Curiosity · Concept
Hybrid Search (BM25 + Vector)
Pure dense vector search is great at paraphrases and concept-level matches but often misses exact identifiers, product codes, or rare technical terms. Pure BM25 is the opposite. Hybrid search runs both retrievers and combines the results — typically via Reciprocal Rank Fusion (RRF) or a weighted score — so the top-k you pass to the LLM contains the best of both. Most modern RAG stacks default to hybrid because it's a cheap, large quality win.
Quick reference
- Proficiency
- Intermediate
- Also known as
- dense-sparse retrieval, BM25 + vector
- Prerequisites
- BM25, embeddings, vector database
Frequently asked questions
What is hybrid search?
Hybrid search is a retrieval strategy that combines sparse keyword scoring (usually BM25) with dense vector similarity and fuses the ranked lists. It catches exact-token matches BM25 is good at and semantic matches dense retrieval is good at.
What is Reciprocal Rank Fusion (RRF)?
RRF scores each document as sum over retrievers of 1/(k + rank), then sorts. It needs no score calibration between retrievers and is the default fusion in Elasticsearch, OpenSearch, Weaviate, and most hybrid libraries. Typical k is 60.
When is hybrid worth the extra complexity?
Whenever your corpus contains proper nouns, codes, identifiers, or jargon that embeddings flatten together — which is most enterprise corpora. On very small or highly conversational corpora the gain is smaller.
How is hybrid different from reranking?
Hybrid fuses two first-stage retrievers. Reranking is a second-stage model (usually a cross-encoder) that rescores a shortlist. They compose: hybrid to get a broader shortlist, rerank to sharpen the top of it.
Sources
- Cormack, Clarke & Buettcher — Reciprocal Rank Fusion — accessed 2026-04-20
- Weaviate — Hybrid search explained — accessed 2026-04-20