Hybrid search and reranking: a deeper look at RAG

Production-grade RAG systems require more than basic vector search. Hybrid search combines dense vector search (semantic similarity) with sparse full-text search (BM25) to improve recall, merging results using algorithms like Reciprocal Rank Fusion (RRF) or weighted average scoring. A reranking model (cross-encoder) then takes the top 50–100 hybrid search candidates and re-scores them by evaluating the query and document together, surfacing the most relevant 5–10 chunks for the LLM. The post includes a PostgreSQL SQL example implementing hybrid search with RRF using pgvector and tsvector, and explains where each stage (embeddings, hybrid search, reranking) belongs in the architecture stack.