Pgvector Rules

FATAL Constraints

• NEVER create an IVFFlat index. HNSW MUST be the default and only index type for vector columns.
• NEVER use Cosine Distance (<=>) on normalized embeddings. MUST use Inner Product (<#>) instead.
• NEVER create an HNSW index without specifying m and ef_construction parameters. Relying on defaults is forbidden.
• NEVER use L2 Distance (<->) unless the operator has explicitly confirmed Euclidean distance is required.
• NEVER query a vector column that lacks an HNSW index. HALT and create the index first.

Mandatory Behaviors

• The agent MUST verify embedding normalization status before selecting a distance operator. Check with SELECT vector_norm(embedding) FROM table LIMIT 5.
• HNSW indexes MUST use m = 16, ef_construction = 64 as baseline parameters. For tables exceeding 1M rows, MUST increase to m = 24, ef_construction = 128.
• Every vector similarity query MUST be validated with EXPLAIN ANALYZE to confirm index usage. A sequential scan on a vector column is a defect.
• Operator class MUST match the chosen distance operator: vector_ip_ops for <#>, vector_cosine_ops for <=>, vector_l2_ops for <->.
• When the embedding model changes, the agent MUST re-evaluate normalization status and rebuild the index if the operator class changes.
• The vector extension MUST be confirmed enabled before any vector DDL. HALT if missing.