K-Schubert
commented
1 month ago @tabee
Initializing and inserting data in pgvector db doesn't require constructing an "index" (search through the vectors is performed with exact nearest neighbours which is exact search with perfect recall, but it's slow if there are many vectors in the db). An index can be built to speed up search (decreasing recall) using 2 different techniques with their advantages/tradeoffs:
- HNSW: better speed-recall tradeoff compared to ivfflat, but takes longer to build index and uses more memory. An index can also be built without data inside.
- IVFFLAT: Lower speed-recall tradeoff compared to hnsw, but faster build times and requires less memory. Requires a training step (and data) to build index.
Both methods require some finetuning with index parameters to evaluate search performance.
Since there probably won't be millions of vectors in the db, we might want to skip this index building?
Description
Configure vectordb index creation (hnsw, ivfflat).
Configure semantic search with pgvector (search params such as "m", "ef_construction" for hnsw, "probes" for ivfflat) through .env file.