Implement Hybrid Search for RAG Systems

Defines the hybrid-search architecture of running vector search and keyword search in parallel then fusing their candidates into one ranked result list, and covers four fusion methods: Reciprocal Rank Fusion for general-purpose use, linear weighted-sum for tunable balance, cross-encoder reranking for highest quality, and cascade filter-then-rerank for efficiency. Ships four templates: Reciprocal Rank Fusion (a reciprocal_rank_fusion function combining multiple ranked doc_id and score lists using the 1/(k+rank+1) formula with an optional per-list weight, plus a companion linear_combination function that normalizes vector-similarity and BM25 scores to a 0-1 range and blends them with an alpha weight), PostgreSQL hybrid search (a PostgresHybridSearch class whose setup_schema creates a documents table with a pgvector embedding column, a generated tsvector column for full-text search, an HNSW index for vector similarity, and a GIN index for text search, then runs a hybrid_search method that fuses vector and full-text results with RRF and supports a JSONB metadata filter), Elasticsearch hybrid search (an ElasticsearchHybridSearch class that creates an index mapping with both an analyzed text field and a dense_vector field set to cosine similarity, then runs a bool query combining a kNN vector clause with a text-match clause, each independently boostable and filterable), and a custom hybrid RAG pipeline wiring retrieval fusion into a generation step.

Use it when combining vector and keyword search results, choosing a fusion method, implementing hybrid search in PostgreSQL or Elasticsearch, or building a hybrid retrieval step for a RAG pipeline. Skip it for anything unrelated to hybrid search implementation or outside this scope.

Input is separate ranked result lists from a vector search and a keyword search, or a query needing both; output is a single fused ranking via RRF or linear combination, or a store-specific hybrid query (PostgreSQL or Elasticsearch) returning already-fused results.

Store-agnostic fusion functions (RRF, linear combination) in plain Python; PostgreSQL via pgvector and full-text search; Elasticsearch via its native dense_vector field type and kNN query clause combined with text matching in a single bool query.

Engineers building retrieval for search or RAG who need concrete fusion algorithms and store-specific hybrid query implementations rather than choosing between vector-only and keyword-only search. Especially useful when vector-only retrieval is missing exact-match queries - product codes, names, or acronyms - that keyword search catches but semantic similarity alone tends to rank low.