Hybrid search
Also: hybrid retrieval, reciprocal rank fusion, RRF
Runs a lexical retriever such as BM25 and a semantic retriever in parallel and fuses their result lists, either by a weighted sum of normalized scores or by reciprocal rank fusion (RRF), which uses only rank positions and needs no tuning.
RRF(d) = Σᵢ 1 / (k + rankᵢ(d)), k ≈ 60
Example
With k = 60, a document ranked #2 by both retrievers scores 2/62 ≈ 0.0323, beating one ranked #1 by BM25 but #5 by dense search: 1/61 + 1/65 ≈ 0.0318. Consensus wins.