pgvector Not Working? Common Errors & Fixes

The server cannot find the vector extension files. Use the name vector, install the package for your Postgres major version, then CREATE EXTENSION vector.

HNSW builds slow down sharply when the graph no longer fits in maintenance_work_mem. Raise it for the session, add parallel workers and load data first.

The operator needs a vector on both sides. Cast the query parameter to vector or register the driver adapter.

A vector column stores up to 16,000 dimensions but HNSW and IVFFlat index at most 2,000. Use halfvec (4,000) or reduce the embedding size.

The column is fixed width, so every vector must match it. Re-embed with one model or create a new column at the new size.

Approximate indexes filter after the scan, capped by ef_search. Enable iterative scan, raise ef_search or use partial indexes.

The planner uses the index only for ORDER BY on a distance operator, ascending, with LIMIT, and a matching operator class.

Parallel HNSW builds need shared memory and RAM. Increase Docker shm_size, lower workers or memory settings.

Enable vector on the Extensions page or run create extension vector with schema extensions. Dimension must match your model.

Run CREATE EXTENSION IF NOT EXISTS vector; in each database. It is available on every plan.

Connect as rds_superuser and run CREATE EXTENSION vector. No restart is needed; update with ALTER EXTENSION vector UPDATE.