Reciprocal Rank Fusion
Also called RRF
A simple way to merge several ranked lists: reward items for placing high in any of them.
Think of it like
Combining critics’ top-ten movie lists by giving points for rank — a film near the top of several lists wins, even if no one made it number one.
Example
Given keyword results and vector results, RRF gives each document a score of 1/(k + its rank) in every list and sums them, so a document ranked #2 in both beats one ranked #1 in only one.
How it actually works
RRF ignores the raw similarity scores, which are hard to compare across different retrievers, and looks only at position. Each list contributes 1/(k + rank) per document, with a small constant k (often 60) to dampen the top. Summing across lists produces one ranking. It’s popular because it needs no tuning and no score normalization, yet holds up surprisingly well.
For product teams
A no-knobs default for combining searches that just works, so the team ships fusion without a tuning project.
For engineers
score(d) = Σ over lists of 1/(k + rank_list(d)); sort descending. k≈60 is the common default.
Related
- Fusion Retrieval — The general technique it powers.
- Hybrid Search — A common pairing it merges.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome