Why do hybrid dense+sparse retrievers beat dense-only, and how does RRF combine them?

A **hybrid retriever** runs two searches in parallel and combines the results. One is dense, using an embedding model and an approximate nearest neighbor index. The other is sparse, almost always BM25 over an inverted index of exact tokens. Dense retrieval is strong on paraphrase. The query *how do I unsubscribe* finds the document titled *cancel my subscription* because both phrases map to similar vectors. BM25 cannot make that jump because the two phrases share no words. Sparse retrieval is strong on rare tokens. A query containing an error code, a customer ID, or a function name pulls back the exact document that contains that token. Dense smooths those tokens into nearby concepts and often misses them. Reciprocal Rank Fusion combines the two ranked lists by adding `1 / (k + rank)` from each retriever, where `k` is a small constant like 60. Because the formula uses each candidate's rank in each list, not the raw score, you do not have to worry that BM25 and cosine produce numbers on different scales.

**Dense embeddings** and **BM25** fail on opposite query classes, which is the whole reason hybrid retrieval works. Dense handles paraphrase: *cancel my subscription* matches a doc titled *how to unsubscribe* because the vectors are close. BM25 handles rare-token lookup: a query containing `INC-48291` or `K8sNodeNotReady` finds documents with that exact token because the inverse document frequency factor lights up. The failure modes are structural. Dense embedding models are trained with a contrastive objective that intentionally smooths surface form. BM25 has no notion of meaning beyond token overlap. Combining them is the only way to cover both query classes with one retriever. Reciprocal Rank Fusion is the standard combiner because it operates on ranks rather than raw scores. For each candidate document, the RRF score sums `1 / (k + rank_r)` over each retriever `r`, with `k` typically set to 60. A candidate that appears in the top results of both retrievers gets a much higher RRF score than one that ranks first in only one list. The critical property is that ranks are invariant to score scale. Dense cosine and BM25 produce numbers on completely different scales whose calibration drifts with corpus and query distribution. Score-averaging fails when those distributions shift. RRF is robust because it never looks at raw scores.