- We used a different index method called HNSW, which is more widely used in vector search area and also faster than ivfflat used by pgvector.
- There are some drawbacks with HNSW. It's designed for memory usage but not so suitable as a disk database index. Currently the storage of the HNSW index is managed outside postgres's buffer system, which is not so ideal. We're exploring more indexing method such as DiskANN to see whether we can integrate with postgres more. Fully integration with postgres can make things work like a charm
Since it was explicitly developed as a pgvector replacement, i feel it would have been useful to have a more in deph comparison though..
- what does it do specifically to achieve the speedup, other than rust? Does it use a better similarity search?
- what are potential drawbacks, other that itβs still in development? Missing features? Things that it does differently?