Azure AI Search runs full-text and vector queries in parallel and merges them with reciprocal rank fusion. Other search engines expose similar combinations with different tuning controls.
How does hybrid search improve RAG?
Keyword search protects exact matches while semantic search broadens recall. The combined list is more robust across mixed business data, where a question may contain both natural language and an invoice number. A vector database must support the text and vector paths or delegate one of them to another index.
When do we recommend hybrid search for a RAG system?
We add a keyword index when the corpus is full of identifiers: error codes, part numbers, clause references, product names. Dense embeddings match meaning and shrug at exact strings, and that gap shows up the first time a user pastes a code. On a corpus of plain prose, hybrid search earns less, and metadata filtering often cuts the candidate set more cheaply. In our guide to enterprise RAG on LangChain, we build retrieval in a fixed order: contextual chunking, then keyword plus vector fusion, then a reranker, then the evaluation loop, the order the published retrieval data supports.
Those gains come from Anthropic’s published retrieval evaluation, which we cite rather than claim as our own: adding contextual BM25 to contextual embeddings took the top-20 retrieval failure rate from 3.7% to 2.9%, and a reranker brought it to 1.9% (Anthropic, Contextual Retrieval, read 5 November 2026).
Does hybrid search remove the need for reranking?
No. Fusion combines initial rankings, while reranking applies a stronger second-pass model to the leading candidates. Measure the complete RAG pipeline, because better retrieval can still produce poor answers when chunking or generation fails. Metadata filtering should also run before results cross tenant or permission boundaries.