annoy VS ann-benchmarks

The focus on the top 10 in vector search is a product of wanting to prove value over keyword search. Keyword search is going to miss some conceptual matches. You can try to work around that with tokenization and complex queries with all variations but it's not easy.

Vector search isn't all that new a concept. For example, the annoy library (https://github.com/spotify/annoy) has been around since 2014. It was one of the first open source approximate nearest neighbor libraries. Recommendations have always been a good use case for vector similarity.

Recommendations are a natural extension of search and transformers models made building the vectors for natural language possible. To prove the worth of vector search over keyword search, the focus was always on showing how the top N matches include results not possible with keyword search.

In 2023, there has been a shift towards acknowledging keyword search also has value and that a combination of vector + keyword search (aka hybrid search) operates in the sweet spot. Once again this is validated through the same benchmarks which focus on the top 10.

On top of all this, there is also the reality that the vector database space is very crowded and some want to use their performance benchmarks for marketing.

Disclaimer: I am the author of txtai (https://github.com/neuml/txtai), an open source embeddings database

If you want to go larger you could still use some simple setup in conjunction with faiss, annoy or hnsw.

I then use annoy to compare them. Annoy can use different measures for distance, like cosine, euclidean and more

Yes you can do this for equality predicates if your row groups are sorted . This blog post (that I didn't write) might add more color. You can't do this for any kind of text searching. If you need to do this with file based storage I'd recommend using a vector based text search and utilize a ANN index library like Annoy.

If you need large scale (1000+ dimension, millions+ source points, >1000 queries per second) and accept imperfect results / approximate nearest neighbors, then other people have already mentioned some of the best libraries (FAISS, Annoy).

Would be possible to further speed up the process with using something like ANNOY? https://github.com/spotify/annoy

I like Faiss but I tried Spotify's annoy[1] for a recent project and was pretty impressed.

Since lots of people don't seem to understand how useful these embedding libraries are here's an example. I built a thing that indexes bouldering and climbing competition videos, then builds an embedding of the climber's body position per frame. I then can automatically match different climbers on the same problem.

It works pretty well. Since the body positions are 3D it works reasonably well across camera angles.

The biggest problem is getting the embedding right. I simplified it a lot above because I actually need to embed the problem shape itself because otherwise it matches too well: you get frames of people in identical positions but on different problems!

[1] https://github.com/spotify/annoy

If you just want quick in memory search then pynndescent is a decent option: it's easy to install, and easy to get running. Another good option is Annoy; it's just as easy to install and get running with python, but it is a little less performant if you want to do a lot of queries, or get a knn-graph quickly.

https://ann-benchmarks.com/ is a good resource covering those libraries and much more.

We utilized the ANN Benchmarks methodology, a standard for benchmarking vector databases. Our tests used the dbpedia dataset of 1,000,000 OpenAI embeddings (1536 dimensions) and inner product distance metric for both Pinecone and pgvector.

Data warehouses are columnar stores. They are very different from row-oriented databases - like Postgres, MySQL. Operations on columns - e.g., aggregations (mean of a column) are very efficient.

Most vector databases use one of a few different vector indexing libraries - FAISS, hnswlib, and scann (google only) are popular. The newer vector dbs, like weaviate, have introduced their own indexes, but i haven't seen any performance difference -

Reference: https://ann-benchmarks.com/

(Blog author here). Thanks for the question. In this case the index for both DiskANN and pgvector HNSW is small enough to fit in memory on the machine (8GB RAM), so there's no need to touch the SSD. We plan to test on a config where the index size is larger than memory (we couldn't this time due to limitations in ANN benchmarks [0], the tool we use).

To your question about RAM usage, we provide a graph of index size. When enabling PQ, our new index is 10x smaller than pgvector HNSW. We don't have numbers for HNSWPQ in FAISS yet.

[0]: https://github.com/erikbern/ann-benchmarks/

In terms of "All You Need" for Vector Search, ANN Benchmarks (https://ann-benchmarks.com/) is a good site to review when deciding what you need. As with anything complex, there often isn't a universal solution.

txtai (https://github.com/neuml/txtai) can build indexes with Faiss, Hnswlib and Annoy. All 3 libraries have been around at least 4 years and are mature. txtai also supports storing metadata in SQLite, DuckDB and the next release will support any JSON-capable database supported by SQLAlchemy (Postgres, MariaDB/MySQL, etc).

pg_vector doesn't perform well compared to other methods, at least according to ANN-Benchmarks (https://ann-benchmarks.com/).

txtai is more than just a vector database. It also has a built-in graph component for topic modeling that utilizes the vector index to autogenerate relationships. It can store metadata in SQLite/DuckDB with support for other databases coming. It has support for running LLM prompts right with the data, similar to a stored procedure, through workflows. And it has built-in support for vectorizing data into vectors.

For vector databases that simply store vectors, I agree that it's nothing more than just a different index type.

The reason https://ann-benchmarks.com is so good, is that we can see a plot of recall vs latency. I can see you have some latency numbers in the leaderboard at the bottom, but it's very difficult to make a decision.

As a practitioner that works with vector databases every day, just latency is meaningless to me, because I need to know if it's fast AND accurate, and what the tradeoff is! You can't have it both ways. So it would be helpful if you showed plots showing this tradeoff, similar to ann-benchmarks.

There are very efficient algorithms for doing this, but of course it may still be expensive if your dataset is very large. See https://ann-benchmarks.com/ for some of the algorithms