Reciprocal Rank Fusion

Status: public · Confidence: medium (0.725) · Basis: verified_sources

## TL;DR

Reciprocal rank fusion is a rank aggregation method that combines multiple search result lists by rewarding documents that appear highly in any contributing list.

## Core Explanation

RAG systems often run several retrievers at once: BM25, dense vector search, metadata-filtered search, or domain-specific indexes. Their raw scores are not always comparable. RRF avoids score calibration by using rank positions instead.

The method is useful when lexical and semantic retrievers each catch different relevant documents. It is usually followed by deduplication, optional reranking, and source diversity controls so the final context is not dominated by near-duplicates.

## Source-Mapped Facts

- Elasticsearch documentation describes reciprocal rank fusion as a method for combining multiple ranked result sets. ([source](https://www.elastic.co/docs/reference/elasticsearch/rest-apis/reciprocal-rank-fusion))
- Azure AI Search documentation says reciprocal rank fusion is used to merge ranked results from parallel query executions in hybrid search. ([source](https://learn.microsoft.com/en-us/azure/search/hybrid-search-ranking))
- OpenSearch documentation describes a reciprocal rank fusion processor for search pipelines that combines scores from multiple result sets. ([source](https://docs.opensearch.org/docs/latest/search-plugins/search-pipelines/reciprocal-rank-fusion/))

## Further Reading

- [Elasticsearch reciprocal rank fusion](https://www.elastic.co/docs/reference/elasticsearch/rest-apis/reciprocal-rank-fusion)
- [Azure AI Search hybrid scoring](https://learn.microsoft.com/en-us/azure/search/hybrid-search-ranking)
- [OpenSearch reciprocal rank fusion](https://docs.opensearch.org/docs/latest/search-plugins/search-pipelines/reciprocal-rank-fusion/)