Sure, but lucene should be able to have limits. We have this discussion with every single limit we attempt to implement :) There will always be extreme use cases using too many dimensions or whatever. It is open source! I think if what you are doing is strange enough, you can modify the sources.
Personally, I'm concerned about increasing this limit: things are quite slow already with hundreds of dimensions. There seems to be no light at the end of the tunnel for the JDK vector api, I think OpenJDK will incubate this API until the sun supernovas and java is dead :) It is frustrating, as that could give current implementation a needed performance boost on basically any hardware. Also, I'm concerned about increasing limit while HNSW is the only implementation. I'd like us to keep the door open to alternative algorithms that might have better performance. On Tue, Feb 15, 2022 at 12:21 PM Michael Wechner <michael.wech...@wyona.com> wrote: > I understand, but if Lucene itself would allow to overwrite the default > max size programmatically, then I think it should be clear that you do this > at your own risk :-) > > Thanks for the links to your blog posts, which sound very interesting. > > Thanks > > Michael > > Am 15.02.22 um 17:25 schrieb Alessandro Benedetti: > > I believe it could make sense, but as Michael pointed out in the Jira > ticket related to the Solr integration, then we'll get complaints like "I > set it to 1.000.000 and my Solr instance doesn't work anymore" (I kept > everything super simple just to simulate a realistic scenario). > So I tend to agree to keep it to 1024 at the moment and potentially extend > it(providing some benchmark on common machines as a reference to justify > the increase). > > In terms of your original question, how are you training/fine-tuning your > models? > Using pre-trained language models won't probably help you that much, on > top of that, queries are short, so you may require a well-curated > fine-tuning step. > We have a series of blog posts on that, and one is coming soon: > https://sease.io/2021/12/using-bert-to-improve-search-relevance.html > > https://sease.io/2022/01/tackling-vocabulary-mismatch-with-document-expansion.html > > Cheers > -------------------------- > Alessandro Benedetti > Apache Lucene/Solr PMC member and Committer > Director, R&D Software Engineer, Search Consultant > > www.sease.io > > > On Tue, 15 Feb 2022 at 09:10, Michael Wechner <michael.wech...@wyona.com> > wrote: > >> fair enough, but wouldn't it make sense that one can increase it >> programmatically, e.g. >> >> .setVectorMaxDimension(2028) >> >> ? >> >> Thanks >> >> Michael >> >> >> Am 14.02.22 um 23:34 schrieb Michael Sokolov: >> > I think we picked the 1024 number as something that seemed so large >> > nobody would ever want to exceed it! Obviously that was naive. Still >> > the limit serves as a cautionary point for users; if your vectors are >> > bigger than this, there is probably a better way to accomplish what >> > you are after (eg better off-line training to reduce dimensionality). >> > Is 1024 the magic number? Maybe not, but before increasing I'd like to >> > see some strong evidence that bigger vectors than that are indeed >> > useful as part of a search application using Lucene. >> > >> > -Mike >> > >> > On Mon, Feb 14, 2022 at 5:08 PM Julie Tibshirani <juliet...@gmail.com> >> wrote: >> >> Sounds good, hope the testing goes well! Memory and CPU (largely from >> more expensive vector distance calculations) are indeed the main factors to >> consider. >> >> >> >> Julie >> >> >> >> On Mon, Feb 14, 2022 at 1:02 PM Michael Wechner < >> michael.wech...@wyona.com> wrote: >> >>> Hi Julie >> >>> >> >>> Thanks again for your feedback! >> >>> >> >>> I will do some more tests with "all-mpnet-base-v2" (768) and >> "all-roberta-large-v1" (1024), so 1024 is enough for me for the moment :-) >> >>> >> >>> But yes, I could imagine, that eventually it might make sense to >> allow more dimensions than 1024. >> >>> >> >>> Beside memory and "CPU", are there other limiting factors re more >> dimensions? >> >>> >> >>> Thanks >> >>> >> >>> Michael >> >>> >> >>> Am 14.02.22 um 21:53 schrieb Julie Tibshirani: >> >>> >> >>> Hello Michael, the max number of dimensions is currently hardcoded >> and can't be changed. I could see an argument for increasing the default a >> bit and would be happy to discuss if you'd like to file a JIRA issue. >> However 12288 dimensions still seems high to me, this is much larger than >> most well-established embedding models and could require a lot of memory. >> >>> >> >>> Julie >> >>> >> >>> On Mon, Feb 14, 2022 at 12:08 PM Michael Wechner < >> michael.wech...@wyona.com> wrote: >> >>>> Hi Julie >> >>>> >> >>>> Thanks very much for this link, which is very interesting! >> >>>> >> >>>> Btw, do you have an idea how to increase the default max size of >> 1024? >> >>>> >> >>>> https://lists.apache.org/thread/hyb6w5c4x5rjt34k3w7zqn3yp5wvf33o >> >>>> >> >>>> Thanks >> >>>> >> >>>> Michael >> >>>> >> >>>> >> >>>> >> >>>> Am 14.02.22 um 17:45 schrieb Julie Tibshirani: >> >>>> >> >>>> Hello Michael, I don't have personal experience with these models, >> but I found this article insightful: >> https://medium.com/@nils_reimers/openai-gpt-3-text-embeddings-really-a-new-state-of-the-art-in-dense-text-embeddings-6571fe3ec9d9. >> It evaluates the OpenAI models against a variety of existing models on >> tasks like sentence similarity and text retrieval. Although the other >> models are cheaper and have fewer dimensions, the OpenAI ones perform >> similarly or worse. This got me thinking that they might not be a good >> cost/ effectiveness trade-off, especially the larger ones with 4096 or >> 12288 dimensions. >> >>>> >> >>>> Julie >> >>>> >> >>>> On Sun, Feb 13, 2022 at 1:55 AM Michael Wechner < >> michael.wech...@wyona.com> wrote: >> >>>>> Re the OpenAI embedding the following recent paper might be of >> interest >> >>>>> >> >>>>> https://arxiv.org/pdf/2201.10005.pdf >> >>>>> >> >>>>> (Text and Code Embeddings by Contrastive Pre-Training, Jan 24, 2022) >> >>>>> >> >>>>> Thanks >> >>>>> >> >>>>> Michael >> >>>>> >> >>>>> Am 13.02.22 um 00:14 schrieb Michael Wechner: >> >>>>> >> >>>>> Here a concrete example where I combine OpenAI model >> "text-similarity-ada-001" with Lucene vector search >> >>>>> >> >>>>> INPUT sentence: "What is your age this year?" >> >>>>> >> >>>>> Result sentences >> >>>>> >> >>>>> 1) How old are you this year? >> >>>>> score '0.98860765' >> >>>>> >> >>>>> 2) What was your age last year? >> >>>>> score '0.97811764' >> >>>>> >> >>>>> 3) What is your age? >> >>>>> score '0.97094905' >> >>>>> >> >>>>> 4) How old are you? >> >>>>> score '0.9600177' >> >>>>> >> >>>>> >> >>>>> Result 1 is great and result 2 looks similar, but is not correct >> from an "understanding" point of view and results 3 and 4 are good again. >> >>>>> >> >>>>> I understand "similarity" is not the same as "understanding", but I >> hope it makes it clearer what I am looking for :-) >> >>>>> >> >>>>> Thanks >> >>>>> >> >>>>> Michael >> >>>>> >> >>>>> >> >>>>> >> >>>>> Am 12.02.22 um 22:38 schrieb Michael Wechner: >> >>>>> >> >>>>> Hi Alessandro >> >>>>> >> >>>>> I am mainly interested in detecting similarity, for example whether >> the following two sentences are similar resp. likely to mean the same thing >> >>>>> >> >>>>> "How old are you?" >> >>>>> "What is your age?" >> >>>>> >> >>>>> and that the following two sentences are not similar, resp. do not >> mean the same thing >> >>>>> >> >>>>> "How old are you this year?" >> >>>>> "How old have you been last year?" >> >>>>> >> >>>>> But also performance or how OpenAI embeddings compare for example >> with SBERT (https://sbert.net/docs/usage/semantic_textual_similarity.html >> ) >> >>>>> >> >>>>> Thanks >> >>>>> >> >>>>> Michael >> >>>>> >> >>>>> >> >>>>> >> >>>>> Am 12.02.22 um 20:41 schrieb Alessandro Benedetti: >> >>>>> >> >>>>> Hi Michael, experience to what extent? >> >>>>> We have been exploring the area for a while given we contributed >> the first neural search milestone to Apache Solr. >> >>>>> What is your curiosity? Performance? Relevance impact? How to >> integrate it? >> >>>>> Regards >> >>>>> >> >>>>> On Fri, 11 Feb 2022, 22:38 Michael Wechner, < >> michael.wech...@wyona.com> wrote: >> >>>>>> Hi >> >>>>>> >> >>>>>> Does anyone have experience using OpenAI embeddings in combination >> with Lucene vector search? >> >>>>>> >> >>>>>> https://beta.openai.com/docs/guides/embeddings >> >>>>>> >> >>>>>> for example comparing performance re vector size >> >>>>>> >> >>>>>> >> https://api.openai.com/v1/engines/text-similarity-ada-001/embeddings >> >>>>>> >> >>>>>> and >> >>>>>> >> >>>>>> >> https://api.openai.com/v1/engines/text-similarity-davinci-001/embeddings >> >>>>>> >> >>>>>> ? >> >>>>>> >> >>>>>> >> >>>>>> Thanks >> >>>>>> >> >>>>>> Michael >> >>>>> >> >>>>> >> >>>>> >> > --------------------------------------------------------------------- >> > To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >> > For additional commands, e-mail: dev-h...@lucene.apache.org >> > >> >> >> --------------------------------------------------------------------- >> To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >> For additional commands, e-mail: dev-h...@lucene.apache.org >> >> >
