What is Reciprocal Rank Fusion?
Reciprocal Rank Fusion is a rank aggregation method that combines multiple retrieval result lists by scoring each document according to the reciprocal of its rank in each list.
How It Works
Reciprocal Rank Fusion, often abbreviated RRF, is a simple and robust way to merge rankings from different retrievers. Instead of requiring comparable raw scores from BM25, dense retrieval, or other systems, it uses rank positions. A document that appears near the top of several lists receives a higher fused score. RRF is popular in hybrid RAG because lexical and vector scores are often not directly comparable, while ranks are easier to combine reliably.
Key Characteristics
- Combines ranked lists without needing score calibration across retrievers
- Rewards documents that appear high in multiple retrieval branches
- Works well for hybrid search that mixes BM25 and dense retrieval
- Uses a smoothing constant to reduce overemphasis on rank-one results
- Simple to implement and often competitive as a production baseline
Common Use Cases
- Merging BM25 and vector-search rankings for RAG
- Combining results from multiple query rewrites
- Fusing retrieval lists from different embedding models
- Building a robust baseline before learned rank fusion
- Reducing dependence on incompatible raw retrieval scores
Example
Loading code...Frequently Asked Questions
Why use Reciprocal Rank Fusion instead of averaging scores?
Raw scores from different retrievers are often on incompatible scales. RRF uses rank positions, which makes fusion more stable.
What does the RRF constant do?
The constant smooths the contribution of ranks so very top positions matter, but lower-ranked results can still contribute.
Is RRF a learned model?
No. It is a deterministic rank aggregation method, which makes it easy to implement and debug.
Does RRF replace reranking?
Not necessarily. RRF can merge candidates first, and a reranker can then perform more precise relevance scoring on the fused list.