pgvector

Should I Use pgvector or Pinecone for My AI App?

Use pgvector if you already run Postgres and your vector count stays under ~1 million rows — you avoid a separate service, separate billing, and the embedding/metadata join is a native SQL query. Use Pinecone when you need sub-10ms ANN search at 10M+ vectors, multi-tenancy with namespace isolation, or managed metadata filtering at scale. The switching cost is real: pgvector embeddings are just a column you can export; Pinecone data requires re-indexing to migrate.

Also see: Pinecone · Supabase

Which pgvector Index Type Should I Use — HNSW or IVFFlat?

Default to HNSW for almost all use cases: CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops). It supports incremental inserts, queries without needing to set ivfflat.probes, and delivers better recall at the same speed. Use IVFFlat only if you're on a very memory-constrained server (IVFFlat uses less RAM at rest) or if your dataset is static and you can set lists = sqrt(row_count) at build time.