinced here and I an not sure that a single person opinion should be allowed to be detrimental for such an important project.
>>> >>>>>>>>
>>> >>>>>>>> The limit as far as I know is literally just raising an exception.
>>> >>>>>>>> Removing it won't alter in any way the current performance for users in low dimensional space.
>>> >>>>>>>> Removing it will just enable more users to use Lucene.
>>> >>>>>>>>
>>> >>>>>>>> If new users in certain situations will be unhappy with the performance, they may contribute improvements.
>>> >>>>>>>> This is how you make progress.
>>> >>>>>>>>
>>> >>>>>>>> If it's a reputation thing, trust me that not allowing users to play with high dimensional space will equally damage it.
>>> >>>>>>>>
>>> >>>>>>>> To me it's really a no brainer.
>>> >>>>>>>> Removing the limit and enable people to use high dimensional vectors will take minutes.
>>> >>>>>>>> Improving the hnsw implementation can take months.
>>> >>>>>>>> Pick one to begin with...
>>> >>>>>>>>
>>> >>>>>>>> And there's no-one paying me here, no company interest whatsoever, actually I pay people to contribute, I am just convinced it's a good idea.
>>> >>>>>>>>
>>> >>>>>>>>
>>> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>>> >>>>>>>>> I disagree with your categorization. I put in plenty of work and
>>> >>>>>>>>> experienced plenty of pain myself, writing tests and fighting these
>>> >>>>>>>>> issues, after i saw that, two releases in a row, vector indexing fell
>>> >>>>>>>>> over and hit integer overflows etc on small datasets:
>>> >>>>>>>>>
>>> >>>>>>>>> https://github.com/apache/lucene/pull/11905
>>> >>>>>>>>>
>>> >>>>>>>>> Attacking me isn't helping the situation.
>>> >>>>>>>>>
>>> >>>>>>>>> PS: when i said the "one guy who wrote the code" I didn't mean it in
>>> >>>>>>>>> any kind of demeaning fashion really. I meant to describe the current
>>> >>>>>>>>> state of usability with respect to indexing a few million docs with
>>> >>>>>>>>> high dimensions. You can scroll up the thread and see that at least
>>> >>>>>>>>> one other committer on the project experienced similar pain as me.
>>> >>>>>>>>> Then, think about users who aren't committers trying to use the
>>> >>>>>>>>> functionality!
>>> >>>>>>>>>
>>> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <msokolov@gmail.com> wrote:
>>> >>>>>>>>>> What you said about increasing dimensions requiring a bigger ram buffer on merge is wrong. That's the point I was trying to make. Your concerns about merge costs are not wrong, but your conclusion that we need to limit dimensions is not justified.
>>> >>>>>>>>>>
>>> >>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when I show it scales linearly with dimension you just ignore that and complain about something entirely different.
>>> >>>>>>>>>>
>>> >>>>>>>>>> You demand that people run all kinds of tests to prove you wrong but when they do, you don't listen and you won't put in the work yourself or complain that it's too hard.
>>> >>>>>>>>>>
>>> >>>>>>>>>> Then you complain about people not meeting you half way. Wow
>>> >>>>>>>>>>
>>> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com> wrote:
>>> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>>> >>>>>>>>>>>> What exactly do you consider reasonable?
>>> >>>>>>>>>>> Let's begin a real discussion by being HONEST about the current
>>> >>>>>>>>>>> status. Please put politically correct or your own company's wishes
>>> >>>>>>>>>>> aside, we know it's not in a good state.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> Current status is the one guy who wrote the code can set a
>>> >>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset with 1024
>>> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> My concerns are everyone else except the one guy, I want it to be
>>> >>>>>>>>>>> usable. Increasing dimensions just means even bigger multi-gigabyte
>>> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>>> >>>>>>>>>>> It is also a permanent backwards compatibility decision, we have to
>>> >>>>>>>>>>> support it once we do this and we can't just say "oops" and flip it
>>> >>>>>>>>>>> back.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer is really to
>>> >>>>>>>>>>> avoid merges because they are so slow and it would be DAYS otherwise,
>>> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>>> >>>>>>>>>>> Also from personal experience, it takes trial and error (means
>>> >>>>>>>>>>> experiencing OOM on merge!!!) before you get those heap values correct
>>> >>>>>>>>>>> for your dataset. This usually means starting over which is
>>> >>>>>>>>>>> frustrating and wastes more time.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> Jim mentioned some ideas about the memory usage in IndexWriter, seems
>>> >>>>>>>>>>> to me like its a good idea. maybe the multigigabyte ram buffer can be
>>> >>>>>>>>>>> avoided in this way and performance improved by writing bigger
>>> >>>>>>>>>>> segments with lucene's defaults. But this doesn't mean we can simply
>>> >>>>>>>>>>> ignore the horrors of what happens on merge. merging needs to scale so
>>> >>>>>>>>>>> that indexing really scales.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts and cause OOM,
>>> >>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU in O(n^2)
>>> >>>>>>>>>>> fashion when indexing.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> ---------------------------------------------------------------------
>>> >>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>>>>>>
>>> >>>>>>>>> ---------------------------------------------------------------------
>>> >>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>>>>
>>> >>>>>>>
>>> >>>>>>> --
>>> >>>>>>> Adrien
>>> >>>>>>>
>>> >>>>>>> ---------------------------------------------------------------------
>>> >>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>>
>>> >>>>>> ---------------------------------------------------------------------
>>> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>
>>> >>>>
>>> >>>> --
>>> >>>> http://www.needhamsoftware.com (work)
>>> >>>> http://www.the111shift.com (play)
>>> >>
>>> > ---------------------------------------------------------------------
>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>> >
>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>
>>

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

kent.fitch at gmail

Apr 12, 2023, 4:35 PM

Post #94 of 99 (942 views)

Hi Michael (and anyone else who wants just over 240K "real world" ada-002
vectors of dimension 1536),
you are welcome to retrieve a tar.gz file which contains:
- 47K embeddings of Canberra Times news article text from 1994
- 38K embeddings of the first paragraphs of wikipedia articles about
organisations
- 156.6K embeddings of the first paragraphs of wikipedia articles about
people

https://drive.google.com/file/d/13JP_5u7E8oZO6vRg0ekaTgBDQOaj-W00/view?usp=sharing

The file is about 1.7GB and will expand to about 4.4GB. This file will be
accessible for at least a week, and I hope you dont hit any google drive
download limits trying to retrieve it.

The embeddings were generated using my openAI account and you are welcome
to use them for any purpose you like.

best wishes,

Kent Fitch

On Wed, Apr 12, 2023 at 4:37?PM Michael Wechner <michael.wechner@wyona.com>
wrote:

> thank you very much for your feedback!
>
> In a previous post (April 7) you wrote you could make availlable the 47K
> ada-002 vectors, which would be great!
>
> Would it make sense to setup a public gitub repo, such that others could
> use or also contribute vectors?
>
> Thanks
>
> Michael Wechner
>
>
> Am 12.04.23 um 04:51 schrieb Kent Fitch:
>
> I only know some characteristics of the openAI ada-002 vectors, although
> they are a very popular as embeddings/text-characterisations as they allow
> more accurate/"human meaningful" semantic search results with fewer
> dimensions than their predecessors - I've evaluated a few different
> embedding models, including some BERT variants, CLIP ViT-L-14 (with 768
> dims, which was quite good), openAI's ada-001 (1024 dims) and babbage-001
> (2048 dims), and ada-002 are qualitatively the best, although that will
> certainly change!
>
> In any case, ada-002 vectors have interesting characteristics that I think
> mean you could confidently create synthetic vectors which would be hard to
> distinguish from "real" vectors. I found this from looking at 47K ada-002
> vectors generated across a full year (1994) of newspaper articles from the
> Canberra Times and 200K wikipedia articles:
> - there is no discernible/significant correlation between values in any
> pair of dimensions
> - all but 5 of the 1536 dimensions have an almost identical distribution
> of values shown in the central blob on these graphs (that just show a few
> of these 1531 dimensions with clumped values and the 5 "outlier"
> dimensions, but all 1531 non-outlier dims are in there, which makes for
> some easy quantisation from float to byte if you dont want to go the full
> kmeans/clustering/Lloyds-algorithm approach):
>
> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
>
> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
>
> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
> - the variance of the value of each dimension is characteristic:
>
> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>
> This probably represents something significant about how the ada-002
> embeddings are created, but I think it also means creating "realistic"
> values is possible. I did not use this information when testing recall &
> performance on Lucene's HNSW implementation on 192m documents, as I
> slightly dithered the values of a "real" set on 47K docs and stored other
> fields in the doc that referenced the "base" document that the dithers were
> made from, and used different dithering magnitudes so that I could test
> recall with different neighbour sizes ("M"), construction-beamwidth and
> search-beamwidths.
>
> best regards
>
> Kent Fitch
>
>
>
>
> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner <michael.wechner@wyona.com>
> wrote:
>
>> I understand what you mean that it seems to be artificial, but I don't
>> understand why this matters to test performance and scalability of the
>> indexing?
>>
>> Let's assume the limit of Lucene would be 4 instead of 1024 and there
>> are only open source models generating vectors with 4 dimensions, for
>> example
>>
>>
>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>>
>>
>> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>
>>
>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>>
>>
>> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>
>> and now I concatenate them to vectors with 8 dimensions
>>
>>
>>
>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>
>>
>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>
>> and normalize them to length 1.
>>
>> Why should this be any different to a model which is acting like a black
>> box generating vectors with 8 dimensions?
>>
>>
>>
>>
>> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>> >> What exactly do you consider real vector data? Vector data which is
>> based on texts written by humans?
>> > We have plenty of text; the problem is coming up with a realistic
>> > vector model that requires as many dimensions as people seem to be
>> > demanding. As I said above, after surveying huggingface I couldn't
>> > find any text-based model using more than 768 dimensions. So far we
>> > have some ideas of generating higher-dimensional data by dithering or
>> > concatenating existing data, but it seems artificial.
>> >
>> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
>> > <michael.wechner@wyona.com> wrote:
>> >> What exactly do you consider real vector data? Vector data which is
>> based on texts written by humans?
>> >>
>> >> I am asking, because I recently attended the following presentation by
>> Anastassia Shaitarova (UZH Institute for Computational Linguistics,
>> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>> >>
>> >> ----
>> >>
>> >> Can we Identify Machine-Generated Text? An Overview of Current
>> Approaches
>> >> by Anastassia Shaitarova (UZH Institute for Computational Linguistics)
>> >>
>> >> The detection of machine-generated text has become increasingly
>> important due to the prevalence of automated content generation and its
>> potential for misuse. In this talk, we will discuss the motivation for
>> automatic detection of generated text. We will present the currently
>> available methods, including feature-based classification as a “first
>> line-of-defense.” We will provide an overview of the detection tools that
>> have been made available so far and discuss their limitations. Finally, we
>> will reflect on some open problems associated with the automatic
>> discrimination of generated texts.
>> >>
>> >> ----
>> >>
>> >> and her conclusion was that it has become basically impossible to
>> differentiate between text generated by humans and text generated by for
>> example ChatGPT.
>> >>
>> >> Whereas others have a slightly different opinion, see for example
>> >>
>> >> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>> >>
>> >> But I would argue that real world and synthetic have become close
>> enough that testing performance and scalability of indexing should be
>> possible with synthetic data.
>> >>
>> >> I completely agree that we have to base our discussions and decisions
>> on scientific methods and that we have to make sure that Lucene performs
>> and scales well and that we understand the limits and what is going on
>> under the hood.
>> >>
>> >> Thanks
>> >>
>> >> Michael W
>> >>
>> >>
>> >>
>> >>
>> >>
>> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>> >>
>> >> +1 to test on real vector data -- if you test on synthetic data you
>> draw synthetic conclusions.
>> >>
>> >> Can someone post the theoretical performance (CPU and RAM required) of
>> HNSW construction? Do we know/believe our HNSW implementation has achieved
>> that theoretical big-O performance? Maybe we have some silly performance
>> bug that's causing it not to?
>> >>
>> >> As I understand it, HNSW makes the tradeoff of costly construction for
>> faster searching, which is typically the right tradeoff for search use
>> cases. We do this in other parts of the Lucene index too.
>> >>
>> >> Lucene will do a logarithmic number of merges over time, i.e. each doc
>> will be merged O(log(N)) times in its lifetime in the index. We need to
>> multiply that by the cost of re-building the whole HNSW graph on each
>> merge. BTW, other things in Lucene, like BKD/dimensional points, also
>> rebuild the whole data structure on each merge, I think? But, as Rob
>> pointed out, stored fields merging do indeed do some sneaky tricks to avoid
>> excessive block decompress/recompress on each merge.
>> >>
>> >>> As I understand it, vetoes must have technical merit. I'm not sure
>> that this veto rises to "technical merit" on 2 counts:
>> >> Actually I think Robert's veto stands on its technical merit already.
>> Robert's take on technical matters very much resonate with me, even if he
>> is sometimes prickly in how he expresses them ;)
>> >>
>> >> His point is that we, as a dev community, are not paying enough
>> attention to the indexing performance of our KNN algo (HNSW) and
>> implementation, and that it is reckless to increase / remove limits in that
>> state. It is indeed a one-way door decision and one must confront such
>> decisions with caution, especially for such a widely used base
>> infrastructure as Lucene. We don't even advertise today in our javadocs
>> that you need XXX heap if you index vectors with dimension Y, fanout X,
>> levels Z, etc.
>> >>
>> >> RAM used during merging is unaffected by dimensionality, but is
>> affected by fanout, because the HNSW graph (not the raw vectors) is memory
>> resident, I think? Maybe we could move it off-heap and let the OS manage
>> the memory (and still document the RAM requirements)? Maybe merge RAM
>> costs should be accounted for in IW's RAM buffer accounting? It is not
>> today, and there are some other things that use non-trivial RAM, e.g. the
>> doc mapping (to compress docid space when deletions are reclaimed).
>> >>
>> >> When we added KNN vector testing to Lucene's nightly benchmarks, the
>> indexing time massively increased -- see annotations DH and DP here:
>> https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly
>> benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of
>> course, that is using a single thread for indexing (on a box that has 128
>> cores!) so we produce a deterministic index every night ...
>> >>
>> >> Stepping out (meta) a bit ... this discussion is precisely one of the
>> awesome benefits of the (informed) veto. It means risky changes to the
>> software, as determined by any single informed developer on the project,
>> can force a healthy discussion about the problem at hand. Robert is
>> legitimately concerned about a real issue and so we should use our creative
>> energies to characterize our HNSW implementation's performance, document it
>> clearly for users, and uncover ways to improve it.
>> >>
>> >> Mike McCandless
>> >>
>> >> http://blog.mikemccandless.com
>> >>
>> >>
>> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <
>> a.benedetti@sease.io> wrote:
>> >>> I think Gus points are on target.
>> >>>
>> >>> I recommend we move this forward in this way:
>> >>> We stop any discussion and everyone interested proposes an option
>> with a motivation, then we aggregate the options and we create a Vote maybe?
>> >>>
>> >>> I am also on the same page on the fact that a veto should come with a
>> clear and reasonable technical merit, which also in my opinion has not come
>> yet.
>> >>>
>> >>> I also apologise if any of my words sounded harsh or personal
>> attacks, never meant to do so.
>> >>>
>> >>> My proposed option:
>> >>>
>> >>> 1) remove the limit and potentially make it configurable,
>> >>> Motivation:
>> >>> The system administrator can enforce a limit its users need to
>> respect that it's in line with whatever the admin decided to be acceptable
>> for them.
>> >>> Default can stay the current one.
>> >>>
>> >>> That's my favourite at the moment, but I agree that potentially in
>> the future this may need to change, as we may optimise the data structures
>> for certain dimensions. I am a big fan of Yagni (you aren't going to need
>> it) so I am ok we'll face a different discussion if that happens in the
>> future.
>> >>>
>> >>>
>> >>>
>> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>> >>>> What I see so far:
>> >>>>
>> >>>> Much positive support for raising the limit
>> >>>> Slightly less support for removing it or making it configurable
>> >>>> A single veto which argues that a (as yet undefined) performance
>> standard must be met before raising the limit
>> >>>> Hot tempers (various) making this discussion difficult
>> >>>>
>> >>>> As I understand it, vetoes must have technical merit. I'm not sure
>> that this veto rises to "technical merit" on 2 counts:
>> >>>>
>> >>>> No standard for the performance is given so it cannot be technically
>> met. Without hard criteria it's a moving target.
>> >>>> It appears to encode a valuation of the user's time, and that
>> valuation is really up to the user. Some users may consider 2hours useless
>> and not worth it, and others might happily wait 2 hours. This is not a
>> technical decision, it's a business decision regarding the relative value
>> of the time invested vs the value of the result. If I can cure cancer by
>> indexing for a year, that might be worth it... (hyperbole of course).
>> >>>>
>> >>>> Things I would consider to have technical merit that I don't hear:
>> >>>>
>> >>>> Impact on the speed of **other** indexing operations. (devaluation
>> of other functionality)
>> >>>> Actual scenarios that work when the limit is low and fail when the
>> limit is high (new failure on the same data with the limit raised).
>> >>>>
>> >>>> One thing that might or might not have technical merit
>> >>>>
>> >>>> If someone feels there is a lack of documentation of the
>> costs/performance implications of using large vectors, possibly including
>> reproducible benchmarks establishing the scaling behavior (there seems to
>> be disagreement on O(n) vs O(n^2)).
>> >>>>
>> >>>> The users *should* know what they are getting into, but if the cost
>> is worth it to them, they should be able to pay it without forking the
>> project. If this veto causes a fork that's not good.
>> >>>>
>> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com>
>> wrote:
>> >>>>> We do have a dataset built from Wikipedia in luceneutil. It comes
>> in 100 and 300 dimensional varieties and can easily enough generate large
>> numbers of vector documents from the articles data. To go higher we could
>> concatenate vectors from that and I believe the performance numbers would
>> be plausible.
>> >>>>>
>> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com>
>> wrote:
>> >>>>>> Can we set up a branch in which the limit is bumped to 2048, then
>> have
>> >>>>>> a realistic, free data set (wikipedia sample or something) that
>> has,
>> >>>>>> say, 5 million docs and vectors created using public data (glove
>> >>>>>> pre-trained embeddings or the like)? We then could run indexing on
>> the
>> >>>>>> same hardware with 512, 1024 and 2048 and see what the numbers,
>> limits
>> >>>>>> and behavior actually are.
>> >>>>>>
>> >>>>>> I can help in writing this but not until after Easter.
>> >>>>>>
>> >>>>>>
>> >>>>>> Dawid
>> >>>>>>
>> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com>
>> wrote:
>> >>>>>>> As Dawid pointed out earlier on this thread, this is the rule for
>> >>>>>>> Apache projects: a single -1 vote on a code change is a veto and
>> >>>>>>> cannot be overridden. Furthermore, Robert is one of the people on
>> this
>> >>>>>>> project who worked the most on debugging subtle bugs, making
>> Lucene
>> >>>>>>> more robust and improving our test framework, so I'm listening
>> when he
>> >>>>>>> voices quality concerns.
>> >>>>>>>
>> >>>>>>> The argument against removing/raising the limit that resonates
>> with me
>> >>>>>>> the most is that it is a one-way door. As MikeS highlighted
>> earlier on
>> >>>>>>> this thread, implementations may want to take advantage of the
>> fact
>> >>>>>>> that there is a limit at some point too. This is why I don't want
>> to
>> >>>>>>> remove the limit and would prefer a slight increase, such as 2048
>> as
>> >>>>>>> suggested in the original issue, which would enable most of the
>> things
>> >>>>>>> that users who have been asking about raising the limit would
>> like to
>> >>>>>>> do.
>> >>>>>>>
>> >>>>>>> I agree that the merge-time memory usage and slow indexing rate
>> are
>> >>>>>>> not great. But it's still possible to index multi-million vector
>> >>>>>>> datasets with a 4GB heap without hitting OOMEs regardless of the
>> >>>>>>> number of dimensions, and the feedback I'm seeing is that many
>> users
>> >>>>>>> are still interested in indexing multi-million vector datasets
>> despite
>> >>>>>>> the slow indexing rate. I wish we could do better, and vector
>> indexing
>> >>>>>>> is certainly more expert than text indexing, but it still is
>> usable in
>> >>>>>>> my opinion. I understand how giving Lucene more information about
>> >>>>>>> vectors prior to indexing (e.g. clustering information as Jim
>> pointed
>> >>>>>>> out) could help make merging faster and more memory-efficient,
>> but I
>> >>>>>>> would really like to avoid making it a requirement for indexing
>> >>>>>>> vectors as it also makes this feature much harder to use.
>> >>>>>>>
>> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>> >>>>>>> <a.benedetti@sease.io> wrote:
>> >>>>>>>> I am very attentive to listen opinions but I am un-convinced
>> here and I an not sure that a single person opinion should be allowed to be
>> detrimental for such an important project.
>> >>>>>>>>
>> >>>>>>>> The limit as far as I know is literally just raising an
>> exception.
>> >>>>>>>> Removing it won't alter in any way the current performance for
>> users in low dimensional space.
>> >>>>>>>> Removing it will just enable more users to use Lucene.
>> >>>>>>>>
>> >>>>>>>> If new users in certain situations will be unhappy with the
>> performance, they may contribute improvements.
>> >>>>>>>> This is how you make progress.
>> >>>>>>>>
>> >>>>>>>> If it's a reputation thing, trust me that not allowing users to
>> play with high dimensional space will equally damage it.
>> >>>>>>>>
>> >>>>>>>> To me it's really a no brainer.
>> >>>>>>>> Removing the limit and enable people to use high dimensional
>> vectors will take minutes.
>> >>>>>>>> Improving the hnsw implementation can take months.
>> >>>>>>>> Pick one to begin with...
>> >>>>>>>>
>> >>>>>>>> And there's no-one paying me here, no company interest
>> whatsoever, actually I pay people to contribute, I am just convinced it's a
>> good idea.
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com> wrote:
>> >>>>>>>>> I disagree with your categorization. I put in plenty of work and
>> >>>>>>>>> experienced plenty of pain myself, writing tests and fighting
>> these
>> >>>>>>>>> issues, after i saw that, two releases in a row, vector
>> indexing fell
>> >>>>>>>>> over and hit integer overflows etc on small datasets:
>> >>>>>>>>>
>> >>>>>>>>> https://github.com/apache/lucene/pull/11905
>> >>>>>>>>>
>> >>>>>>>>> Attacking me isn't helping the situation.
>> >>>>>>>>>
>> >>>>>>>>> PS: when i said the "one guy who wrote the code" I didn't mean
>> it in
>> >>>>>>>>> any kind of demeaning fashion really. I meant to describe the
>> current
>> >>>>>>>>> state of usability with respect to indexing a few million docs
>> with
>> >>>>>>>>> high dimensions. You can scroll up the thread and see that at
>> least
>> >>>>>>>>> one other committer on the project experienced similar pain as
>> me.
>> >>>>>>>>> Then, think about users who aren't committers trying to use the
>> >>>>>>>>> functionality!
>> >>>>>>>>>
>> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <
>> msokolov@gmail.com> wrote:
>> >>>>>>>>>> What you said about increasing dimensions requiring a bigger
>> ram buffer on merge is wrong. That's the point I was trying to make. Your
>> concerns about merge costs are not wrong, but your conclusion that we need
>> to limit dimensions is not justified.
>> >>>>>>>>>>
>> >>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when I show
>> it scales linearly with dimension you just ignore that and complain about
>> something entirely different.
>> >>>>>>>>>>
>> >>>>>>>>>> You demand that people run all kinds of tests to prove you
>> wrong but when they do, you don't listen and you won't put in the work
>> yourself or complain that it's too hard.
>> >>>>>>>>>>
>> >>>>>>>>>> Then you complain about people not meeting you half way. Wow
>> >>>>>>>>>>
>> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
>> wrote:
>> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>> >>>>>>>>>>>> What exactly do you consider reasonable?
>> >>>>>>>>>>> Let's begin a real discussion by being HONEST about the
>> current
>> >>>>>>>>>>> status. Please put politically correct or your own company's
>> wishes
>> >>>>>>>>>>> aside, we know it's not in a good state.
>> >>>>>>>>>>>
>> >>>>>>>>>>> Current status is the one guy who wrote the code can set a
>> >>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset with 1024
>> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>> >>>>>>>>>>>
>> >>>>>>>>>>> My concerns are everyone else except the one guy, I want it
>> to be
>> >>>>>>>>>>> usable. Increasing dimensions just means even bigger
>> multi-gigabyte
>> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>> >>>>>>>>>>> It is also a permanent backwards compatibility decision, we
>> have to
>> >>>>>>>>>>> support it once we do this and we can't just say "oops" and
>> flip it
>> >>>>>>>>>>> back.
>> >>>>>>>>>>>
>> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer is
>> really to
>> >>>>>>>>>>> avoid merges because they are so slow and it would be DAYS
>> otherwise,
>> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>> >>>>>>>>>>> Also from personal experience, it takes trial and error (means
>> >>>>>>>>>>> experiencing OOM on merge!!!) before you get those heap
>> values correct
>> >>>>>>>>>>> for your dataset. This usually means starting over which is
>> >>>>>>>>>>> frustrating and wastes more time.
>> >>>>>>>>>>>
>> >>>>>>>>>>> Jim mentioned some ideas about the memory usage in
>> IndexWriter, seems
>> >>>>>>>>>>> to me like its a good idea. maybe the multigigabyte ram
>> buffer can be
>> >>>>>>>>>>> avoided in this way and performance improved by writing bigger
>> >>>>>>>>>>> segments with lucene's defaults. But this doesn't mean we can
>> simply
>> >>>>>>>>>>> ignore the horrors of what happens on merge. merging needs to
>> scale so
>> >>>>>>>>>>> that indexing really scales.
>> >>>>>>>>>>>
>> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts and
>> cause OOM,
>> >>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU in
>> O(n^2)
>> >>>>>>>>>>> fashion when indexing.
>> >>>>>>>>>>>
>> >>>>>>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>>>>>>
>> >>>>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>>>>
>> >>>>>>>
>> >>>>>>> --
>> >>>>>>> Adrien
>> >>>>>>>
>> >>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>>
>> >>>>>>
>> ---------------------------------------------------------------------
>> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>> >>>>>>
>> >>>>
>> >>>> --
>> >>>> http://www.needhamsoftware.com (work)
>> >>>> http://www.the111shift.com (play)
>> >>
>> > ---------------------------------------------------------------------
>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> > For additional commands, e-mail: dev-help@lucene.apache.org
>> >
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> For additional commands, e-mail: dev-help@lucene.apache.org
>>
>>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 12, 2023, 10:58 PM

Post #95 of 99 (942 views)

Hi Kent

Great, thank you very much!

Will download it later today :-)

All the best

Michael

Am 13.04.23 um 01:35 schrieb Kent Fitch:
> Hi Michael (and anyone else who wants just over 240K "real world"
> ada-002 vectors of dimension 1536),
> you are welcome to retrieve a tar.gz file which contains:
> - 47K embeddings of Canberra Times news article text from 1994
> - 38K embeddings of the first paragraphs of wikipedia articles about
> organisations
> - 156.6K embeddings of the first paragraphs of wikipedia articles
> about people
>
> https://drive.google.com/file/d/13JP_5u7E8oZO6vRg0ekaTgBDQOaj-W00/view?usp=sharing
>
> The file is about 1.7GB and will expand to about 4.4GB. This file will
> be accessible for at least a week, and I hope you dont hit any google
> drive download limits trying to retrieve it.
>
> The embeddings were generated using my openAI account and you are
> welcome to use them for any purpose you like.
>
> best wishes,
>
> Kent Fitch
>
> On Wed, Apr 12, 2023 at 4:37?PM Michael Wechner
> <michael.wechner@wyona.com> wrote:
>
> thank you very much for your feedback!
>
> In a previous post (April 7) you wrote you could make availlable
> the 47K ada-002 vectors, which would be great!
>
> Would it make sense to setup a public gitub repo, such that others
> could use or also contribute vectors?
>
> Thanks
>
> Michael Wechner
>
>
> Am 12.04.23 um 04:51 schrieb Kent Fitch:
>> I only know some characteristics of the openAI ada-002 vectors,
>> although they are a very popular as
>> embeddings/text-characterisations as they allow more
>> accurate/"human meaningful" semantic search results with fewer
>> dimensions than their predecessors - I've evaluated a few
>> different embedding models, including some BERT variants, CLIP
>> ViT-L-14 (with 768 dims, which was quite good), openAI's ada-001
>> (1024 dims) and babbage-001 (2048 dims), and ada-002 are
>> qualitatively the best, although that will certainly change!
>>
>> In any case, ada-002 vectors have interesting characteristics
>> that I think mean you could confidently create synthetic vectors
>> which would be hard to distinguish from "real" vectors. I found
>> this from looking at 47K ada-002 vectors generated across a full
>> year (1994) of newspaper articles from the Canberra Times and
>> 200K wikipedia articles:
>> - there is no discernible/significant correlation between values
>> in any pair of dimensions
>> - all but 5 of the 1536 dimensions have an almost identical
>> distribution of values shown in the central blob on these graphs
>> (that just show a few of these 1531 dimensions with clumped
>> values and the 5 "outlier" dimensions, but all 1531 non-outlier
>> dims are in there, which makes for some easy quantisation from
>> float to byte if you dont want to go the full
>> kmeans/clustering/Lloyds-algorithm approach):
>> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
>> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
>> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
>> - the variance of the value of each dimension is characteristic:
>> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>>
>> This probably represents something significant about how the
>> ada-002 embeddings are created, but I think it also means
>> creating "realistic" values is possible. I did not use this
>> information when testing recall & performance on Lucene's HNSW
>> implementation on 192m documents, as I slightly dithered the
>> values of a "real" set on 47K docs and stored other fields in the
>> doc that referenced the "base" document that the dithers were
>> made from, and used different dithering magnitudes so that I
>> could test recall with different neighbour sizes ("M"),
>> construction-beamwidth and search-beamwidths.
>>
>> best regards
>>
>> Kent Fitch
>>
>>
>>
>>
>> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner
>> <michael.wechner@wyona.com> wrote:
>>
>> I understand what you mean that it seems to be artificial,
>> but I don't
>> understand why this matters to test performance and
>> scalability of the
>> indexing?
>>
>> Let's assume the limit of Lucene would be 4 instead of 1024
>> and there
>> are only open source models generating vectors with 4
>> dimensions, for
>> example
>>
>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>>
>> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>
>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>>
>> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>
>> and now I concatenate them to vectors with 8 dimensions
>>
>>
>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>
>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>
>> and normalize them to length 1.
>>
>> Why should this be any different to a model which is acting
>> like a black
>> box generating vectors with 8 dimensions?
>>
>>
>>
>>
>> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>> >> What exactly do you consider real vector data? Vector data
>> which is based on texts written by humans?
>> > We have plenty of text; the problem is coming up with a
>> realistic
>> > vector model that requires as many dimensions as people
>> seem to be
>> > demanding. As I said above, after surveying huggingface I
>> couldn't
>> > find any text-based model using more than 768 dimensions.
>> So far we
>> > have some ideas of generating higher-dimensional data by
>> dithering or
>> > concatenating existing data, but it seems artificial.
>> >
>> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
>> > <michael.wechner@wyona.com> wrote:
>> >> What exactly do you consider real vector data? Vector data
>> which is based on texts written by humans?
>> >>
>> >> I am asking, because I recently attended the following
>> presentation by Anastassia Shaitarova (UZH Institute for
>> Computational Linguistics,
>> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>> >>
>> >> ----
>> >>
>> >> Can we Identify Machine-Generated Text? An Overview of
>> Current Approaches
>> >> by Anastassia Shaitarova (UZH Institute for Computational
>> Linguistics)
>> >>
>> >> The detection of machine-generated text has become
>> increasingly important due to the prevalence of automated
>> content generation and its potential for misuse. In this
>> talk, we will discuss the motivation for automatic detection
>> of generated text. We will present the currently available
>> methods, including feature-based classification as a “first
>> line-of-defense.” We will provide an overview of the
>> detection tools that have been made available so far and
>> discuss their limitations. Finally, we will reflect on some
>> open problems associated with the automatic discrimination of
>> generated texts.
>> >>
>> >> ----
>> >>
>> >> and her conclusion was that it has become basically
>> impossible to differentiate between text generated by humans
>> and text generated by for example ChatGPT.
>> >>
>> >> Whereas others have a slightly different opinion, see for
>> example
>> >>
>> >>
>> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>> >>
>> >> But I would argue that real world and synthetic have
>> become close enough that testing performance and scalability
>> of indexing should be possible with synthetic data.
>> >>
>> >> I completely agree that we have to base our discussions
>> and decisions on scientific methods and that we have to make
>> sure that Lucene performs and scales well and that we
>> understand the limits and what is going on under the hood.
>> >>
>> >> Thanks
>> >>
>> >> Michael W
>> >>
>> >>
>> >>
>> >>
>> >>
>> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>> >>
>> >> +1 to test on real vector data -- if you test on synthetic
>> data you draw synthetic conclusions.
>> >>
>> >> Can someone post the theoretical performance (CPU and RAM
>> required) of HNSW construction? Do we know/believe our HNSW
>> implementation has achieved that theoretical big-O
>> performance? Maybe we have some silly performance bug that's
>> causing it not to?
>> >>
>> >> As I understand it, HNSW makes the tradeoff of costly
>> construction for faster searching, which is typically the
>> right tradeoff for search use cases. We do this in other
>> parts of the Lucene index too.
>> >>
>> >> Lucene will do a logarithmic number of merges over time,
>> i.e. each doc will be merged O(log(N)) times in its lifetime
>> in the index. We need to multiply that by the cost of
>> re-building the whole HNSW graph on each merge. BTW, other
>> things in Lucene, like BKD/dimensional points, also rebuild
>> the whole data structure on each merge, I think? But, as Rob
>> pointed out, stored fields merging do indeed do some sneaky
>> tricks to avoid excessive block decompress/recompress on each
>> merge.
>> >>
>> >>> As I understand it, vetoes must have technical merit. I'm
>> not sure that this veto rises to "technical merit" on 2 counts:
>> >> Actually I think Robert's veto stands on its technical
>> merit already. Robert's take on technical matters very much
>> resonate with me, even if he is sometimes prickly in how he
>> expresses them ;)
>> >>
>> >> His point is that we, as a dev community, are not paying
>> enough attention to the indexing performance of our KNN algo
>> (HNSW) and implementation, and that it is reckless to
>> increase / remove limits in that state. It is indeed a
>> one-way door decision and one must confront such decisions
>> with caution, especially for such a widely used base
>> infrastructure as Lucene. We don't even advertise today in
>> our javadocs that you need XXX heap if you index vectors with
>> dimension Y, fanout X, levels Z, etc.
>> >>
>> >> RAM used during merging is unaffected by dimensionality,
>> but is affected by fanout, because the HNSW graph (not the
>> raw vectors) is memory resident, I think? Maybe we could
>> move it off-heap and let the OS manage the memory (and still
>> document the RAM requirements)? Maybe merge RAM costs should
>> be accounted for in IW's RAM buffer accounting? It is not
>> today, and there are some other things that use non-trivial
>> RAM, e.g. the doc mapping (to compress docid space when
>> deletions are reclaimed).
>> >>
>> >> When we added KNN vector testing to Lucene's nightly
>> benchmarks, the indexing time massively increased -- see
>> annotations DH and DP here:
>> https://home.apache.org/~mikemccand/lucenebench/indexing.html.
>> Nightly benchmarks now start at 6 PM and don't finish until
>> ~14.5 hours later. Of course, that is using a single thread
>> for indexing (on a box that has 128 cores!) so we produce a
>> deterministic index every night ...
>> >>
>> >> Stepping out (meta) a bit ... this discussion is precisely
>> one of the awesome benefits of the (informed) veto. It means
>> risky changes to the software, as determined by any single
>> informed developer on the project, can force a healthy
>> discussion about the problem at hand. Robert is legitimately
>> concerned about a real issue and so we should use our
>> creative energies to characterize our HNSW implementation's
>> performance, document it clearly for users, and uncover ways
>> to improve it.
>> >>
>> >> Mike McCandless
>> >>
>> >> http://blog.mikemccandless.com
>> >>
>> >>
>> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti
>> <a.benedetti@sease.io> wrote:
>> >>> I think Gus points are on target.
>> >>>
>> >>> I recommend we move this forward in this way:
>> >>> We stop any discussion and everyone interested proposes
>> an option with a motivation, then we aggregate the options
>> and we create a Vote maybe?
>> >>>
>> >>> I am also on the same page on the fact that a veto should
>> come with a clear and reasonable technical merit, which also
>> in my opinion has not come yet.
>> >>>
>> >>> I also apologise if any of my words sounded harsh or
>> personal attacks, never meant to do so.
>> >>>
>> >>> My proposed option:
>> >>>
>> >>> 1) remove the limit and potentially make it configurable,
>> >>> Motivation:
>> >>> The system administrator can enforce a limit its users
>> need to respect that it's in line with whatever the admin
>> decided to be acceptable for them.
>> >>> Default can stay the current one.
>> >>>
>> >>> That's my favourite at the moment, but I agree that
>> potentially in the future this may need to change, as we may
>> optimise the data structures for certain dimensions. I am a
>> big fan of Yagni (you aren't going to need it) so I am ok
>> we'll face a different discussion if that happens in the future.
>> >>>
>> >>>
>> >>>
>> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com>
>> wrote:
>> >>>> What I see so far:
>> >>>>
>> >>>> Much positive support for raising the limit
>> >>>> Slightly less support for removing it or making it
>> configurable
>> >>>> A single veto which argues that a (as yet undefined)
>> performance standard must be met before raising the limit
>> >>>> Hot tempers (various) making this discussion difficult
>> >>>>
>> >>>> As I understand it, vetoes must have technical merit.
>> I'm not sure that this veto rises to "technical merit" on 2
>> counts:
>> >>>>
>> >>>> No standard for the performance is given so it cannot be
>> technically met. Without hard criteria it's a moving target.
>> >>>> It appears to encode a valuation of the user's time, and
>> that valuation is really up to the user. Some users may
>> consider 2hours useless and not worth it, and others might
>> happily wait 2 hours. This is not a technical decision, it's
>> a business decision regarding the relative value of the time
>> invested vs the value of the result. If I can cure cancer by
>> indexing for a year, that might be worth it... (hyperbole of
>> course).
>> >>>>
>> >>>> Things I would consider to have technical merit that I
>> don't hear:
>> >>>>
>> >>>> Impact on the speed of **other** indexing operations.
>> (devaluation of other functionality)
>> >>>> Actual scenarios that work when the limit is low and
>> fail when the limit is high (new failure on the same data
>> with the limit raised).
>> >>>>
>> >>>> One thing that might or might not have technical merit
>> >>>>
>> >>>> If someone feels there is a lack of documentation of the
>> costs/performance implications of using large vectors,
>> possibly including reproducible benchmarks establishing the
>> scaling behavior (there seems to be disagreement on O(n) vs
>> O(n^2)).
>> >>>>
>> >>>> The users *should* know what they are getting into, but
>> if the cost is worth it to them, they should be able to pay
>> it without forking the project. If this veto causes a fork
>> that's not good.
>> >>>>
>> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov
>> <msokolov@gmail.com> wrote:
>> >>>>> We do have a dataset built from Wikipedia in
>> luceneutil. It comes in 100 and 300 dimensional varieties and
>> can easily enough generate large numbers of vector documents
>> from the articles data. To go higher we could concatenate
>> vectors from that and I believe the performance numbers would
>> be plausible.
>> >>>>>
>> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss
>> <dawid.weiss@gmail.com> wrote:
>> >>>>>> Can we set up a branch in which the limit is bumped to
>> 2048, then have
>> >>>>>> a realistic, free data set (wikipedia sample or
>> something) that has,
>> >>>>>> say, 5 million docs and vectors created using public
>> data (glove
>> >>>>>> pre-trained embeddings or the like)? We then could run
>> indexing on the
>> >>>>>> same hardware with 512, 1024 and 2048 and see what the
>> numbers, limits
>> >>>>>> and behavior actually are.
>> >>>>>>
>> >>>>>> I can help in writing this but not until after Easter.
>> >>>>>>
>> >>>>>>
>> >>>>>> Dawid
>> >>>>>>
>> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand
>> <jpountz@gmail.com> wrote:
>> >>>>>>> As Dawid pointed out earlier on this thread, this is
>> the rule for
>> >>>>>>> Apache projects: a single -1 vote on a code change is
>> a veto and
>> >>>>>>> cannot be overridden. Furthermore, Robert is one of
>> the people on this
>> >>>>>>> project who worked the most on debugging subtle bugs,
>> making Lucene
>> >>>>>>> more robust and improving our test framework, so I'm
>> listening when he
>> >>>>>>> voices quality concerns.
>> >>>>>>>
>> >>>>>>> The argument against removing/raising the limit that
>> resonates with me
>> >>>>>>> the most is that it is a one-way door. As MikeS
>> highlighted earlier on
>> >>>>>>> this thread, implementations may want to take
>> advantage of the fact
>> >>>>>>> that there is a limit at some point too. This is why
>> I don't want to
>> >>>>>>> remove the limit and would prefer a slight increase,
>> such as 2048 as
>> >>>>>>> suggested in the original issue, which would enable
>> most of the things
>> >>>>>>> that users who have been asking about raising the
>> limit would like to
>> >>>>>>> do.
>> >>>>>>>
>> >>>>>>> I agree that the merge-time memory usage and slow
>> indexing rate are
>> >>>>>>> not great. But it's still possible to index
>> multi-million vector
>> >>>>>>> datasets with a 4GB heap without hitting OOMEs
>> regardless of the
>> >>>>>>> number of dimensions, and the feedback I'm seeing is
>> that many users
>> >>>>>>> are still interested in indexing multi-million vector
>> datasets despite
>> >>>>>>> the slow indexing rate. I wish we could do better,
>> and vector indexing
>> >>>>>>> is certainly more expert than text indexing, but it
>> still is usable in
>> >>>>>>> my opinion. I understand how giving Lucene more
>> information about
>> >>>>>>> vectors prior to indexing (e.g. clustering
>> information as Jim pointed
>> >>>>>>> out) could help make merging faster and more
>> memory-efficient, but I
>> >>>>>>> would really like to avoid making it a requirement
>> for indexing
>> >>>>>>> vectors as it also makes this feature much harder to use.
>> >>>>>>>
>> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>> >>>>>>> <a.benedetti@sease.io> wrote:
>> >>>>>>>> I am very attentive to listen opinions but I am
>> un-convinced here and I an not sure that a single person
>> opinion should be allowed to be detrimental for such an
>> important project.
>> >>>>>>>>
>> >>>>>>>> The limit as far as I know is literally just raising
>> an exception.
>> >>>>>>>> Removing it won't alter in any way the current
>> performance for users in low dimensional space.
>> >>>>>>>> Removing it will just enable more users to use Lucene.
>> >>>>>>>>
>> >>>>>>>> If new users in certain situations will be unhappy
>> with the performance, they may contribute improvements.
>> >>>>>>>> This is how you make progress.
>> >>>>>>>>
>> >>>>>>>> If it's a reputation thing, trust me that not
>> allowing users to play with high dimensional space will
>> equally damage it.
>> >>>>>>>>
>> >>>>>>>> To me it's really a no brainer.
>> >>>>>>>> Removing the limit and enable people to use high
>> dimensional vectors will take minutes.
>> >>>>>>>> Improving the hnsw implementation can take months.
>> >>>>>>>> Pick one to begin with...
>> >>>>>>>>
>> >>>>>>>> And there's no-one paying me here, no company
>> interest whatsoever, actually I pay people to contribute, I
>> am just convinced it's a good idea.
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir,
>> <rcmuir@gmail.com> wrote:
>> >>>>>>>>> I disagree with your categorization. I put in
>> plenty of work and
>> >>>>>>>>> experienced plenty of pain myself, writing tests
>> and fighting these
>> >>>>>>>>> issues, after i saw that, two releases in a row,
>> vector indexing fell
>> >>>>>>>>> over and hit integer overflows etc on small datasets:
>> >>>>>>>>>
>> >>>>>>>>> https://github.com/apache/lucene/pull/11905
>> >>>>>>>>>
>> >>>>>>>>> Attacking me isn't helping the situation.
>> >>>>>>>>>
>> >>>>>>>>> PS: when i said the "one guy who wrote the code" I
>> didn't mean it in
>> >>>>>>>>> any kind of demeaning fashion really. I meant to
>> describe the current
>> >>>>>>>>> state of usability with respect to indexing a few
>> million docs with
>> >>>>>>>>> high dimensions. You can scroll up the thread and
>> see that at least
>> >>>>>>>>> one other committer on the project experienced
>> similar pain as me.
>> >>>>>>>>> Then, think about users who aren't committers
>> trying to use the
>> >>>>>>>>> functionality!
>> >>>>>>>>>
>> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov
>> <msokolov@gmail.com> wrote:
>> >>>>>>>>>> What you said about increasing dimensions
>> requiring a bigger ram buffer on merge is wrong. That's the
>> point I was trying to make. Your concerns about merge costs
>> are not wrong, but your conclusion that we need to limit
>> dimensions is not justified.
>> >>>>>>>>>>
>> >>>>>>>>>> You complain that hnsw sucks it doesn't scale, but
>> when I show it scales linearly with dimension you just ignore
>> that and complain about something entirely different.
>> >>>>>>>>>>
>> >>>>>>>>>> You demand that people run all kinds of tests to
>> prove you wrong but when they do, you don't listen and you
>> won't put in the work yourself or complain that it's too hard.
>> >>>>>>>>>>
>> >>>>>>>>>> Then you complain about people not meeting you
>> half way. Wow
>> >>>>>>>>>>
>> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir
>> <rcmuir@gmail.com> wrote:
>> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>> >>>>>>>>>>>> What exactly do you consider reasonable?
>> >>>>>>>>>>> Let's begin a real discussion by being HONEST
>> about the current
>> >>>>>>>>>>> status. Please put politically correct or your
>> own company's wishes
>> >>>>>>>>>>> aside, we know it's not in a good state.
>> >>>>>>>>>>>
>> >>>>>>>>>>> Current status is the one guy who wrote the code
>> can set a
>> >>>>>>>>>>> multi-gigabyte ram buffer and index a small
>> dataset with 1024
>> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>> >>>>>>>>>>>
>> >>>>>>>>>>> My concerns are everyone else except the one guy,
>> I want it to be
>> >>>>>>>>>>> usable. Increasing dimensions just means even
>> bigger multi-gigabyte
>> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>> >>>>>>>>>>> It is also a permanent backwards compatibility
>> decision, we have to
>> >>>>>>>>>>> support it once we do this and we can't just say
>> "oops" and flip it
>> >>>>>>>>>>> back.
>> >>>>>>>>>>>
>> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram
>> buffer is really to
>> >>>>>>>>>>> avoid merges because they are so slow and it
>> would be DAYS otherwise,
>> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>> >>>>>>>>>>> Also from personal experience, it takes trial and
>> error (means
>> >>>>>>>>>>> experiencing OOM on merge!!!) before you get
>> those heap values correct
>> >>>>>>>>>>> for your dataset. This usually means starting
>> over which is
>> >>>>>>>>>>> frustrating and wastes more time.
>> >>>>>>>>>>>
>> >>>>>>>>>>> Jim mentioned some ideas about the memory usage
>> in IndexWriter, seems
>> >>>>>>>>>>> to me like its a good idea. maybe the
>> multigigabyte ram buffer can be
>> >>>>>>>>>>> avoided in this way and performance improved by
>> writing bigger
>> >>>>>>>>>>> segments with lucene's defaults. But this doesn't
>> mean we can simply
>> >>>>>>>>>>> ignore the horrors of what happens on merge.
>> merging needs to scale so
>> >>>>>>>>>>> that indexing really scales.
>> >>>>>>>>>>>
>> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data
>> amounts and cause OOM,
>> >>>>>>>>>>> and definitely it shouldnt burn hours and hours
>> of CPU in O(n^2)
>> >>>>>>>>>>> fashion when indexing.
>> >>>>>>>>>>>
>> >>>>>>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>>>>>> To unsubscribe, e-mail:
>> dev-unsubscribe@lucene.apache.org
>> >>>>>>>>>>> For additional commands, e-mail:
>> dev-help@lucene.apache.org
>> >>>>>>>>>>>
>> >>>>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>>>> To unsubscribe, e-mail:
>> dev-unsubscribe@lucene.apache.org
>> >>>>>>>>> For additional commands, e-mail:
>> dev-help@lucene.apache.org
>> >>>>>>>>>
>> >>>>>>>
>> >>>>>>> --
>> >>>>>>> Adrien
>> >>>>>>>
>> >>>>>>>
>> ---------------------------------------------------------------------
>> >>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>>> For additional commands, e-mail:
>> dev-help@lucene.apache.org
>> >>>>>>>
>> >>>>>>
>> ---------------------------------------------------------------------
>> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> >>>>>> For additional commands, e-mail:
>> dev-help@lucene.apache.org
>> >>>>>>
>> >>>>
>> >>>> --
>> >>>> http://www.needhamsoftware.com (work)
>> >>>> http://www.the111shift.com (play)
>> >>
>> >
>> ---------------------------------------------------------------------
>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> > For additional commands, e-mail: dev-help@lucene.apache.org
>> >
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>> For additional commands, e-mail: dev-help@lucene.apache.org
>>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 20, 2023, 11:17 PM

Post #96 of 99 (941 views)

Hi Together

Cohere just published approx. 100Mio embeddings based on Wikipedia content

https://txt.cohere.com/embedding-archives-wikipedia/

resp.

https://huggingface.co/datasets/Cohere/wikipedia-22-12-en-embeddings
https://huggingface.co/datasets/Cohere/wikipedia-22-12-de-embeddings
....

HTH

Michael

Am 13.04.23 um 07:58 schrieb Michael Wechner:
> Hi Kent
>
> Great, thank you very much!
>
> Will download it later today :-)
>
> All the best
>
> Michael
>
> Am 13.04.23 um 01:35 schrieb Kent Fitch:
>> Hi Michael (and anyone else who wants just over 240K "real world"
>> ada-002 vectors of dimension 1536),
>> you are welcome to retrieve a tar.gz file which contains:
>> - 47K embeddings of Canberra Times news article text from 1994
>> - 38K embeddings of the first paragraphs of wikipedia articles about
>> organisations
>> - 156.6K embeddings of the first paragraphs of wikipedia articles
>> about people
>>
>> https://drive.google.com/file/d/13JP_5u7E8oZO6vRg0ekaTgBDQOaj-W00/view?usp=sharing
>>
>> The file is about 1.7GB and will expand to about 4.4GB. This file
>> will be accessible for at least a week, and I hope you dont hit any
>> google drive download limits trying to retrieve it.
>>
>> The embeddings were generated using my openAI account and you are
>> welcome to use them for any purpose you like.
>>
>> best wishes,
>>
>> Kent Fitch
>>
>> On Wed, Apr 12, 2023 at 4:37?PM Michael Wechner
>> <michael.wechner@wyona.com> wrote:
>>
>> thank you very much for your feedback!
>>
>> In a previous post (April 7) you wrote you could make availlable
>> the 47K ada-002 vectors, which would be great!
>>
>> Would it make sense to setup a public gitub repo, such that
>> others could use or also contribute vectors?
>>
>> Thanks
>>
>> Michael Wechner
>>
>>
>> Am 12.04.23 um 04:51 schrieb Kent Fitch:
>>> I only know some characteristics of the openAI ada-002 vectors,
>>> although they are a very popular as
>>> embeddings/text-characterisations as they allow more
>>> accurate/"human meaningful" semantic search results with fewer
>>> dimensions than their predecessors - I've evaluated a few
>>> different embedding models, including some BERT variants, CLIP
>>> ViT-L-14 (with 768 dims, which was quite good), openAI's ada-001
>>> (1024 dims) and babbage-001 (2048 dims), and ada-002 are
>>> qualitatively the best, although that will certainly change!
>>>
>>> In any case, ada-002 vectors have interesting characteristics
>>> that I think mean you could confidently create synthetic vectors
>>> which would be hard to distinguish from "real" vectors. I found
>>> this from looking at 47K ada-002 vectors generated across a full
>>> year (1994) of newspaper articles from the Canberra Times and
>>> 200K wikipedia articles:
>>> - there is no discernible/significant correlation between values
>>> in any pair of dimensions
>>> - all but 5 of the 1536 dimensions have an almost identical
>>> distribution of values shown in the central blob on these graphs
>>> (that just show a few of these 1531 dimensions with clumped
>>> values and the 5 "outlier" dimensions, but all 1531 non-outlier
>>> dims are in there, which makes for some easy quantisation from
>>> float to byte if you dont want to go the full
>>> kmeans/clustering/Lloyds-algorithm approach):
>>> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
>>> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
>>> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
>>> - the variance of the value of each dimension is characteristic:
>>> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>>>
>>> This probably represents something significant about how the
>>> ada-002 embeddings are created, but I think it also means
>>> creating "realistic" values is possible. I did not use this
>>> information when testing recall & performance on Lucene's HNSW
>>> implementation on 192m documents, as I slightly dithered the
>>> values of a "real" set on 47K docs and stored other fields in
>>> the doc that referenced the "base" document that the dithers
>>> were made from, and used different dithering magnitudes so that
>>> I could test recall with different neighbour sizes ("M"),
>>> construction-beamwidth and search-beamwidths.
>>>
>>> best regards
>>>
>>> Kent Fitch
>>>
>>>
>>>
>>>
>>> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner
>>> <michael.wechner@wyona.com> wrote:
>>>
>>> I understand what you mean that it seems to be artificial,
>>> but I don't
>>> understand why this matters to test performance and
>>> scalability of the
>>> indexing?
>>>
>>> Let's assume the limit of Lucene would be 4 instead of 1024
>>> and there
>>> are only open source models generating vectors with 4
>>> dimensions, for
>>> example
>>>
>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>>>
>>> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>
>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>>>
>>> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>
>>> and now I concatenate them to vectors with 8 dimensions
>>>
>>>
>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>
>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>
>>> and normalize them to length 1.
>>>
>>> Why should this be any different to a model which is acting
>>> like a black
>>> box generating vectors with 8 dimensions?
>>>
>>>
>>>
>>>
>>> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>>> >> What exactly do you consider real vector data? Vector
>>> data which is based on texts written by humans?
>>> > We have plenty of text; the problem is coming up with a
>>> realistic
>>> > vector model that requires as many dimensions as people
>>> seem to be
>>> > demanding. As I said above, after surveying huggingface I
>>> couldn't
>>> > find any text-based model using more than 768 dimensions.
>>> So far we
>>> > have some ideas of generating higher-dimensional data by
>>> dithering or
>>> > concatenating existing data, but it seems artificial.
>>> >
>>> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
>>> > <michael.wechner@wyona.com> wrote:
>>> >> What exactly do you consider real vector data? Vector
>>> data which is based on texts written by humans?
>>> >>
>>> >> I am asking, because I recently attended the following
>>> presentation by Anastassia Shaitarova (UZH Institute for
>>> Computational Linguistics,
>>> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>>> >>
>>> >> ----
>>> >>
>>> >> Can we Identify Machine-Generated Text? An Overview of
>>> Current Approaches
>>> >> by Anastassia Shaitarova (UZH Institute for Computational
>>> Linguistics)
>>> >>
>>> >> The detection of machine-generated text has become
>>> increasingly important due to the prevalence of automated
>>> content generation and its potential for misuse. In this
>>> talk, we will discuss the motivation for automatic detection
>>> of generated text. We will present the currently available
>>> methods, including feature-based classification as a “first
>>> line-of-defense.” We will provide an overview of the
>>> detection tools that have been made available so far and
>>> discuss their limitations. Finally, we will reflect on some
>>> open problems associated with the automatic discrimination
>>> of generated texts.
>>> >>
>>> >> ----
>>> >>
>>> >> and her conclusion was that it has become basically
>>> impossible to differentiate between text generated by humans
>>> and text generated by for example ChatGPT.
>>> >>
>>> >> Whereas others have a slightly different opinion, see for
>>> example
>>> >>
>>> >>
>>> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>>> >>
>>> >> But I would argue that real world and synthetic have
>>> become close enough that testing performance and scalability
>>> of indexing should be possible with synthetic data.
>>> >>
>>> >> I completely agree that we have to base our discussions
>>> and decisions on scientific methods and that we have to make
>>> sure that Lucene performs and scales well and that we
>>> understand the limits and what is going on under the hood.
>>> >>
>>> >> Thanks
>>> >>
>>> >> Michael W
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>>> >>
>>> >> +1 to test on real vector data -- if you test on
>>> synthetic data you draw synthetic conclusions.
>>> >>
>>> >> Can someone post the theoretical performance (CPU and RAM
>>> required) of HNSW construction? Do we know/believe our HNSW
>>> implementation has achieved that theoretical big-O
>>> performance? Maybe we have some silly performance bug
>>> that's causing it not to?
>>> >>
>>> >> As I understand it, HNSW makes the tradeoff of costly
>>> construction for faster searching, which is typically the
>>> right tradeoff for search use cases. We do this in other
>>> parts of the Lucene index too.
>>> >>
>>> >> Lucene will do a logarithmic number of merges over time,
>>> i.e. each doc will be merged O(log(N)) times in its lifetime
>>> in the index. We need to multiply that by the cost of
>>> re-building the whole HNSW graph on each merge. BTW, other
>>> things in Lucene, like BKD/dimensional points, also rebuild
>>> the whole data structure on each merge, I think? But, as Rob
>>> pointed out, stored fields merging do indeed do some sneaky
>>> tricks to avoid excessive block decompress/recompress on
>>> each merge.
>>> >>
>>> >>> As I understand it, vetoes must have technical merit.
>>> I'm not sure that this veto rises to "technical merit" on 2
>>> counts:
>>> >> Actually I think Robert's veto stands on its technical
>>> merit already. Robert's take on technical matters very much
>>> resonate with me, even if he is sometimes prickly in how he
>>> expresses them ;)
>>> >>
>>> >> His point is that we, as a dev community, are not paying
>>> enough attention to the indexing performance of our KNN algo
>>> (HNSW) and implementation, and that it is reckless to
>>> increase / remove limits in that state. It is indeed a
>>> one-way door decision and one must confront such decisions
>>> with caution, especially for such a widely used base
>>> infrastructure as Lucene. We don't even advertise today in
>>> our javadocs that you need XXX heap if you index vectors
>>> with dimension Y, fanout X, levels Z, etc.
>>> >>
>>> >> RAM used during merging is unaffected by dimensionality,
>>> but is affected by fanout, because the HNSW graph (not the
>>> raw vectors) is memory resident, I think? Maybe we could
>>> move it off-heap and let the OS manage the memory (and still
>>> document the RAM requirements)? Maybe merge RAM costs
>>> should be accounted for in IW's RAM buffer accounting? It
>>> is not today, and there are some other things that use
>>> non-trivial RAM, e.g. the doc mapping (to compress docid
>>> space when deletions are reclaimed).
>>> >>
>>> >> When we added KNN vector testing to Lucene's nightly
>>> benchmarks, the indexing time massively increased -- see
>>> annotations DH and DP here:
>>> https://home.apache.org/~mikemccand/lucenebench/indexing.html.
>>> Nightly benchmarks now start at 6 PM and don't finish until
>>> ~14.5 hours later. Of course, that is using a single thread
>>> for indexing (on a box that has 128 cores!) so we produce a
>>> deterministic index every night ...
>>> >>
>>> >> Stepping out (meta) a bit ... this discussion is
>>> precisely one of the awesome benefits of the (informed)
>>> veto. It means risky changes to the software, as determined
>>> by any single informed developer on the project, can force a
>>> healthy discussion about the problem at hand. Robert is
>>> legitimately concerned about a real issue and so we should
>>> use our creative energies to characterize our HNSW
>>> implementation's performance, document it clearly for users,
>>> and uncover ways to improve it.
>>> >>
>>> >> Mike McCandless
>>> >>
>>> >> http://blog.mikemccandless.com
>>> >>
>>> >>
>>> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti
>>> <a.benedetti@sease.io> wrote:
>>> >>> I think Gus points are on target.
>>> >>>
>>> >>> I recommend we move this forward in this way:
>>> >>> We stop any discussion and everyone interested proposes
>>> an option with a motivation, then we aggregate the options
>>> and we create a Vote maybe?
>>> >>>
>>> >>> I am also on the same page on the fact that a veto
>>> should come with a clear and reasonable technical merit,
>>> which also in my opinion has not come yet.
>>> >>>
>>> >>> I also apologise if any of my words sounded harsh or
>>> personal attacks, never meant to do so.
>>> >>>
>>> >>> My proposed option:
>>> >>>
>>> >>> 1) remove the limit and potentially make it configurable,
>>> >>> Motivation:
>>> >>> The system administrator can enforce a limit its users
>>> need to respect that it's in line with whatever the admin
>>> decided to be acceptable for them.
>>> >>> Default can stay the current one.
>>> >>>
>>> >>> That's my favourite at the moment, but I agree that
>>> potentially in the future this may need to change, as we may
>>> optimise the data structures for certain dimensions. I am a
>>> big fan of Yagni (you aren't going to need it) so I am ok
>>> we'll face a different discussion if that happens in the future.
>>> >>>
>>> >>>
>>> >>>
>>> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com>
>>> wrote:
>>> >>>> What I see so far:
>>> >>>>
>>> >>>> Much positive support for raising the limit
>>> >>>> Slightly less support for removing it or making it
>>> configurable
>>> >>>> A single veto which argues that a (as yet undefined)
>>> performance standard must be met before raising the limit
>>> >>>> Hot tempers (various) making this discussion difficult
>>> >>>>
>>> >>>> As I understand it, vetoes must have technical merit.
>>> I'm not sure that this veto rises to "technical merit" on 2
>>> counts:
>>> >>>>
>>> >>>> No standard for the performance is given so it cannot
>>> be technically met. Without hard criteria it's a moving target.
>>> >>>> It appears to encode a valuation of the user's time,
>>> and that valuation is really up to the user. Some users may
>>> consider 2hours useless and not worth it, and others might
>>> happily wait 2 hours. This is not a technical decision, it's
>>> a business decision regarding the relative value of the time
>>> invested vs the value of the result. If I can cure cancer by
>>> indexing for a year, that might be worth it... (hyperbole of
>>> course).
>>> >>>>
>>> >>>> Things I would consider to have technical merit that I
>>> don't hear:
>>> >>>>
>>> >>>> Impact on the speed of **other** indexing operations.
>>> (devaluation of other functionality)
>>> >>>> Actual scenarios that work when the limit is low and
>>> fail when the limit is high (new failure on the same data
>>> with the limit raised).
>>> >>>>
>>> >>>> One thing that might or might not have technical merit
>>> >>>>
>>> >>>> If someone feels there is a lack of documentation of
>>> the costs/performance implications of using large vectors,
>>> possibly including reproducible benchmarks establishing the
>>> scaling behavior (there seems to be disagreement on O(n) vs
>>> O(n^2)).
>>> >>>>
>>> >>>> The users *should* know what they are getting into, but
>>> if the cost is worth it to them, they should be able to pay
>>> it without forking the project. If this veto causes a fork
>>> that's not good.
>>> >>>>
>>> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov
>>> <msokolov@gmail.com> wrote:
>>> >>>>> We do have a dataset built from Wikipedia in
>>> luceneutil. It comes in 100 and 300 dimensional varieties
>>> and can easily enough generate large numbers of vector
>>> documents from the articles data. To go higher we could
>>> concatenate vectors from that and I believe the performance
>>> numbers would be plausible.
>>> >>>>>
>>> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss
>>> <dawid.weiss@gmail.com> wrote:
>>> >>>>>> Can we set up a branch in which the limit is bumped
>>> to 2048, then have
>>> >>>>>> a realistic, free data set (wikipedia sample or
>>> something) that has,
>>> >>>>>> say, 5 million docs and vectors created using public
>>> data (glove
>>> >>>>>> pre-trained embeddings or the like)? We then could
>>> run indexing on the
>>> >>>>>> same hardware with 512, 1024 and 2048 and see what
>>> the numbers, limits
>>> >>>>>> and behavior actually are.
>>> >>>>>>
>>> >>>>>> I can help in writing this but not until after Easter.
>>> >>>>>>
>>> >>>>>>
>>> >>>>>> Dawid
>>> >>>>>>
>>> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand
>>> <jpountz@gmail.com> wrote:
>>> >>>>>>> As Dawid pointed out earlier on this thread, this is
>>> the rule for
>>> >>>>>>> Apache projects: a single -1 vote on a code change
>>> is a veto and
>>> >>>>>>> cannot be overridden. Furthermore, Robert is one of
>>> the people on this
>>> >>>>>>> project who worked the most on debugging subtle
>>> bugs, making Lucene
>>> >>>>>>> more robust and improving our test framework, so I'm
>>> listening when he
>>> >>>>>>> voices quality concerns.
>>> >>>>>>>
>>> >>>>>>> The argument against removing/raising the limit that
>>> resonates with me
>>> >>>>>>> the most is that it is a one-way door. As MikeS
>>> highlighted earlier on
>>> >>>>>>> this thread, implementations may want to take
>>> advantage of the fact
>>> >>>>>>> that there is a limit at some point too. This is why
>>> I don't want to
>>> >>>>>>> remove the limit and would prefer a slight increase,
>>> such as 2048 as
>>> >>>>>>> suggested in the original issue, which would enable
>>> most of the things
>>> >>>>>>> that users who have been asking about raising the
>>> limit would like to
>>> >>>>>>> do.
>>> >>>>>>>
>>> >>>>>>> I agree that the merge-time memory usage and slow
>>> indexing rate are
>>> >>>>>>> not great. But it's still possible to index
>>> multi-million vector
>>> >>>>>>> datasets with a 4GB heap without hitting OOMEs
>>> regardless of the
>>> >>>>>>> number of dimensions, and the feedback I'm seeing is
>>> that many users
>>> >>>>>>> are still interested in indexing multi-million
>>> vector datasets despite
>>> >>>>>>> the slow indexing rate. I wish we could do better,
>>> and vector indexing
>>> >>>>>>> is certainly more expert than text indexing, but it
>>> still is usable in
>>> >>>>>>> my opinion. I understand how giving Lucene more
>>> information about
>>> >>>>>>> vectors prior to indexing (e.g. clustering
>>> information as Jim pointed
>>> >>>>>>> out) could help make merging faster and more
>>> memory-efficient, but I
>>> >>>>>>> would really like to avoid making it a requirement
>>> for indexing
>>> >>>>>>> vectors as it also makes this feature much harder to
>>> use.
>>> >>>>>>>
>>> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>> >>>>>>> <a.benedetti@sease.io> wrote:
>>> >>>>>>>> I am very attentive to listen opinions but I am
>>> un-convinced here and I an not sure that a single person
>>> opinion should be allowed to be detrimental for such an
>>> important project.
>>> >>>>>>>>
>>> >>>>>>>> The limit as far as I know is literally just
>>> raising an exception.
>>> >>>>>>>> Removing it won't alter in any way the current
>>> performance for users in low dimensional space.
>>> >>>>>>>> Removing it will just enable more users to use Lucene.
>>> >>>>>>>>
>>> >>>>>>>> If new users in certain situations will be unhappy
>>> with the performance, they may contribute improvements.
>>> >>>>>>>> This is how you make progress.
>>> >>>>>>>>
>>> >>>>>>>> If it's a reputation thing, trust me that not
>>> allowing users to play with high dimensional space will
>>> equally damage it.
>>> >>>>>>>>
>>> >>>>>>>> To me it's really a no brainer.
>>> >>>>>>>> Removing the limit and enable people to use high
>>> dimensional vectors will take minutes.
>>> >>>>>>>> Improving the hnsw implementation can take months.
>>> >>>>>>>> Pick one to begin with...
>>> >>>>>>>>
>>> >>>>>>>> And there's no-one paying me here, no company
>>> interest whatsoever, actually I pay people to contribute, I
>>> am just convinced it's a good idea.
>>> >>>>>>>>
>>> >>>>>>>>
>>> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir,
>>> <rcmuir@gmail.com> wrote:
>>> >>>>>>>>> I disagree with your categorization. I put in
>>> plenty of work and
>>> >>>>>>>>> experienced plenty of pain myself, writing tests
>>> and fighting these
>>> >>>>>>>>> issues, after i saw that, two releases in a row,
>>> vector indexing fell
>>> >>>>>>>>> over and hit integer overflows etc on small datasets:
>>> >>>>>>>>>
>>> >>>>>>>>> https://github.com/apache/lucene/pull/11905
>>> >>>>>>>>>
>>> >>>>>>>>> Attacking me isn't helping the situation.
>>> >>>>>>>>>
>>> >>>>>>>>> PS: when i said the "one guy who wrote the code" I
>>> didn't mean it in
>>> >>>>>>>>> any kind of demeaning fashion really. I meant to
>>> describe the current
>>> >>>>>>>>> state of usability with respect to indexing a few
>>> million docs with
>>> >>>>>>>>> high dimensions. You can scroll up the thread and
>>> see that at least
>>> >>>>>>>>> one other committer on the project experienced
>>> similar pain as me.
>>> >>>>>>>>> Then, think about users who aren't committers
>>> trying to use the
>>> >>>>>>>>> functionality!
>>> >>>>>>>>>
>>> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov
>>> <msokolov@gmail.com> wrote:
>>> >>>>>>>>>> What you said about increasing dimensions
>>> requiring a bigger ram buffer on merge is wrong. That's the
>>> point I was trying to make. Your concerns about merge costs
>>> are not wrong, but your conclusion that we need to limit
>>> dimensions is not justified.
>>> >>>>>>>>>>
>>> >>>>>>>>>> You complain that hnsw sucks it doesn't scale,
>>> but when I show it scales linearly with dimension you just
>>> ignore that and complain about something entirely different.
>>> >>>>>>>>>>
>>> >>>>>>>>>> You demand that people run all kinds of tests to
>>> prove you wrong but when they do, you don't listen and you
>>> won't put in the work yourself or complain that it's too hard.
>>> >>>>>>>>>>
>>> >>>>>>>>>> Then you complain about people not meeting you
>>> half way. Wow
>>> >>>>>>>>>>
>>> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir
>>> <rcmuir@gmail.com> wrote:
>>> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>>> >>>>>>>>>>>> What exactly do you consider reasonable?
>>> >>>>>>>>>>> Let's begin a real discussion by being HONEST
>>> about the current
>>> >>>>>>>>>>> status. Please put politically correct or your
>>> own company's wishes
>>> >>>>>>>>>>> aside, we know it's not in a good state.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> Current status is the one guy who wrote the code
>>> can set a
>>> >>>>>>>>>>> multi-gigabyte ram buffer and index a small
>>> dataset with 1024
>>> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> My concerns are everyone else except the one
>>> guy, I want it to be
>>> >>>>>>>>>>> usable. Increasing dimensions just means even
>>> bigger multi-gigabyte
>>> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>>> >>>>>>>>>>> It is also a permanent backwards compatibility
>>> decision, we have to
>>> >>>>>>>>>>> support it once we do this and we can't just say
>>> "oops" and flip it
>>> >>>>>>>>>>> back.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram
>>> buffer is really to
>>> >>>>>>>>>>> avoid merges because they are so slow and it
>>> would be DAYS otherwise,
>>> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>>> >>>>>>>>>>> Also from personal experience, it takes trial
>>> and error (means
>>> >>>>>>>>>>> experiencing OOM on merge!!!) before you get
>>> those heap values correct
>>> >>>>>>>>>>> for your dataset. This usually means starting
>>> over which is
>>> >>>>>>>>>>> frustrating and wastes more time.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> Jim mentioned some ideas about the memory usage
>>> in IndexWriter, seems
>>> >>>>>>>>>>> to me like its a good idea. maybe the
>>> multigigabyte ram buffer can be
>>> >>>>>>>>>>> avoided in this way and performance improved by
>>> writing bigger
>>> >>>>>>>>>>> segments with lucene's defaults. But this
>>> doesn't mean we can simply
>>> >>>>>>>>>>> ignore the horrors of what happens on merge.
>>> merging needs to scale so
>>> >>>>>>>>>>> that indexing really scales.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data
>>> amounts and cause OOM,
>>> >>>>>>>>>>> and definitely it shouldnt burn hours and hours
>>> of CPU in O(n^2)
>>> >>>>>>>>>>> fashion when indexing.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>>>>>>> To unsubscribe, e-mail:
>>> dev-unsubscribe@lucene.apache.org
>>> >>>>>>>>>>> For additional commands, e-mail:
>>> dev-help@lucene.apache.org
>>> >>>>>>>>>>>
>>> >>>>>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>>>>> To unsubscribe, e-mail:
>>> dev-unsubscribe@lucene.apache.org
>>> >>>>>>>>> For additional commands, e-mail:
>>> dev-help@lucene.apache.org
>>> >>>>>>>>>
>>> >>>>>>>
>>> >>>>>>> --
>>> >>>>>>> Adrien
>>> >>>>>>>
>>> >>>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>>> To unsubscribe, e-mail:
>>> dev-unsubscribe@lucene.apache.org
>>> >>>>>>> For additional commands, e-mail:
>>> dev-help@lucene.apache.org
>>> >>>>>>>
>>> >>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>> For additional commands, e-mail:
>>> dev-help@lucene.apache.org
>>> >>>>>>
>>> >>>>
>>> >>>> --
>>> >>>> http://www.needhamsoftware.com (work)
>>> >>>> http://www.the111shift.com (play)
>>> >>
>>> >
>>> ---------------------------------------------------------------------
>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>> >
>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>
>>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

ichattopadhyaya at gmail

Apr 21, 2023, 12:24 AM

Post #97 of 99 (941 views)

Seems like they were all 768 dimensions.

On Fri, 21 Apr, 2023, 11:48 am Michael Wechner, <michael.wechner@wyona.com>
wrote:

> Hi Together
>
> Cohere just published approx. 100Mio embeddings based on Wikipedia content
>
> https://txt.cohere.com/embedding-archives-wikipedia/
>
> resp.
>
> https://huggingface.co/datasets/Cohere/wikipedia-22-12-en-embeddings
> https://huggingface.co/datasets/Cohere/wikipedia-22-12-de-embeddings
> ....
>
> HTH
>
> Michael
>
>
>
> Am 13.04.23 um 07:58 schrieb Michael Wechner:
>
> Hi Kent
>
> Great, thank you very much!
>
> Will download it later today :-)
>
> All the best
>
> Michael
>
> Am 13.04.23 um 01:35 schrieb Kent Fitch:
>
> Hi Michael (and anyone else who wants just over 240K "real world" ada-002
> vectors of dimension 1536),
> you are welcome to retrieve a tar.gz file which contains:
> - 47K embeddings of Canberra Times news article text from 1994
> - 38K embeddings of the first paragraphs of wikipedia articles about
> organisations
> - 156.6K embeddings of the first paragraphs of wikipedia articles about
> people
>
>
> https://drive.google.com/file/d/13JP_5u7E8oZO6vRg0ekaTgBDQOaj-W00/view?usp=sharing
>
> The file is about 1.7GB and will expand to about 4.4GB. This file will be
> accessible for at least a week, and I hope you dont hit any google drive
> download limits trying to retrieve it.
>
> The embeddings were generated using my openAI account and you are welcome
> to use them for any purpose you like.
>
> best wishes,
>
> Kent Fitch
>
> On Wed, Apr 12, 2023 at 4:37?PM Michael Wechner <michael.wechner@wyona.com>
> wrote:
>
>> thank you very much for your feedback!
>>
>> In a previous post (April 7) you wrote you could make availlable the 47K
>> ada-002 vectors, which would be great!
>>
>> Would it make sense to setup a public gitub repo, such that others could
>> use or also contribute vectors?
>>
>> Thanks
>>
>> Michael Wechner
>>
>>
>> Am 12.04.23 um 04:51 schrieb Kent Fitch:
>>
>> I only know some characteristics of the openAI ada-002 vectors, although
>> they are a very popular as embeddings/text-characterisations as they allow
>> more accurate/"human meaningful" semantic search results with fewer
>> dimensions than their predecessors - I've evaluated a few different
>> embedding models, including some BERT variants, CLIP ViT-L-14 (with 768
>> dims, which was quite good), openAI's ada-001 (1024 dims) and babbage-001
>> (2048 dims), and ada-002 are qualitatively the best, although that will
>> certainly change!
>>
>> In any case, ada-002 vectors have interesting characteristics that I
>> think mean you could confidently create synthetic vectors which would be
>> hard to distinguish from "real" vectors. I found this from looking at 47K
>> ada-002 vectors generated across a full year (1994) of newspaper articles
>> from the Canberra Times and 200K wikipedia articles:
>> - there is no discernible/significant correlation between values in any
>> pair of dimensions
>> - all but 5 of the 1536 dimensions have an almost identical distribution
>> of values shown in the central blob on these graphs (that just show a few
>> of these 1531 dimensions with clumped values and the 5 "outlier"
>> dimensions, but all 1531 non-outlier dims are in there, which makes for
>> some easy quantisation from float to byte if you dont want to go the full
>> kmeans/clustering/Lloyds-algorithm approach):
>>
>> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
>>
>> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
>>
>> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
>> - the variance of the value of each dimension is characteristic:
>>
>> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>>
>> This probably represents something significant about how the ada-002
>> embeddings are created, but I think it also means creating "realistic"
>> values is possible. I did not use this information when testing recall &
>> performance on Lucene's HNSW implementation on 192m documents, as I
>> slightly dithered the values of a "real" set on 47K docs and stored other
>> fields in the doc that referenced the "base" document that the dithers were
>> made from, and used different dithering magnitudes so that I could test
>> recall with different neighbour sizes ("M"), construction-beamwidth and
>> search-beamwidths.
>>
>> best regards
>>
>> Kent Fitch
>>
>>
>>
>>
>> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner <
>> michael.wechner@wyona.com> wrote:
>>
>>> I understand what you mean that it seems to be artificial, but I don't
>>> understand why this matters to test performance and scalability of the
>>> indexing?
>>>
>>> Let's assume the limit of Lucene would be 4 instead of 1024 and there
>>> are only open source models generating vectors with 4 dimensions, for
>>> example
>>>
>>>
>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>>>
>>>
>>> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>
>>>
>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>>>
>>>
>>> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>
>>> and now I concatenate them to vectors with 8 dimensions
>>>
>>>
>>>
>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>
>>>
>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>
>>> and normalize them to length 1.
>>>
>>> Why should this be any different to a model which is acting like a black
>>> box generating vectors with 8 dimensions?
>>>
>>>
>>>
>>>
>>> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>>> >> What exactly do you consider real vector data? Vector data which is
>>> based on texts written by humans?
>>> > We have plenty of text; the problem is coming up with a realistic
>>> > vector model that requires as many dimensions as people seem to be
>>> > demanding. As I said above, after surveying huggingface I couldn't
>>> > find any text-based model using more than 768 dimensions. So far we
>>> > have some ideas of generating higher-dimensional data by dithering or
>>> > concatenating existing data, but it seems artificial.
>>> >
>>> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
>>> > <michael.wechner@wyona.com> wrote:
>>> >> What exactly do you consider real vector data? Vector data which is
>>> based on texts written by humans?
>>> >>
>>> >> I am asking, because I recently attended the following presentation
>>> by Anastassia Shaitarova (UZH Institute for Computational Linguistics,
>>> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>>> >>
>>> >> ----
>>> >>
>>> >> Can we Identify Machine-Generated Text? An Overview of Current
>>> Approaches
>>> >> by Anastassia Shaitarova (UZH Institute for Computational Linguistics)
>>> >>
>>> >> The detection of machine-generated text has become increasingly
>>> important due to the prevalence of automated content generation and its
>>> potential for misuse. In this talk, we will discuss the motivation for
>>> automatic detection of generated text. We will present the currently
>>> available methods, including feature-based classification as a “first
>>> line-of-defense.” We will provide an overview of the detection tools that
>>> have been made available so far and discuss their limitations. Finally, we
>>> will reflect on some open problems associated with the automatic
>>> discrimination of generated texts.
>>> >>
>>> >> ----
>>> >>
>>> >> and her conclusion was that it has become basically impossible to
>>> differentiate between text generated by humans and text generated by for
>>> example ChatGPT.
>>> >>
>>> >> Whereas others have a slightly different opinion, see for example
>>> >>
>>> >> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>>> >>
>>> >> But I would argue that real world and synthetic have become close
>>> enough that testing performance and scalability of indexing should be
>>> possible with synthetic data.
>>> >>
>>> >> I completely agree that we have to base our discussions and decisions
>>> on scientific methods and that we have to make sure that Lucene performs
>>> and scales well and that we understand the limits and what is going on
>>> under the hood.
>>> >>
>>> >> Thanks
>>> >>
>>> >> Michael W
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>>> >>
>>> >> +1 to test on real vector data -- if you test on synthetic data you
>>> draw synthetic conclusions.
>>> >>
>>> >> Can someone post the theoretical performance (CPU and RAM required)
>>> of HNSW construction? Do we know/believe our HNSW implementation has
>>> achieved that theoretical big-O performance? Maybe we have some silly
>>> performance bug that's causing it not to?
>>> >>
>>> >> As I understand it, HNSW makes the tradeoff of costly construction
>>> for faster searching, which is typically the right tradeoff for search use
>>> cases. We do this in other parts of the Lucene index too.
>>> >>
>>> >> Lucene will do a logarithmic number of merges over time, i.e. each
>>> doc will be merged O(log(N)) times in its lifetime in the index. We need
>>> to multiply that by the cost of re-building the whole HNSW graph on each
>>> merge. BTW, other things in Lucene, like BKD/dimensional points, also
>>> rebuild the whole data structure on each merge, I think? But, as Rob
>>> pointed out, stored fields merging do indeed do some sneaky tricks to avoid
>>> excessive block decompress/recompress on each merge.
>>> >>
>>> >>> As I understand it, vetoes must have technical merit. I'm not sure
>>> that this veto rises to "technical merit" on 2 counts:
>>> >> Actually I think Robert's veto stands on its technical merit
>>> already. Robert's take on technical matters very much resonate with me,
>>> even if he is sometimes prickly in how he expresses them ;)
>>> >>
>>> >> His point is that we, as a dev community, are not paying enough
>>> attention to the indexing performance of our KNN algo (HNSW) and
>>> implementation, and that it is reckless to increase / remove limits in that
>>> state. It is indeed a one-way door decision and one must confront such
>>> decisions with caution, especially for such a widely used base
>>> infrastructure as Lucene. We don't even advertise today in our javadocs
>>> that you need XXX heap if you index vectors with dimension Y, fanout X,
>>> levels Z, etc.
>>> >>
>>> >> RAM used during merging is unaffected by dimensionality, but is
>>> affected by fanout, because the HNSW graph (not the raw vectors) is memory
>>> resident, I think? Maybe we could move it off-heap and let the OS manage
>>> the memory (and still document the RAM requirements)? Maybe merge RAM
>>> costs should be accounted for in IW's RAM buffer accounting? It is not
>>> today, and there are some other things that use non-trivial RAM, e.g. the
>>> doc mapping (to compress docid space when deletions are reclaimed).
>>> >>
>>> >> When we added KNN vector testing to Lucene's nightly benchmarks, the
>>> indexing time massively increased -- see annotations DH and DP here:
>>> https://home.apache.org/~mikemccand/lucenebench/indexing.html. Nightly
>>> benchmarks now start at 6 PM and don't finish until ~14.5 hours later. Of
>>> course, that is using a single thread for indexing (on a box that has 128
>>> cores!) so we produce a deterministic index every night ...
>>> >>
>>> >> Stepping out (meta) a bit ... this discussion is precisely one of the
>>> awesome benefits of the (informed) veto. It means risky changes to the
>>> software, as determined by any single informed developer on the project,
>>> can force a healthy discussion about the problem at hand. Robert is
>>> legitimately concerned about a real issue and so we should use our creative
>>> energies to characterize our HNSW implementation's performance, document it
>>> clearly for users, and uncover ways to improve it.
>>> >>
>>> >> Mike McCandless
>>> >>
>>> >> http://blog.mikemccandless.com
>>> >>
>>> >>
>>> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti <
>>> a.benedetti@sease.io> wrote:
>>> >>> I think Gus points are on target.
>>> >>>
>>> >>> I recommend we move this forward in this way:
>>> >>> We stop any discussion and everyone interested proposes an option
>>> with a motivation, then we aggregate the options and we create a Vote maybe?
>>> >>>
>>> >>> I am also on the same page on the fact that a veto should come with
>>> a clear and reasonable technical merit, which also in my opinion has not
>>> come yet.
>>> >>>
>>> >>> I also apologise if any of my words sounded harsh or personal
>>> attacks, never meant to do so.
>>> >>>
>>> >>> My proposed option:
>>> >>>
>>> >>> 1) remove the limit and potentially make it configurable,
>>> >>> Motivation:
>>> >>> The system administrator can enforce a limit its users need to
>>> respect that it's in line with whatever the admin decided to be acceptable
>>> for them.
>>> >>> Default can stay the current one.
>>> >>>
>>> >>> That's my favourite at the moment, but I agree that potentially in
>>> the future this may need to change, as we may optimise the data structures
>>> for certain dimensions. I am a big fan of Yagni (you aren't going to need
>>> it) so I am ok we'll face a different discussion if that happens in the
>>> future.
>>> >>>
>>> >>>
>>> >>>
>>> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck, <gus.heck@gmail.com> wrote:
>>> >>>> What I see so far:
>>> >>>>
>>> >>>> Much positive support for raising the limit
>>> >>>> Slightly less support for removing it or making it configurable
>>> >>>> A single veto which argues that a (as yet undefined) performance
>>> standard must be met before raising the limit
>>> >>>> Hot tempers (various) making this discussion difficult
>>> >>>>
>>> >>>> As I understand it, vetoes must have technical merit. I'm not sure
>>> that this veto rises to "technical merit" on 2 counts:
>>> >>>>
>>> >>>> No standard for the performance is given so it cannot be
>>> technically met. Without hard criteria it's a moving target.
>>> >>>> It appears to encode a valuation of the user's time, and that
>>> valuation is really up to the user. Some users may consider 2hours useless
>>> and not worth it, and others might happily wait 2 hours. This is not a
>>> technical decision, it's a business decision regarding the relative value
>>> of the time invested vs the value of the result. If I can cure cancer by
>>> indexing for a year, that might be worth it... (hyperbole of course).
>>> >>>>
>>> >>>> Things I would consider to have technical merit that I don't hear:
>>> >>>>
>>> >>>> Impact on the speed of **other** indexing operations. (devaluation
>>> of other functionality)
>>> >>>> Actual scenarios that work when the limit is low and fail when the
>>> limit is high (new failure on the same data with the limit raised).
>>> >>>>
>>> >>>> One thing that might or might not have technical merit
>>> >>>>
>>> >>>> If someone feels there is a lack of documentation of the
>>> costs/performance implications of using large vectors, possibly including
>>> reproducible benchmarks establishing the scaling behavior (there seems to
>>> be disagreement on O(n) vs O(n^2)).
>>> >>>>
>>> >>>> The users *should* know what they are getting into, but if the cost
>>> is worth it to them, they should be able to pay it without forking the
>>> project. If this veto causes a fork that's not good.
>>> >>>>
>>> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov <msokolov@gmail.com>
>>> wrote:
>>> >>>>> We do have a dataset built from Wikipedia in luceneutil. It comes
>>> in 100 and 300 dimensional varieties and can easily enough generate large
>>> numbers of vector documents from the articles data. To go higher we could
>>> concatenate vectors from that and I believe the performance numbers would
>>> be plausible.
>>> >>>>>
>>> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss <dawid.weiss@gmail.com>
>>> wrote:
>>> >>>>>> Can we set up a branch in which the limit is bumped to 2048, then
>>> have
>>> >>>>>> a realistic, free data set (wikipedia sample or something) that
>>> has,
>>> >>>>>> say, 5 million docs and vectors created using public data (glove
>>> >>>>>> pre-trained embeddings or the like)? We then could run indexing
>>> on the
>>> >>>>>> same hardware with 512, 1024 and 2048 and see what the numbers,
>>> limits
>>> >>>>>> and behavior actually are.
>>> >>>>>>
>>> >>>>>> I can help in writing this but not until after Easter.
>>> >>>>>>
>>> >>>>>>
>>> >>>>>> Dawid
>>> >>>>>>
>>> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand <jpountz@gmail.com>
>>> wrote:
>>> >>>>>>> As Dawid pointed out earlier on this thread, this is the rule for
>>> >>>>>>> Apache projects: a single -1 vote on a code change is a veto and
>>> >>>>>>> cannot be overridden. Furthermore, Robert is one of the people
>>> on this
>>> >>>>>>> project who worked the most on debugging subtle bugs, making
>>> Lucene
>>> >>>>>>> more robust and improving our test framework, so I'm listening
>>> when he
>>> >>>>>>> voices quality concerns.
>>> >>>>>>>
>>> >>>>>>> The argument against removing/raising the limit that resonates
>>> with me
>>> >>>>>>> the most is that it is a one-way door. As MikeS highlighted
>>> earlier on
>>> >>>>>>> this thread, implementations may want to take advantage of the
>>> fact
>>> >>>>>>> that there is a limit at some point too. This is why I don't
>>> want to
>>> >>>>>>> remove the limit and would prefer a slight increase, such as
>>> 2048 as
>>> >>>>>>> suggested in the original issue, which would enable most of the
>>> things
>>> >>>>>>> that users who have been asking about raising the limit would
>>> like to
>>> >>>>>>> do.
>>> >>>>>>>
>>> >>>>>>> I agree that the merge-time memory usage and slow indexing rate
>>> are
>>> >>>>>>> not great. But it's still possible to index multi-million vector
>>> >>>>>>> datasets with a 4GB heap without hitting OOMEs regardless of the
>>> >>>>>>> number of dimensions, and the feedback I'm seeing is that many
>>> users
>>> >>>>>>> are still interested in indexing multi-million vector datasets
>>> despite
>>> >>>>>>> the slow indexing rate. I wish we could do better, and vector
>>> indexing
>>> >>>>>>> is certainly more expert than text indexing, but it still is
>>> usable in
>>> >>>>>>> my opinion. I understand how giving Lucene more information about
>>> >>>>>>> vectors prior to indexing (e.g. clustering information as Jim
>>> pointed
>>> >>>>>>> out) could help make merging faster and more memory-efficient,
>>> but I
>>> >>>>>>> would really like to avoid making it a requirement for indexing
>>> >>>>>>> vectors as it also makes this feature much harder to use.
>>> >>>>>>>
>>> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>> >>>>>>> <a.benedetti@sease.io> wrote:
>>> >>>>>>>> I am very attentive to listen opinions but I am un-convinced
>>> here and I an not sure that a single person opinion should be allowed to be
>>> detrimental for such an important project.
>>> >>>>>>>>
>>> >>>>>>>> The limit as far as I know is literally just raising an
>>> exception.
>>> >>>>>>>> Removing it won't alter in any way the current performance for
>>> users in low dimensional space.
>>> >>>>>>>> Removing it will just enable more users to use Lucene.
>>> >>>>>>>>
>>> >>>>>>>> If new users in certain situations will be unhappy with the
>>> performance, they may contribute improvements.
>>> >>>>>>>> This is how you make progress.
>>> >>>>>>>>
>>> >>>>>>>> If it's a reputation thing, trust me that not allowing users to
>>> play with high dimensional space will equally damage it.
>>> >>>>>>>>
>>> >>>>>>>> To me it's really a no brainer.
>>> >>>>>>>> Removing the limit and enable people to use high dimensional
>>> vectors will take minutes.
>>> >>>>>>>> Improving the hnsw implementation can take months.
>>> >>>>>>>> Pick one to begin with...
>>> >>>>>>>>
>>> >>>>>>>> And there's no-one paying me here, no company interest
>>> whatsoever, actually I pay people to contribute, I am just convinced it's a
>>> good idea.
>>> >>>>>>>>
>>> >>>>>>>>
>>> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir, <rcmuir@gmail.com>
>>> wrote:
>>> >>>>>>>>> I disagree with your categorization. I put in plenty of work
>>> and
>>> >>>>>>>>> experienced plenty of pain myself, writing tests and fighting
>>> these
>>> >>>>>>>>> issues, after i saw that, two releases in a row, vector
>>> indexing fell
>>> >>>>>>>>> over and hit integer overflows etc on small datasets:
>>> >>>>>>>>>
>>> >>>>>>>>> https://github.com/apache/lucene/pull/11905
>>> >>>>>>>>>
>>> >>>>>>>>> Attacking me isn't helping the situation.
>>> >>>>>>>>>
>>> >>>>>>>>> PS: when i said the "one guy who wrote the code" I didn't mean
>>> it in
>>> >>>>>>>>> any kind of demeaning fashion really. I meant to describe the
>>> current
>>> >>>>>>>>> state of usability with respect to indexing a few million docs
>>> with
>>> >>>>>>>>> high dimensions. You can scroll up the thread and see that at
>>> least
>>> >>>>>>>>> one other committer on the project experienced similar pain as
>>> me.
>>> >>>>>>>>> Then, think about users who aren't committers trying to use the
>>> >>>>>>>>> functionality!
>>> >>>>>>>>>
>>> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael Sokolov <
>>> msokolov@gmail.com> wrote:
>>> >>>>>>>>>> What you said about increasing dimensions requiring a bigger
>>> ram buffer on merge is wrong. That's the point I was trying to make. Your
>>> concerns about merge costs are not wrong, but your conclusion that we need
>>> to limit dimensions is not justified.
>>> >>>>>>>>>>
>>> >>>>>>>>>> You complain that hnsw sucks it doesn't scale, but when I
>>> show it scales linearly with dimension you just ignore that and complain
>>> about something entirely different.
>>> >>>>>>>>>>
>>> >>>>>>>>>> You demand that people run all kinds of tests to prove you
>>> wrong but when they do, you don't listen and you won't put in the work
>>> yourself or complain that it's too hard.
>>> >>>>>>>>>>
>>> >>>>>>>>>> Then you complain about people not meeting you half way. Wow
>>> >>>>>>>>>>
>>> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir <rcmuir@gmail.com>
>>> wrote:
>>> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>>> >>>>>>>>>>>> What exactly do you consider reasonable?
>>> >>>>>>>>>>> Let's begin a real discussion by being HONEST about the
>>> current
>>> >>>>>>>>>>> status. Please put politically correct or your own company's
>>> wishes
>>> >>>>>>>>>>> aside, we know it's not in a good state.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> Current status is the one guy who wrote the code can set a
>>> >>>>>>>>>>> multi-gigabyte ram buffer and index a small dataset with 1024
>>> >>>>>>>>>>> dimensions in HOURS (i didn't ask what hardware).
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> My concerns are everyone else except the one guy, I want it
>>> to be
>>> >>>>>>>>>>> usable. Increasing dimensions just means even bigger
>>> multi-gigabyte
>>> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on merge.
>>> >>>>>>>>>>> It is also a permanent backwards compatibility decision, we
>>> have to
>>> >>>>>>>>>>> support it once we do this and we can't just say "oops" and
>>> flip it
>>> >>>>>>>>>>> back.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte ram buffer is
>>> really to
>>> >>>>>>>>>>> avoid merges because they are so slow and it would be DAYS
>>> otherwise,
>>> >>>>>>>>>>> or if its to avoid merges so it doesn't hit OOM.
>>> >>>>>>>>>>> Also from personal experience, it takes trial and error
>>> (means
>>> >>>>>>>>>>> experiencing OOM on merge!!!) before you get those heap
>>> values correct
>>> >>>>>>>>>>> for your dataset. This usually means starting over which is
>>> >>>>>>>>>>> frustrating and wastes more time.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> Jim mentioned some ideas about the memory usage in
>>> IndexWriter, seems
>>> >>>>>>>>>>> to me like its a good idea. maybe the multigigabyte ram
>>> buffer can be
>>> >>>>>>>>>>> avoided in this way and performance improved by writing
>>> bigger
>>> >>>>>>>>>>> segments with lucene's defaults. But this doesn't mean we
>>> can simply
>>> >>>>>>>>>>> ignore the horrors of what happens on merge. merging needs
>>> to scale so
>>> >>>>>>>>>>> that indexing really scales.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>> At least it shouldnt spike RAM on trivial data amounts and
>>> cause OOM,
>>> >>>>>>>>>>> and definitely it shouldnt burn hours and hours of CPU in
>>> O(n^2)
>>> >>>>>>>>>>> fashion when indexing.
>>> >>>>>>>>>>>
>>> >>>>>>>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>>>>>>
>>> >>>>>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>>>>
>>> >>>>>>>
>>> >>>>>>> --
>>> >>>>>>> Adrien
>>> >>>>>>>
>>> >>>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>>
>>> >>>>>>
>>> ---------------------------------------------------------------------
>>> >>>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> >>>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>> >>>>>>
>>> >>>>
>>> >>>> --
>>> >>>> http://www.needhamsoftware.com (work)
>>> >>>> http://www.the111shift.com (play)
>>> >>
>>> > ---------------------------------------------------------------------
>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> > For additional commands, e-mail: dev-help@lucene.apache.org
>>> >
>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>
>>>
>>
>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

Apr 21, 2023, 12:30 AM

Post #98 of 99 (941 views)

yes, they are, whereas it should help us to test performance and
scalability :-)

Am 21.04.23 um 09:24 schrieb Ishan Chattopadhyaya:
> Seems like they were all 768 dimensions.
>
> On Fri, 21 Apr, 2023, 11:48 am Michael Wechner,
> <michael.wechner@wyona.com> wrote:
>
> Hi Together
>
> Cohere just published approx. 100Mio embeddings based on Wikipedia
> content
>
> https://txt.cohere.com/embedding-archives-wikipedia/
>
> resp.
>
> https://huggingface.co/datasets/Cohere/wikipedia-22-12-en-embeddings
> https://huggingface.co/datasets/Cohere/wikipedia-22-12-de-embeddings
> ....
>
> HTH
>
> Michael
>
>
>
> Am 13.04.23 um 07:58 schrieb Michael Wechner:
>> Hi Kent
>>
>> Great, thank you very much!
>>
>> Will download it later today :-)
>>
>> All the best
>>
>> Michael
>>
>> Am 13.04.23 um 01:35 schrieb Kent Fitch:
>>> Hi Michael (and anyone else who wants just over 240K "real
>>> world" ada-002 vectors of dimension 1536),
>>> you are welcome to retrieve a tar.gz file which contains:
>>> - 47K embeddings of Canberra Times news article text from 1994
>>> - 38K embeddings of the first paragraphs of wikipedia articles
>>> about organisations
>>> - 156.6K embeddings of the first paragraphs of wikipedia
>>> articles about people
>>>
>>> https://drive.google.com/file/d/13JP_5u7E8oZO6vRg0ekaTgBDQOaj-W00/view?usp=sharing
>>>
>>> The file is about 1.7GB and will expand to about 4.4GB. This
>>> file will be accessible for at least a week, and I hope you dont
>>> hit any google drive download limits trying to retrieve it.
>>>
>>> The embeddings were generated using my openAI account and you
>>> are welcome to use them for any purpose you like.
>>>
>>> best wishes,
>>>
>>> Kent Fitch
>>>
>>> On Wed, Apr 12, 2023 at 4:37?PM Michael Wechner
>>> <michael.wechner@wyona.com> wrote:
>>>
>>> thank you very much for your feedback!
>>>
>>> In a previous post (April 7) you wrote you could make
>>> availlable the 47K ada-002 vectors, which would be great!
>>>
>>> Would it make sense to setup a public gitub repo, such that
>>> others could use or also contribute vectors?
>>>
>>> Thanks
>>>
>>> Michael Wechner
>>>
>>>
>>> Am 12.04.23 um 04:51 schrieb Kent Fitch:
>>>> I only know some characteristics of the openAI ada-002
>>>> vectors, although they are a very popular as
>>>> embeddings/text-characterisations as they allow more
>>>> accurate/"human meaningful" semantic search results with
>>>> fewer dimensions than their predecessors - I've evaluated a
>>>> few different embedding models, including some BERT
>>>> variants, CLIP ViT-L-14 (with 768 dims, which was quite
>>>> good), openAI's ada-001 (1024 dims) and babbage-001 (2048
>>>> dims), and ada-002 are qualitatively the best, although
>>>> that will certainly change!
>>>>
>>>> In any case, ada-002 vectors have interesting
>>>> characteristics that I think mean you could confidently
>>>> create synthetic vectors which would be hard to distinguish
>>>> from "real" vectors. I found this from looking at 47K
>>>> ada-002 vectors generated across a full year (1994) of
>>>> newspaper articles from the Canberra Times and 200K
>>>> wikipedia articles:
>>>> - there is no discernible/significant correlation between
>>>> values in any pair of dimensions
>>>> - all but 5 of the 1536 dimensions have an almost identical
>>>> distribution of values shown in the central blob on these
>>>> graphs (that just show a few of these 1531 dimensions with
>>>> clumped values and the 5 "outlier" dimensions, but all 1531
>>>> non-outlier dims are in there, which makes for some easy
>>>> quantisation from float to byte if you dont want to go the
>>>> full kmeans/clustering/Lloyds-algorithm approach):
>>>> https://docs.google.com/spreadsheets/d/1DyyBCbirETZSUAEGcMK__mfbUNzsU_L48V9E0SyJYGg/edit?usp=sharing
>>>> https://docs.google.com/spreadsheets/d/1czEAlzYdyKa6xraRLesXjNZvEzlj27TcDGiEFS1-MPs/edit?usp=sharing
>>>> https://docs.google.com/spreadsheets/d/1RxTjV7Sj14etCNLk1GB-m44CXJVKdXaFlg2Y6yvj3z4/edit?usp=sharing
>>>> - the variance of the value of each dimension is
>>>> characteristic:
>>>> https://docs.google.com/spreadsheets/d/1w5LnRUXt1cRzI9Qwm07LZ6UfszjMOgPaJot9cOGLHok/edit#gid=472178228
>>>>
>>>> This probably represents something significant about how
>>>> the ada-002 embeddings are created, but I think it also
>>>> means creating "realistic" values is possible. I did not
>>>> use this information when testing recall & performance on
>>>> Lucene's HNSW implementation on 192m documents, as I
>>>> slightly dithered the values of a "real" set on 47K docs
>>>> and stored other fields in the doc that referenced the
>>>> "base" document that the dithers were made from, and used
>>>> different dithering magnitudes so that I could test recall
>>>> with different neighbour sizes ("M"),
>>>> construction-beamwidth and search-beamwidths.
>>>>
>>>> best regards
>>>>
>>>> Kent Fitch
>>>>
>>>>
>>>>
>>>>
>>>> On Wed, Apr 12, 2023 at 5:08?AM Michael Wechner
>>>> <michael.wechner@wyona.com> wrote:
>>>>
>>>> I understand what you mean that it seems to be
>>>> artificial, but I don't
>>>> understand why this matters to test performance and
>>>> scalability of the
>>>> indexing?
>>>>
>>>> Let's assume the limit of Lucene would be 4 instead of
>>>> 1024 and there
>>>> are only open source models generating vectors with 4
>>>> dimensions, for
>>>> example
>>>>
>>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814
>>>>
>>>> 0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>>
>>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106
>>>>
>>>> -0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>>
>>>> and now I concatenate them to vectors with 8 dimensions
>>>>
>>>>
>>>> 0.02150459587574005,0.11223817616701126,-0.007903356105089188,0.03795722872018814,0.026009393855929375,0.006306684575974941,0.020492585375905037,-0.029064252972602844
>>>>
>>>> -0.08239810913801193,-0.01947402022778988,0.03827739879488945,-0.020566290244460106,-0.007012288551777601,-0.026665858924388885,0.044495150446891785,-0.038030195981264114
>>>>
>>>> and normalize them to length 1.
>>>>
>>>> Why should this be any different to a model which is
>>>> acting like a black
>>>> box generating vectors with 8 dimensions?
>>>>
>>>>
>>>>
>>>>
>>>> Am 11.04.23 um 19:05 schrieb Michael Sokolov:
>>>> >> What exactly do you consider real vector data?
>>>> Vector data which is based on texts written by humans?
>>>> > We have plenty of text; the problem is coming up with
>>>> a realistic
>>>> > vector model that requires as many dimensions as
>>>> people seem to be
>>>> > demanding. As I said above, after surveying
>>>> huggingface I couldn't
>>>> > find any text-based model using more than 768
>>>> dimensions. So far we
>>>> > have some ideas of generating higher-dimensional data
>>>> by dithering or
>>>> > concatenating existing data, but it seems artificial.
>>>> >
>>>> > On Tue, Apr 11, 2023 at 9:31?AM Michael Wechner
>>>> > <michael.wechner@wyona.com> wrote:
>>>> >> What exactly do you consider real vector data?
>>>> Vector data which is based on texts written by humans?
>>>> >>
>>>> >> I am asking, because I recently attended the
>>>> following presentation by Anastassia Shaitarova (UZH
>>>> Institute for Computational Linguistics,
>>>> https://www.cl.uzh.ch/de/people/team/compling/shaitarova.html)
>>>> >>
>>>> >> ----
>>>> >>
>>>> >> Can we Identify Machine-Generated Text? An Overview
>>>> of Current Approaches
>>>> >> by Anastassia Shaitarova (UZH Institute for
>>>> Computational Linguistics)
>>>> >>
>>>> >> The detection of machine-generated text has become
>>>> increasingly important due to the prevalence of
>>>> automated content generation and its potential for
>>>> misuse. In this talk, we will discuss the motivation
>>>> for automatic detection of generated text. We will
>>>> present the currently available methods, including
>>>> feature-based classification as a “first
>>>> line-of-defense.” We will provide an overview of the
>>>> detection tools that have been made available so far
>>>> and discuss their limitations. Finally, we will reflect
>>>> on some open problems associated with the automatic
>>>> discrimination of generated texts.
>>>> >>
>>>> >> ----
>>>> >>
>>>> >> and her conclusion was that it has become basically
>>>> impossible to differentiate between text generated by
>>>> humans and text generated by for example ChatGPT.
>>>> >>
>>>> >> Whereas others have a slightly different opinion,
>>>> see for example
>>>> >>
>>>> >>
>>>> https://www.wired.com/story/how-to-spot-generative-ai-text-chatgpt/
>>>> >>
>>>> >> But I would argue that real world and synthetic have
>>>> become close enough that testing performance and
>>>> scalability of indexing should be possible with
>>>> synthetic data.
>>>> >>
>>>> >> I completely agree that we have to base our
>>>> discussions and decisions on scientific methods and
>>>> that we have to make sure that Lucene performs and
>>>> scales well and that we understand the limits and what
>>>> is going on under the hood.
>>>> >>
>>>> >> Thanks
>>>> >>
>>>> >> Michael W
>>>> >>
>>>> >>
>>>> >>
>>>> >>
>>>> >>
>>>> >> Am 11.04.23 um 14:29 schrieb Michael McCandless:
>>>> >>
>>>> >> +1 to test on real vector data -- if you test on
>>>> synthetic data you draw synthetic conclusions.
>>>> >>
>>>> >> Can someone post the theoretical performance (CPU
>>>> and RAM required) of HNSW construction? Do we
>>>> know/believe our HNSW implementation has achieved that
>>>> theoretical big-O performance? Maybe we have some
>>>> silly performance bug that's causing it not to?
>>>> >>
>>>> >> As I understand it, HNSW makes the tradeoff of
>>>> costly construction for faster searching, which is
>>>> typically the right tradeoff for search use cases. We
>>>> do this in other parts of the Lucene index too.
>>>> >>
>>>> >> Lucene will do a logarithmic number of merges over
>>>> time, i.e. each doc will be merged O(log(N)) times in
>>>> its lifetime in the index. We need to multiply that by
>>>> the cost of re-building the whole HNSW graph on each
>>>> merge. BTW, other things in Lucene, like
>>>> BKD/dimensional points, also rebuild the whole data
>>>> structure on each merge, I think? But, as Rob pointed
>>>> out, stored fields merging do indeed do some sneaky
>>>> tricks to avoid excessive block decompress/recompress
>>>> on each merge.
>>>> >>
>>>> >>> As I understand it, vetoes must have technical
>>>> merit. I'm not sure that this veto rises to "technical
>>>> merit" on 2 counts:
>>>> >> Actually I think Robert's veto stands on its
>>>> technical merit already. Robert's take on technical
>>>> matters very much resonate with me, even if he is
>>>> sometimes prickly in how he expresses them ;)
>>>> >>
>>>> >> His point is that we, as a dev community, are not
>>>> paying enough attention to the indexing performance of
>>>> our KNN algo (HNSW) and implementation, and that it is
>>>> reckless to increase / remove limits in that state. It
>>>> is indeed a one-way door decision and one must confront
>>>> such decisions with caution, especially for such a
>>>> widely used base infrastructure as Lucene. We don't
>>>> even advertise today in our javadocs that you need XXX
>>>> heap if you index vectors with dimension Y, fanout X,
>>>> levels Z, etc.
>>>> >>
>>>> >> RAM used during merging is unaffected by
>>>> dimensionality, but is affected by fanout, because the
>>>> HNSW graph (not the raw vectors) is memory resident, I
>>>> think? Maybe we could move it off-heap and let the OS
>>>> manage the memory (and still document the RAM
>>>> requirements)? Maybe merge RAM costs should be
>>>> accounted for in IW's RAM buffer accounting? It is not
>>>> today, and there are some other things that use
>>>> non-trivial RAM, e.g. the doc mapping (to compress
>>>> docid space when deletions are reclaimed).
>>>> >>
>>>> >> When we added KNN vector testing to Lucene's nightly
>>>> benchmarks, the indexing time massively increased --
>>>> see annotations DH and DP here:
>>>> https://home.apache.org/~mikemccand/lucenebench/indexing.html.
>>>> Nightly benchmarks now start at 6 PM and don't finish
>>>> until ~14.5 hours later. Of course, that is using a
>>>> single thread for indexing (on a box that has 128
>>>> cores!) so we produce a deterministic index every night ...
>>>> >>
>>>> >> Stepping out (meta) a bit ... this discussion is
>>>> precisely one of the awesome benefits of the (informed)
>>>> veto. It means risky changes to the software, as
>>>> determined by any single informed developer on the
>>>> project, can force a healthy discussion about the
>>>> problem at hand. Robert is legitimately concerned
>>>> about a real issue and so we should use our creative
>>>> energies to characterize our HNSW implementation's
>>>> performance, document it clearly for users, and uncover
>>>> ways to improve it.
>>>> >>
>>>> >> Mike McCandless
>>>> >>
>>>> >> http://blog.mikemccandless.com
>>>> >>
>>>> >>
>>>> >> On Mon, Apr 10, 2023 at 6:41?PM Alessandro Benedetti
>>>> <a.benedetti@sease.io> wrote:
>>>> >>> I think Gus points are on target.
>>>> >>>
>>>> >>> I recommend we move this forward in this way:
>>>> >>> We stop any discussion and everyone interested
>>>> proposes an option with a motivation, then we aggregate
>>>> the options and we create a Vote maybe?
>>>> >>>
>>>> >>> I am also on the same page on the fact that a veto
>>>> should come with a clear and reasonable technical
>>>> merit, which also in my opinion has not come yet.
>>>> >>>
>>>> >>> I also apologise if any of my words sounded harsh
>>>> or personal attacks, never meant to do so.
>>>> >>>
>>>> >>> My proposed option:
>>>> >>>
>>>> >>> 1) remove the limit and potentially make it
>>>> configurable,
>>>> >>> Motivation:
>>>> >>> The system administrator can enforce a limit its
>>>> users need to respect that it's in line with whatever
>>>> the admin decided to be acceptable for them.
>>>> >>> Default can stay the current one.
>>>> >>>
>>>> >>> That's my favourite at the moment, but I agree that
>>>> potentially in the future this may need to change, as
>>>> we may optimise the data structures for certain
>>>> dimensions. I am a big fan of Yagni (you aren't going
>>>> to need it) so I am ok we'll face a different
>>>> discussion if that happens in the future.
>>>> >>>
>>>> >>>
>>>> >>>
>>>> >>> On Sun, 9 Apr 2023, 18:46 Gus Heck,
>>>> <gus.heck@gmail.com> wrote:
>>>> >>>> What I see so far:
>>>> >>>>
>>>> >>>> Much positive support for raising the limit
>>>> >>>> Slightly less support for removing it or making it
>>>> configurable
>>>> >>>> A single veto which argues that a (as yet
>>>> undefined) performance standard must be met before
>>>> raising the limit
>>>> >>>> Hot tempers (various) making this discussion difficult
>>>> >>>>
>>>> >>>> As I understand it, vetoes must have technical
>>>> merit. I'm not sure that this veto rises to "technical
>>>> merit" on 2 counts:
>>>> >>>>
>>>> >>>> No standard for the performance is given so it
>>>> cannot be technically met. Without hard criteria it's a
>>>> moving target.
>>>> >>>> It appears to encode a valuation of the user's
>>>> time, and that valuation is really up to the user. Some
>>>> users may consider 2hours useless and not worth it, and
>>>> others might happily wait 2 hours. This is not a
>>>> technical decision, it's a business decision regarding
>>>> the relative value of the time invested vs the value of
>>>> the result. If I can cure cancer by indexing for a
>>>> year, that might be worth it... (hyperbole of course).
>>>> >>>>
>>>> >>>> Things I would consider to have technical merit
>>>> that I don't hear:
>>>> >>>>
>>>> >>>> Impact on the speed of **other** indexing
>>>> operations. (devaluation of other functionality)
>>>> >>>> Actual scenarios that work when the limit is low
>>>> and fail when the limit is high (new failure on the
>>>> same data with the limit raised).
>>>> >>>>
>>>> >>>> One thing that might or might not have technical merit
>>>> >>>>
>>>> >>>> If someone feels there is a lack of documentation
>>>> of the costs/performance implications of using large
>>>> vectors, possibly including reproducible benchmarks
>>>> establishing the scaling behavior (there seems to be
>>>> disagreement on O(n) vs O(n^2)).
>>>> >>>>
>>>> >>>> The users *should* know what they are getting
>>>> into, but if the cost is worth it to them, they should
>>>> be able to pay it without forking the project. If this
>>>> veto causes a fork that's not good.
>>>> >>>>
>>>> >>>> On Sun, Apr 9, 2023 at 7:55?AM Michael Sokolov
>>>> <msokolov@gmail.com> wrote:
>>>> >>>>> We do have a dataset built from Wikipedia in
>>>> luceneutil. It comes in 100 and 300 dimensional
>>>> varieties and can easily enough generate large numbers
>>>> of vector documents from the articles data. To go
>>>> higher we could concatenate vectors from that and I
>>>> believe the performance numbers would be plausible.
>>>> >>>>>
>>>> >>>>> On Sun, Apr 9, 2023, 1:32 AM Dawid Weiss
>>>> <dawid.weiss@gmail.com> wrote:
>>>> >>>>>> Can we set up a branch in which the limit is
>>>> bumped to 2048, then have
>>>> >>>>>> a realistic, free data set (wikipedia sample or
>>>> something) that has,
>>>> >>>>>> say, 5 million docs and vectors created using
>>>> public data (glove
>>>> >>>>>> pre-trained embeddings or the like)? We then
>>>> could run indexing on the
>>>> >>>>>> same hardware with 512, 1024 and 2048 and see
>>>> what the numbers, limits
>>>> >>>>>> and behavior actually are.
>>>> >>>>>>
>>>> >>>>>> I can help in writing this but not until after
>>>> Easter.
>>>> >>>>>>
>>>> >>>>>>
>>>> >>>>>> Dawid
>>>> >>>>>>
>>>> >>>>>> On Sat, Apr 8, 2023 at 11:29?PM Adrien Grand
>>>> <jpountz@gmail.com> wrote:
>>>> >>>>>>> As Dawid pointed out earlier on this thread,
>>>> this is the rule for
>>>> >>>>>>> Apache projects: a single -1 vote on a code
>>>> change is a veto and
>>>> >>>>>>> cannot be overridden. Furthermore, Robert is
>>>> one of the people on this
>>>> >>>>>>> project who worked the most on debugging subtle
>>>> bugs, making Lucene
>>>> >>>>>>> more robust and improving our test framework,
>>>> so I'm listening when he
>>>> >>>>>>> voices quality concerns.
>>>> >>>>>>>
>>>> >>>>>>> The argument against removing/raising the limit
>>>> that resonates with me
>>>> >>>>>>> the most is that it is a one-way door. As MikeS
>>>> highlighted earlier on
>>>> >>>>>>> this thread, implementations may want to take
>>>> advantage of the fact
>>>> >>>>>>> that there is a limit at some point too. This
>>>> is why I don't want to
>>>> >>>>>>> remove the limit and would prefer a slight
>>>> increase, such as 2048 as
>>>> >>>>>>> suggested in the original issue, which would
>>>> enable most of the things
>>>> >>>>>>> that users who have been asking about raising
>>>> the limit would like to
>>>> >>>>>>> do.
>>>> >>>>>>>
>>>> >>>>>>> I agree that the merge-time memory usage and
>>>> slow indexing rate are
>>>> >>>>>>> not great. But it's still possible to index
>>>> multi-million vector
>>>> >>>>>>> datasets with a 4GB heap without hitting OOMEs
>>>> regardless of the
>>>> >>>>>>> number of dimensions, and the feedback I'm
>>>> seeing is that many users
>>>> >>>>>>> are still interested in indexing multi-million
>>>> vector datasets despite
>>>> >>>>>>> the slow indexing rate. I wish we could do
>>>> better, and vector indexing
>>>> >>>>>>> is certainly more expert than text indexing,
>>>> but it still is usable in
>>>> >>>>>>> my opinion. I understand how giving Lucene more
>>>> information about
>>>> >>>>>>> vectors prior to indexing (e.g. clustering
>>>> information as Jim pointed
>>>> >>>>>>> out) could help make merging faster and more
>>>> memory-efficient, but I
>>>> >>>>>>> would really like to avoid making it a
>>>> requirement for indexing
>>>> >>>>>>> vectors as it also makes this feature much
>>>> harder to use.
>>>> >>>>>>>
>>>> >>>>>>> On Sat, Apr 8, 2023 at 9:28?PM Alessandro Benedetti
>>>> >>>>>>> <a.benedetti@sease.io> wrote:
>>>> >>>>>>>> I am very attentive to listen opinions but I
>>>> am un-convinced here and I an not sure that a single
>>>> person opinion should be allowed to be detrimental for
>>>> such an important project.
>>>> >>>>>>>>
>>>> >>>>>>>> The limit as far as I know is literally just
>>>> raising an exception.
>>>> >>>>>>>> Removing it won't alter in any way the current
>>>> performance for users in low dimensional space.
>>>> >>>>>>>> Removing it will just enable more users to use
>>>> Lucene.
>>>> >>>>>>>>
>>>> >>>>>>>> If new users in certain situations will be
>>>> unhappy with the performance, they may contribute
>>>> improvements.
>>>> >>>>>>>> This is how you make progress.
>>>> >>>>>>>>
>>>> >>>>>>>> If it's a reputation thing, trust me that not
>>>> allowing users to play with high dimensional space will
>>>> equally damage it.
>>>> >>>>>>>>
>>>> >>>>>>>> To me it's really a no brainer.
>>>> >>>>>>>> Removing the limit and enable people to use
>>>> high dimensional vectors will take minutes.
>>>> >>>>>>>> Improving the hnsw implementation can take months.
>>>> >>>>>>>> Pick one to begin with...
>>>> >>>>>>>>
>>>> >>>>>>>> And there's no-one paying me here, no company
>>>> interest whatsoever, actually I pay people to
>>>> contribute, I am just convinced it's a good idea.
>>>> >>>>>>>>
>>>> >>>>>>>>
>>>> >>>>>>>> On Sat, 8 Apr 2023, 18:57 Robert Muir,
>>>> <rcmuir@gmail.com> wrote:
>>>> >>>>>>>>> I disagree with your categorization. I put in
>>>> plenty of work and
>>>> >>>>>>>>> experienced plenty of pain myself, writing
>>>> tests and fighting these
>>>> >>>>>>>>> issues, after i saw that, two releases in a
>>>> row, vector indexing fell
>>>> >>>>>>>>> over and hit integer overflows etc on small
>>>> datasets:
>>>> >>>>>>>>>
>>>> >>>>>>>>> https://github.com/apache/lucene/pull/11905
>>>> >>>>>>>>>
>>>> >>>>>>>>> Attacking me isn't helping the situation.
>>>> >>>>>>>>>
>>>> >>>>>>>>> PS: when i said the "one guy who wrote the
>>>> code" I didn't mean it in
>>>> >>>>>>>>> any kind of demeaning fashion really. I meant
>>>> to describe the current
>>>> >>>>>>>>> state of usability with respect to indexing a
>>>> few million docs with
>>>> >>>>>>>>> high dimensions. You can scroll up the thread
>>>> and see that at least
>>>> >>>>>>>>> one other committer on the project
>>>> experienced similar pain as me.
>>>> >>>>>>>>> Then, think about users who aren't committers
>>>> trying to use the
>>>> >>>>>>>>> functionality!
>>>> >>>>>>>>>
>>>> >>>>>>>>> On Sat, Apr 8, 2023 at 12:51?PM Michael
>>>> Sokolov <msokolov@gmail.com> wrote:
>>>> >>>>>>>>>> What you said about increasing dimensions
>>>> requiring a bigger ram buffer on merge is wrong. That's
>>>> the point I was trying to make. Your concerns about
>>>> merge costs are not wrong, but your conclusion that we
>>>> need to limit dimensions is not justified.
>>>> >>>>>>>>>>
>>>> >>>>>>>>>> You complain that hnsw sucks it doesn't
>>>> scale, but when I show it scales linearly with
>>>> dimension you just ignore that and complain about
>>>> something entirely different.
>>>> >>>>>>>>>>
>>>> >>>>>>>>>> You demand that people run all kinds of
>>>> tests to prove you wrong but when they do, you don't
>>>> listen and you won't put in the work yourself or
>>>> complain that it's too hard.
>>>> >>>>>>>>>>
>>>> >>>>>>>>>> Then you complain about people not meeting
>>>> you half way. Wow
>>>> >>>>>>>>>>
>>>> >>>>>>>>>> On Sat, Apr 8, 2023, 12:40 PM Robert Muir
>>>> <rcmuir@gmail.com> wrote:
>>>> >>>>>>>>>>> On Sat, Apr 8, 2023 at 8:33?AM Michael Wechner
>>>> >>>>>>>>>>> <michael.wechner@wyona.com> wrote:
>>>> >>>>>>>>>>>> What exactly do you consider reasonable?
>>>> >>>>>>>>>>> Let's begin a real discussion by being
>>>> HONEST about the current
>>>> >>>>>>>>>>> status. Please put politically correct or
>>>> your own company's wishes
>>>> >>>>>>>>>>> aside, we know it's not in a good state.
>>>> >>>>>>>>>>>
>>>> >>>>>>>>>>> Current status is the one guy who wrote the
>>>> code can set a
>>>> >>>>>>>>>>> multi-gigabyte ram buffer and index a small
>>>> dataset with 1024
>>>> >>>>>>>>>>> dimensions in HOURS (i didn't ask what
>>>> hardware).
>>>> >>>>>>>>>>>
>>>> >>>>>>>>>>> My concerns are everyone else except the
>>>> one guy, I want it to be
>>>> >>>>>>>>>>> usable. Increasing dimensions just means
>>>> even bigger multi-gigabyte
>>>> >>>>>>>>>>> ram buffer and bigger heap to avoid OOM on
>>>> merge.
>>>> >>>>>>>>>>> It is also a permanent backwards
>>>> compatibility decision, we have to
>>>> >>>>>>>>>>> support it once we do this and we can't
>>>> just say "oops" and flip it
>>>> >>>>>>>>>>> back.
>>>> >>>>>>>>>>>
>>>> >>>>>>>>>>> It is unclear to me, if the multi-gigabyte
>>>> ram buffer is really to
>>>> >>>>>>>>>>> avoid merges because they are so slow and
>>>> it would be DAYS otherwise,
>>>> >>>>>>>>>>> or if its to avoid merges so it doesn't hit
>>>> OOM.
>>>> >>>>>>>>>>> Also from personal experience, it takes
>>>> trial and error (means
>>>> >>>>>>>>>>> experiencing OOM on merge!!!) before you
>>>> get those heap values correct
>>>> >>>>>>>>>>> for your dataset. This usually means
>>>> starting over which is
>>>> >>>>>>>>>>> frustrating and wastes more time.
>>>> >>>>>>>>>>>
>>>> >>>>>>>>>>> Jim mentioned some ideas about the memory
>>>> usage in IndexWriter, seems
>>>> >>>>>>>>>>> to me like its a good idea. maybe the
>>>> multigigabyte ram buffer can be
>>>> >>>>>>>>>>> avoided in this way and performance
>>>> improved by writing bigger
>>>> >>>>>>>>>>> segments with lucene's defaults. But this
>>>> doesn't mean we can simply
>>>> >>>>>>>>>>> ignore the horrors of what happens on
>>>> merge. merging needs to scale so
>>>> >>>>>>>>>>> that indexing really scales.
>>>> >>>>>>>>>>>
>>>> >>>>>>>>>>> At least it shouldnt spike RAM on trivial
>>>> data amounts and cause OOM,
>>>> >>>>>>>>>>> and definitely it shouldnt burn hours and
>>>> hours of CPU in O(n^2)
>>>> >>>>>>>>>>> fashion when indexing.
>>>> >>>>>>>>>>>
>>>> >>>>>>>>>>>
>>>> ---------------------------------------------------------------------
>>>> >>>>>>>>>>> To unsubscribe, e-mail:
>>>> dev-unsubscribe@lucene.apache.org
>>>> >>>>>>>>>>> For additional commands, e-mail:
>>>> dev-help@lucene.apache.org
>>>> >>>>>>>>>>>
>>>> >>>>>>>>>
>>>> ---------------------------------------------------------------------
>>>> >>>>>>>>> To unsubscribe, e-mail:
>>>> dev-unsubscribe@lucene.apache.org
>>>> >>>>>>>>> For additional commands, e-mail:
>>>> dev-help@lucene.apache.org
>>>> >>>>>>>>>
>>>> >>>>>>>
>>>> >>>>>>> --
>>>> >>>>>>> Adrien
>>>> >>>>>>>
>>>> >>>>>>>
>>>> ---------------------------------------------------------------------
>>>> >>>>>>> To unsubscribe, e-mail:
>>>> dev-unsubscribe@lucene.apache.org
>>>> >>>>>>> For additional commands, e-mail:
>>>> dev-help@lucene.apache.org
>>>> >>>>>>>
>>>> >>>>>>
>>>> ---------------------------------------------------------------------
>>>> >>>>>> To unsubscribe, e-mail:
>>>> dev-unsubscribe@lucene.apache.org
>>>> >>>>>> For additional commands, e-mail:
>>>> dev-help@lucene.apache.org
>>>> >>>>>>
>>>> >>>>
>>>> >>>> --
>>>> >>>> http://www.needhamsoftware.com (work)
>>>> >>>> http://www.the111shift.com (play)
>>>> >>
>>>> >
>>>> ---------------------------------------------------------------------
>>>> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> > For additional commands, e-mail:
>>>> dev-help@lucene.apache.org
>>>> >
>>>>
>>>>
>>>> ---------------------------------------------------------------------
>>>> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
>>>> For additional commands, e-mail: dev-help@lucene.apache.org
>>>>
>>>
>>
>

Re: [Proposal] Remove max number of dimensions for KNN vectors [ In reply to ]

a.benedetti at sease

May 9, 2023, 5:11 AM

Post #99 of 99 (912 views)