Pgvector HNSW Index Selection

Enforces HNSW index selection over IVFFlat and correct distance operator usage for pgvector.

Overview

This tile mandates HNSW as the only vector index type and enforces the correct distance operator based on embedding normalization. Normalized embeddings MUST use Inner Product (<#>), which is mathematically equivalent to cosine similarity but faster. IVFFlat indexes are explicitly forbidden.

Reference

Distance Operator Selection

HNSW Parameter Guidelines

Index Creation Template

CREATE INDEX idx_{table}_embedding_hnsw
ON {table}
USING hnsw (embedding vector_ip_ops)
WITH (m = 16, ef_construction = 64);

Normalization Check

SELECT vector_norm(embedding) FROM {table} LIMIT 5;
-- All values ~1.0 = normalized

Dependencies

• supabase-mcp-verification -- root prerequisite.
• pgvector (vector) extension MUST be enabled.

Composition Position

• Stage: database-schema
• Priority: HIGH
• Incompatibilities: supabase/ivfflat-index -- MUST NOT be installed alongside this tile.
• Executes after table creation tiles and before any application-layer vector search integration.