Definition
Hybrid search uses multiple retrieval signals to produce one ranked result set. A common setup runs sparse retrieval, which can weight term matches, alongside dense retrieval, which compares learned vector representations. Each produces a candidate list for the same query.
With reciprocal rank fusion (RRF), a result receives a contribution based on its rank in each list where it appears. A result found by only one method can still enter the final ranking. Another approach combines scores after normalization or calibration. Simply adding raw sparse and dense scores is unreliable. Putting them on the same numerical scale does not guarantee useful weighting either.
Simple example
A support assistant receives the question “What does MaxRetryCount do when requests are throttled?” Sparse retrieval may surface the setting’s reference page. Dense retrieval may surface a passage on retry behavior under throttling, even if it does not mention MaxRetryCount. The combined results can include both passages.
Why it matters
A query can contain an exact identifier and describe a problem in words the documentation does not use. Sparse retrieval helps with the identifier. Dense retrieval may find the explanation. Hybrid search gives the ranking both signals. Whether that helps depends on the corpus and query mix. Compare it with each method alone on representative questions, including identifier-heavy and paraphrased queries.
One important nuance
When fusion operates only on returned candidate lists, it cannot recover a passage absent from both. In that setup, the number of results requested from each retriever matters: a small limit can cut off a useful passage before fusion. Evaluate candidate recall separately from the final ordering. A later reranker works on the fused candidates only.