What is reciprocal rank fusion?
Reciprocal rank fusion combines the ranked results of several retrieval methods into a single list. Each passage receives credit according to its position in every list where it appears, and those credits are added together. Because it relies on rank positions instead of raw scores, it can merge methods whose scores are on entirely different scales, such as keyword ranking and vector similarity.
Also called: RRF, Weighted RRF rerank, Rank fusion
Why scores from different methods cannot simply be added
A keyword score and a vector distance measure different things on different scales, and adding them lets whichever method produces larger numbers dominate. Normalising the scores helps only until the distribution shifts with a new query type. Fusing by rank avoids the problem entirely. A passage that several methods place near the top rises, while one favoured by a single method is kept but not over-trusted. Weighted fusion then lets each method's contribution be set according to measured performance.
- Rank positions are comparable across methods where raw scores are not.
- Agreement between methods is rewarded without any training data.
- Weights express a measured judgement about which method to trust more.
- The calculation is cheap and fully deterministic.
How reciprocal rank fusion works
- Each retrieval method returns its own ranked list for the question.
- Every passage is credited according to its position in each list.
- A smoothing constant limits how much the very top positions dominate.
- Each method's credit is multiplied by its configured weight.
- Credits are summed per passage and the passages are sorted by the total.
- The fused list is passed to reranking or directly to context assembly.
How reciprocal rank fusion is evaluated
Fusion is evaluated by comparing the fused list with each method's list on its own. Measure the following on labelled questions:
- hit rate of the fused list against each input method
- mean reciprocal rank of the fused list
- sensitivity of results to the weights assigned to each method
- question types where fusion performs worse than the best single method
- stability of results when the smoothing constant changes
Limitations, risks and common failure modes
- Weights tuned on one question set are applied unchanged to a different workload.
- A weak method included with equal weight drags strong results down.
- Fusion cannot recover a passage that no method retrieved.
- Rank-only fusion discards information about how confident each method was.