hybrid-search-implementation

Combine vector and keyword search using Reciprocal Rank Fusion, which scores each document as the sum of 1/(k + rank) across both result lists. Run both searches in parallel, then merge rankings with the RRF formula using k=60 as the default constant.

Reciprocal Rank Fusion is a method that combines multiple ranked result lists by summing 1/(k + rank) for each document across lists. It works well without tuning and handles the different score scales of vector similarity and BM25 keyword scores.

Create a table with a vector column indexed by HNSW and a generated tsvector column indexed by GIN. Query both with CTEs for vector and full-text search, then join results with FULL OUTER JOIN and compute an RRF score from the two rankings.

Elasticsearch 8.x supports native RRF through the rank parameter with sub_searches combining a match query and a kNN query. For older versions, combine script_score cosine similarity with a boosted match query in a bool should clause.

Use hybrid search when queries contain exact terms like names, codes, or domain-specific vocabulary that embeddings miss. Pure vector search handles semantic similarity well but often fails on precise keyword matching requirements.

Cross-encoder reranking scores query-document pairs jointly with a neural model, capturing relevance signals that independent vector and keyword scoring miss. Apply it to the top fused candidates, typically 50, then return the reranked top results.