[esafak]

HNSW seems better in the criteria that matter.

I think similarity search is a commodity now; I would not invest in developing an in-house solution given the abundance of good commercial solutions.

[bobvanluijt]

That depends a bit on the scale and use case specifics. But commoditized billion-scale vector search is indeed a thing. We published this for Weaviate in December last year https://weaviate.io/blog/sphere-dataset-in-weaviate

[fzliu]

We've seen Milvus used in a variety of recommender systems running in production.

[cjbgkagh]

They’re embeddings so they’re dense. There are few things easier than dense vector similarity.

[esafak]

Embeddings for retrieval don't have to be. It is not unheard of to transform the raw embeddings to optimize them for retrieval; e.g., through binarization or hashing.

[cjbgkagh]

I was more making a distinction between embeddings and bag of words which are very very sparse matrices. The embedding dimensionality will not be anywhere near as high so this level of sparsity is a minor inconvenience.

Edit: also CPUs for this, yikes…

[jakearmitage]

Such as...?

[kartoolOz]

Vespa.ai is pretty crazy too, a bit unkown we run a huge vespa cluster serving 1k+ queries with <100ms latency ...

[jabo]

Typesense: https://typesense.org/docs/0.24.1/api/vector-search.html

[esafak]

Qdrant, milvus, weaviate