Hybrid RAG

Maximize RAG retrieval quality by fusing sparse (BM25) and dense (vector embedding) search results into a single, highly-relevant context set.

Hybrid RAG is the operational standard for robust Retrieval-Augmented Generation. It simultaneously executes two distinct search methods: sparse retrieval (keyword-based, like BM25) for high precision on specific terms, and dense retrieval (vector embeddings) for comprehensive semantic context. The system then fuses these results, often via Reciprocal Rank Fusion (RRF), to produce a unified, re-ranked document list. This dual-methodology approach consistently boosts both precision and recall, directly mitigating Large Language Model (LLM) hallucinations and making it critical for jargon-heavy domains (e.g., legal, medical).

https://www.meilisearch.com/blog/understanding-hybrid-search-rag-for-better-ai-answers