GILD analysis
pgvector Review
pgvector adds vector search to the Postgres a team already runs, free under the PostgreSQL License, with HNSW and IVFFlat indexes, half-precision and binary vectors, and replication through the write-ahead log. Discourse and Instacart report it in production, while filtered queries, index memory, release regressions and host versions draw most complaints. Adopt it when vectors belong with your data, and test filtered plans and your host's pgvector release before scale.
- Vector Databases
- AI Costs
The decision
- Best fit
- Engineering teams on Postgres who want vectors in the same database, from small RAG sites to Instacart-scale search.
- Consider another workflow
- If selective filters and more than roughly 10 million vectors dominate, a dedicated vector database fits: Mem0 moved its vectors to turbopuffer and Halcyon found Postgres unsuitable at its scale.
- Adoption condition
- Filtered queries tested with EXPLAIN at production size, iterative scans or partial indexes where filters are selective, and a host that ships a current pgvector release.