pgvector

Rather than bolting a vector database onto a Postgres stack, pgvector puts embedding storage and similarity search inside the database as a C extension: a dedicated column type and distance operators for L2, cosine, inner product, and binary metrics, so semantic search becomes an ORDER BY clause that joins, filters, and transacts like any other SQL. Searches are exact unless an HNSW or IVFFlat index is added, trading some recall for big speedups, and the 0.8 series introduced iterative index scans to fix filtered queries that came back short. Indexing tops out at 2,000 dimensions for standard vectors, with half-precision and binary quantization stretching that to 4,000 and 64,000. Install means compiling against Postgres 13 or newer, though RDS, Supabase, Neon, and most managed hosts bundle it, and client libraries cover every major language. It ships under the permissive PostgreSQL License. Against dedicated engines like Qdrant or Milvus the tradeoff is honest: index builds are memory-hungry and tuning is manual, but at low millions of vectors, keeping embeddings beside relational data usually wins on operational simplicity.