Definition
Reciprocal Rank Fusion (RRF) is a rank aggregation algorithm used in hybrid search to merge multiple retrieved document lists (e.g., keyword and semantic) into a single ranked output by summing the inverse of a document's rank across all lists. While it eliminates the need for complex score normalization between disparate retrieval methods, it trades off granular relevance magnitude for a simplified, position-based consensus.
It is a rank-merging logic, not a similarity metric or a document embedding model.
"A multi-heat race where the final winner is determined by the lowest combined rank across all heats, regardless of the time gap between runners."
- Hybrid Search(Prerequisite)
- BM25(Component)
- Dense Vector Retrieval(Component)
- Cross-Encoder Reranking(Post-processing)
Conceptual Overview
Reciprocal Rank Fusion (RRF) is a rank aggregation algorithm used in hybrid search to merge multiple retrieved document lists (e.g., keyword and semantic) into a single ranked output by summing the inverse of a document's rank across all lists. While it eliminates the need for complex score normalization between disparate retrieval methods, it trades off granular relevance magnitude for a simplified, position-based consensus.
Disambiguation
It is a rank-merging logic, not a similarity metric or a document embedding model.
Visual Analog
A multi-heat race where the final winner is determined by the lowest combined rank across all heats, regardless of the time gap between runners.