What is Reciprocal Rank Fusion (RRF)?

Reciprocal Rank Fusion (RRF) is a method for combining multiple ranked lists of search results into a single unified ranking. Rather than relying on raw similarity scores – which can vary in scale and meaning across different retrieval systems – RRF assigns each result a score based solely on its position in each list, then sums these scores to produce a final ordering.

Because it operates on ranks rather than scores, RRF requires no calibration or normalization between systems. This makes it a practical and reliable choice for merging results from heterogeneous sources, such as keyword search, semantic search, or multiple vector indexes.

How RRF works

For each result in each ranked list, RRF computes a score using the formula 1 / (rank + k), where rank is the result’s position in the list and k is a constant (typically 60) that dampens the influence of very high-ranked results. These scores are then summed across all lists for each unique result, and the results are re-ranked accordingly.

The intuition is straightforward: a document that ranks highly across multiple retrieval methods is more likely to be genuinely relevant than one that ranks highly in only one. Results that appear consistently across lists accumulate higher scores, while those appearing in only one list are naturally discounted.

What RRF is used for

RRF is widely used in hybrid retrieval systems that combine keyword-based and semantic search, where each method produces its own ranked list and scores are not directly comparable. It is also applied in multi-index RAG architectures, where the same document corpus is indexed at different chunk sizes or with different embedding models. In these settings, RRF allows the system to benefit from multiple retrieval strategies simultaneously without requiring any model retraining or score normalization.