Reciprocal Rank Fusion (RRF)

What is Reciprocal Rank Fusion (RRF)?

Reciprocal Rank Fusion (RRF) is an information retrieval technique that combines multiple ranked lists into a single result ranking by assigning scores based on each document's position in those lists.

Rather than requiring the original search methods to produce comparable relevance scores, RRF uses rank positions to calculate a combined score, with higher-ranked results receiving greater weight. It is commonly used in hybrid search systems that combine keyword-based retrieval with vector-based retrieval.

Key Features

Rank-Based Scoring: Calculates a combined score using a document's position in each ranked result list rather than relying on the original relevance scores.

Multiple Retrieval Sources: Combines results from different search methods, such as BM25-based keyword retrieval and vector search.

Rank Contribution Aggregation: Rewards documents that rank highly across one or more result lists by adding their reciprocal rank contributions.

Reduced Score Calibration Requirements: Avoids the need to directly compare raw scores from retrieval systems that may use different scoring scales.

Configurable Ranking Behaviour: Uses a constant parameter to control how strongly rank positions influence the combined score, alongside implementation-specific options such as the number of results considered.

Benefits

Improved Hybrid Search: Brings keyword-based and semantic retrieval results together in one ranked list.

More Balanced Result Coverage: Can surface relevant documents found by one retrieval method even when another method ranks them poorly or does not retrieve them.

Simplified Result Combination: Reduces the need to normalize or calibrate raw scores from different retrieval methods.

Greater Retrieval Flexibility: Allows organisations to combine multiple search approaches without replacing their underlying retrieval systems.

Support for Advanced Document Discovery: Helps search pipelines retrieve and consolidate candidates from multiple sources before additional reranking or downstream processing.

Conclusion

Reciprocal Rank Fusion provides a practical way to combine ranked results from multiple retrieval methods into one list. By using rank positions instead of comparing raw relevance scores, it can help hybrid search systems integrate keyword-based and vector-based retrieval.

In document management environments, RRF can improve how candidate documents are consolidated for discovery, while the final search quality still depends on the retrieval methods, indexed content and relevance of the returned results.