What is Reciprocal Rank Fusion (RRF)?
Also called RRF
Definition
Reciprocal rank fusion is a method for merging ranked result lists using document positions rather than directly comparing their raw relevance scores. It is commonly used to combine results from complementary retrieval methods.
The operation of Reciprocal Rank Fusion (RRF) is explained.
RRF assigns contributions based on each document’s rank in each input list and adds them together. This allows lists with different scoring scales to be combined without treating those raw scores as equivalent.
The relevance of Reciprocal Rank Fusion (RRF) is explained.
A SaaS knowledge-base search can merge keyword results with vector results. An article that performs well in either list can contribute to the final ordering, but the merged list still requires relevance evaluation.
The evaluation of Reciprocal Rank Fusion (RRF) is explained.
Representative questions support comparisons of fused results with individual retrievers. Evaluation includes candidate limits and ranking settings. Fusion combines candidates but does not verify their factual claims.
Frequently asked questions about Reciprocal Rank Fusion (RRF)
Does RRF require identical score scales?
RRF uses rank positions rather than requiring comparable raw score scales.
Is RRF a language model?
RRF is a ranking-combination method, not a language model.
Can RRF fix missing candidates?
A document absent from every input result list cannot appear through fusion alone.