pgvector

pgvector is a PostgreSQL extension that adds vector storage and similarity search to a database many teams already run. It supports halfvec, bit, and sparsevec types, HNSW and IVFFlat indexes, and L2 / inner-product / cosine / Hamming / Jaccard distances. Combined with tsvector for full-text search, you get hybrid retrieval inside a single SQL query — no second datastore to deploy, monitor, or back up.

For AI agents, pgvector's biggest argument is operational simplicity. ACID guarantees, joins against your existing relational data, point-in-time recovery, and per-row row-level security all come for free. You can scope a vector query to a tenant with a WHERE tenant_id = $1 clause and let Postgres do what it has done for decades.

The honest tradeoffs are performance and feature surface. At tens of millions of vectors with high QPS, dedicated engines (Qdrant, Pinecone, Milvus) typically deliver lower p99 latency and higher recall per dollar. pgvector itself has no built-in embedding generation, no MCP server, and no native multi-tenancy beyond Postgres schemas — though host platforms (Supabase, Neon, Crunchy Bridge) layer some of these on top.

Agent-specific features:

• Hybrid retrieval via tsvector + vector in a single SQL query
• Row-level security for per-tenant or per-user isolation
• Joins against existing relational data — no dual-write or sync pipeline
• Available on every major managed Postgres (Supabase, Neon, RDS, Cloud SQL, Crunchy Bridge)
• Same backups, replication, and observability as the rest of your Postgres fleet