Definition
Reciprocal Rank Fusion (RRF) is a deterministic reranking algorithm used to combine multiple ranked result sets—typically from keyword-based BM25 and semantic vector search—into a single unified list. It calculates a final score by summing the reciprocals of a document's rank across all source lists, eliminating the need for complex score normalization.
It is a rank-based merging strategy, not a similarity metric or a model-based cross-encoder.
"A talent show where the winner is chosen by looking at the leaderboard positions from multiple different judges, rather than trying to average the judges' varied scoring points."
- Hybrid Search(Implementation Context)
- BM25(Prerequisite)
- Vector Retrieval(Prerequisite)
- Cross-Encoder(Alternative/Next Step)
Conceptual Overview
Reciprocal Rank Fusion (RRF) is a deterministic reranking algorithm used to combine multiple ranked result sets—typically from keyword-based BM25 and semantic vector search—into a single unified list. It calculates a final score by summing the reciprocals of a document's rank across all source lists, eliminating the need for complex score normalization.
Disambiguation
It is a rank-based merging strategy, not a similarity metric or a model-based cross-encoder.
Visual Analog
A talent show where the winner is chosen by looking at the leaderboard positions from multiple different judges, rather than trying to average the judges' varied scoring points.