Hybrid Search RRF Pattern

Fuses semantic vector search with lexical full-text search using Reciprocal Rank Fusion in a PL/pgSQL RPC.

Overview

This tile produces a single PL/pgSQL RPC that combines pgvector cosine-distance results with PostgreSQL tsvector full-text results. Reciprocal Rank Fusion (RRF) merges the two ranked lists into one unified score using the formula score = 1/(k + rank), where k defaults to 60. The fused score surfaces rows that rank well in either or both retrieval methods.

Reference

RRF Formula

semantic_score  = 1.0 / (k + semantic_rank)
fulltext_score  = 1.0 / (k + fulltext_rank)
combined_score  = semantic_score + fulltext_score

Required Extensions

Required Indexes

RPC Signature

hybrid_search(
  query_embedding vector,
  query_text      text,
  match_count     int DEFAULT 20,
  k               int DEFAULT 60
) RETURNS TABLE (id, combined_score, semantic_score, fulltext_score)

Dependencies

• pgvector-hnsw-index-selection — MUST be installed to guarantee correct HNSW index configuration before this tile operates.

Incompatibilities

• supabase/ivfflat-index — IVFFlat indexes on the same column conflict with the HNSW requirement. NEVER install both on the same vector column.

Composition Position

This tile executes after pgvector-hnsw-index-selection has confirmed the HNSW index exists. It produces a callable RPC that downstream tiles or application code invoke for search. It does not manage index creation itself.