Reciprocal Rank Fusion (RRF) — AI Glossary

Reference details

Related terms

• Hybrid Search
• BM25
• Semantic Search

How it works

1. Each retriever returns its own ranked list, and only the rank positions are taken forward — the underlying scores are discarded.
2. A document's contribution from one list is 1 / (k + rank), where k is a constant that damps the influence of the very top positions.
3. Contributions are summed across lists, so a document ranked moderately by several retrievers can outrank one ranked first by a single retriever.
4. Because only ranks are used, no score normalisation is needed — which is the point, since a BM25 score and a cosine similarity are not on comparable scales.
5. The constant k controls how sharply rank one is favoured: small k makes the top position dominant, large k flattens the curve.

Example: Why consensus beats a single first place

With k = 60, a document ranked first in one list and absent from the other scores 1/61 ≈ 0.0164. A document ranked third in both scores 1/63 + 1/63 ≈ 0.0317 — nearly double, despite never placing first. That is the intended behaviour: agreement across independent retrievers is stronger evidence than one retriever's confidence. Lower k to 10 and the first-place document scores 1/11 ≈ 0.0909 against 2/13 ≈ 0.1538, so consensus still wins but by less — the constant is how you tune that balance.

Commonly confused with

The usual way hybrid search combines its two result lists, and its defining property is that it uses only rank position, never the underlying scores. That is deliberate: BM25 scores and cosine similarities are on incomparable scales, so any attempt to combine them numerically requires a normalisation that has to be tuned and re-tuned. Ranks need no such calibration, which is why this is the default.

When to use it

Reach for it when:

• Combining retrievers whose scores are on incomparable scales, such as lexical and dense.
• A robust default fusion is wanted without tuning per-retriever score normalisation.
• Agreement between retrievers should count for more than one retriever's top hit.

Reach for something else when:

• Score magnitudes carry information you want to preserve — rank discards it.
• Only one retriever exists, where there is nothing to fuse.
• A trained reranker is available and affordable, which generally outperforms rank fusion.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Reciprocal Rank Fusion (RRF) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path