Also known as: hybrid retrieval, lexical + dense, BM25 + embeddings
Hybrid search combines lexical retrieval (BM25) with dense retrieval (embeddings) into one ranked candidate set. Each method catches what the other misses, so the union is more recall-complete than either alone.
Hybrid search runs two retrieval methods over the same corpus and merges their results. The classical pairing is BM25 (lexical, keyword-based) plus dense embedding similarity. The two methods are good at different things:
Combine them and you cover both failure modes.
Three combination strategies, in order of adoption:
In practice teams ship RRF as the baseline or skip straight to a reranker on the union.
The k=60 is from the original Cormack-Clarke-Buettcher paper (2009) and is mostly a smoothing constant: it dampens the contribution of top-1 hits relative to top-10 hits, so a document ranked first in BM25 doesn’t dominate the fusion if it’s ranked 50th in dense. Larger k flattens the contribution curve (everyone matters about equally); smaller k sharpens it (top-1 dominates). In practice the choice between k=10 and k=100 moves NDCG@10 by less than half a point on most retrieval evals, which is why nobody tunes it. The constant is robust because RRF is rank-only — the actual score magnitudes never enter the fusion, just the ordering — and that scale-invariance is the point.
BM25 and dense embeddings fail in different directions, which is exactly what makes hybrid work. The union recovers more than either alone because the failure modes are uncorrelated, not because the scores are individually better.
Hybrid search improves recall (the relevant document is in the top-N) but doesn’t necessarily improve precision at the top (the most relevant document is at #1). Hybrid surfaces a better candidate set; a reranker reorders that set to put the actually-relevant document first. Production stacks almost always layer a reranker on top of hybrid first-pass.
Start with reciprocal rank fusion — no weighting needed. Move to a learned linear or logistic combiner only after you have an eval set with at least a few hundred labeled queries; tuning a single α on a small set tends to overfit.
Yes — the reranker can only reorder what first-pass surfaces. Lexical and dense fail in different ways, so dropping one usually shows up as a recall regression on the rare-token or paraphrased slice of your eval.
BM25 needs language-specific tokenization and stemming to work well, while a multilingual embedding model handles cross-lingual queries natively. In multilingual stacks, the dense leg often does most of the work and BM25 catches code, identifiers, and proper nouns.
Posts on the ZeroEntropy blog that reference hybrid search.
