Hybrid Search with RRF Fixes Blind Spots in RAG Retrieval Pipelines

Production RAG systems that rely solely on vector search often fail to retrieve exact matches for error codes, product IDs, or other precise identifiers, while pure keyword search (BM25) breaks down when users paraphrase queries. Running both BM25 and dense vector search in parallel, then merging results using Reciprocal Rank Fusion (RRF), addresses both failure modes simultaneously. RRF ranks documents by their relative position across both result lists rather than trying to normalize their incompatible raw scores, using a smoothing constant typically set to 60. A minimal Python implementation combining BM25Okapi, SentenceTransformers, and RRF can be built in under 25 lines of code. This hybrid approach improves retrieval reliability and, in turn, the quality of responses generated by downstream language models.