Reciprocal Rank Fusion
Quick Take
- In brief: RRF combines search rankings using each result’s position.
- Useful for: Merging retrievers whose scores use different scales.
- Main limit: It cannot recover a document that every retriever missed.
Overview
Keyword search may rank an exact error-code page first. Vector search may rank a related incident second. RRF can combine these lists without comparing their raw scores.
Core Definition
Reciprocal rank fusion (RRF) combines ranked lists by adding a score based on each result’s rank position.
Background and Evolution
Search systems often combine lexical and vector retrievers. Their scores may not be directly comparable, so rank-based fusion offers a simple alternative to score calibration.
How It Works
Each result gets a contribution based on its position in a list. Results ranked near the top of several lists receive stronger combined scores and move higher in the fused ranking.
Applications in Software Development and Testing
Use RRF to merge BM25 results for exact identifiers with vector results for related descriptions. Evaluate whether the merged list retrieves the evidence needed by real questions.
Strengths and Limitations
RRF does not require score scales to match and is simple to apply. Its result depends on the candidate lists and rank positions; it does not assess truth.
Comparison with Similar Terms
Hybrid search combines retrieval methods. RRF is one ranking-fusion method that can merge their result lists. Reranking applies a separate relevance judgment to candidates.
Best Practices
Compare RRF with each retriever alone and other fusion methods on the same query set. Track recall, answer support, latency, and candidate count.
Common Pitfalls
Do not expect fusion to find evidence missed by every retriever. Do not describe a high fused rank as proof that a passage is correct.
FAQ
Does RRF combine the raw search scores?
No. It combines rank positions, which avoids directly comparing scores on different scales.