Reciprocal Rank Fusion (RRF)
Definition
Reciprocal Rank Fusion (RRF) is a rank-fusion method that combines multiple ranked result lists by assigning each item a score based on its rank position across those lists.
Why it matters
RRF provides a practical way to combine results from different retrievers without requiring their raw relevance scores to be directly comparable. This is useful in hybrid systems where lexical and dense retrieval may use very different scoring scales.
How it works
Multiple ranked lists
- Two or more retrievers produce ordered candidate lists, such as a BM25 list and a vector-search list.
Reciprocal rank scoring
- Each candidate receives a contribution based on its position in each ranked list.
- Higher-ranked appearances contribute more than lower-ranked appearances.
Fusion
- The contributions are combined to produce a new fused ranking.
- Candidates appearing strongly across multiple retrievers can rise in the fused result.
How Netsleek uses the term
Netsleek uses RRF as an established example of Retrieval Fusion within Hybrid Search. It helps explain that combining lexical and dense retrieval is not simply averaging unrelated raw scores.
Comparisons
- RRF vs Hybrid Search: Hybrid Search is the broader architecture combining retrieval methods. RRF is one method for fusing their ranked outputs.
- RRF vs Reranking: RRF fuses multiple ranked lists. Reranking applies a new ranking model or scoring stage to a candidate set.
- RRF vs BM25: BM25 produces lexical relevance scores. RRF can combine a BM25 ranking with rankings from other retrievers.
How Reciprocal Rank Fusion Connects Across AI Systems
Reciprocal Rank Fusion (RRF) does not operate in isolation. It connects to other established concepts across the AI search and retrieval process:
- Combines retrieval lists: Lexical, Sparse and Dense Retrieval can each contribute a ranked list.
- Supports hybrid search: Hybrid Search can use RRF to merge results from different retrieval families.
- Precedes reranking: A fused candidate set may then pass through a more expensive Reranking stage.
- Reduces score-scale dependence: RRF operates on rank positions rather than assuming all retriever scores share the same scale.
Related glossary concepts
- Hybrid Search
- Retrieval Fusion
- Lexical Retrieval
- Dense Retrieval
- BM25
- Reranking / Re-ranking
- Ranking Functions
Common misinterpretations
- RRF is not the only rank-fusion method.
- RRF does not prove that every hybrid search system uses it.
- Fusion and reranking are separate stages even when both affect final order.
Summary
Reciprocal Rank Fusion combines multiple ranked result lists using rank positions rather than raw score equivalence. It is a widely used method for fusing lexical and dense results in hybrid retrieval systems.