4 ms·
> The index is tiny What was the largest index you've had on Faiss? That seems to affect whether people think of it as more than adequate or terribly inadequat
by gk1 5y ago
> The index is tiny
What was the largest index you've had on Faiss? That seems to affect whether people think of it as more than adequate or terribly inadequate.
- freediver 5y agoThis demo is only about million vectors. The largest I had in Faiss was embeddings of the entire Wikipedia (scale in the neighborhood of ~30 million vectors). I know people running few billion vectors in Faiss.
- leobg 5y agoSo one vector per article? Doesn’t this skew results? A short article with 0.9 relevance score would rank higher than a long article containing one paragraph with 1.0 relevance. Am I mistaken?
- visarga 5y agoIt's a known problem. Choosing appropriately between cosine similarity or dot product depending on your use case can help.
- marginalia_nu 5y agoYeah, this is also a problem traditional search ranking algorithms needs to compensate for.
- leobg 5y agoAlso, BERT on cheap hardware? I thought that without a GPU, vectorizing millions of snippets or doing sub-second queries was basically out of the question.
- visarga 5y agoCPU BERT inference is fast enough to embed 50 examples per second. Your large index is built offline, the query is embedded live but it's just one at a time. Approximate similarity search complexity is logarithmic so it's super fast on large collections.
- jkb79 5y agoIt's about choosing the right Transformer model, there are several models which are smaller, with fewer parameters than bert-base which gives the exact same accuracy as bert-base, which you can use on a modern CPU single digit ms, even with a single intra-thread. See for example, https://github.com/vespa-engine/sample-apps/blob/master/msmarco-ranking/passage-ranking.md https://github.com/vespa-engine/sample-apps/blob/master/msma...
- leobg 5y agoThanks for the link! I compared BERT[1], distilbert[2], mpnet[3] and minilm[4] in the past. But the results I got "out of the box" for semantic search were not better than using fastText, which is orders of magnitude faster. BERT and distilbert are 400x slower than fastText, minilm 300x, and mpnet 700x. At least if you are using a CPU-only machine. USE, xlmroberta and elmo were even worse (5,000 - 18,000x slower). I also love how fast and easy it is to train your own fastText model. [1]: distiluse-base-multilingual-cased-v2 [2]: multi-qa-MiniLM-L6-cos-v1 [3]: multi-qa-mpnet-base-cos-v1 [4]: multi-qa-MiniLM-L6-cos-v1
- jkb79 5y agoVector models are nothing but representation learning and applying the model out-of-domain usually gives worse results than plain old BM25. See https://arxiv.org/abs/2104.08663 https://arxiv.org/abs/2104.08663 A concrete example is DPR which is a state of the art dense retriever model for wikipedia for question answering, when applying that model on MS Marco passage ranking it performs worse than plain BM25.
- m3at 5y agoNot the author, but at work we've had in the hundreds of millions. Faiss can certainly scale. If you do have a tiny index and want to try Google's version of vector search (as an alternative to Faiss), you can easily run ScaNN locally [1] (linked in the article, that's the underlying tech). On small scale I had better perf with ScaNN [1] https://github.com/google-research/google-research/tree/master/scann https://github.com/google-research/google-research/tree/mast...
- freediver 5y agoThanks for mentioning it. Just tried it and it seems to be faster than Faiss indeed on smaller scale searches.
- leobg 5y agoThere's also hnswlib[1], which has supposedly lower memory requirements and allows for adding new vectors to an existing index. [1]: https://github.com/nmslib/hnswlib/ https://github.com/nmslib/hnswlib/