If you'd like an alternate approach, txtai can store content in Postgres. This approach stores metadata/fields in the database and builds a Faiss/HNSW index alongside. The idea is to lean on two mature projects vs having to rewrite a full vector index implementation.

https://neuml.hashnode.dev/external-database-integration

https://github.com/neuml/txtai

Disclaimer: I'm the primary author of txtai

Really great to see how the different config parameters (m, ef_construction, ef_search) impact latency and recall.

Interesting blog! Makes me wonder what a comparison between Pinecone + ChromaDB + Pgvector would look like..

in summary it seems pgvector is faster where it counts

Lantern argues that being able to create indexes quickly is useful for tuning the hnsw paramters.

Does pgvector use parallel index build in this benchmark ?

It does not! Would be interesting to try that!