Pgvector

pgvector extends Postgres with vector similarity search, making it practical for agents to store and query embeddings alongside relational data without a separate vector database. It supports cosine, L2, and inner-product distance operators, HNSW and IVFFlat indexes, and integrates with any Postgres client. Agent-native use case is strong: agents can insert, index, and query embeddings in the same transaction as structured data. The extension is widely available on managed Postgres hosts (Supabase, Neon, Timescale, RDS). Confidence is docs-derived; runtime probe would require a live Postgres instance with the extension enabled.

pgvector adds SQL operators (<->, <=>, <#>) and index types (HNSW, IVFFlat) to standard Postgres. Agents interact via any PostgreSQL client using familiar SQL SELECT/INSERT semantics. No REST API — integration is at the client-library level. Vector columns are declared as VECTOR(n) where n is the embedding dimension. HNSW indexes offer good recall with tunable ef_construction and ef_search parameters. Bulk insert via COPY is supported. The SQL surface is clean and predictable; schema design is the integration challenge, not the API.

Authentication is inherited from Postgres: role-based access control, connection-level auth (password, SSL, IAM on managed hosts). No pgvector-specific auth surface — all security is at the Postgres layer. Agents should use read-only roles for query-only workloads and scoped write roles for embedding ingestion. Connection pooling (PgBouncer) is recommended for high-concurrency agent workloads.

pgvector inherits Postgres error semantics: well-defined error codes, transaction rollback on failure, and ACID guarantees. Index build failures (e.g., dimension mismatch) produce clear error messages. ANN recall is probabilistic — HNSW with default settings may miss some near neighbors; ef_search tuning trades recall for latency. No streaming or long-poll patterns; all operations are synchronous SQL queries.

The pgvector GitHub README is the primary documentation: clear installation instructions, supported operators, and index types explained with examples. Third-party documentation (Supabase, Neon, LangChain, LlamaIndex) is extensive and higher quality for agent-specific use cases. No official API reference beyond the README. The extension is mature and widely documented in AI ecosystem guides. Docs confidence is moderate — docs-derived only.