Reciprocal Rank Fusion (RRF)

A method for combining multiple ranked lists into a single ranking without learned parameters.

For each item, RRF computes a score based on its rank in each input list: items ranked highly across multiple lists score higher than items ranked highly in only one list. The formula penalizes low ranks exponentially, so an item at rank 1 in one list and rank 100 in another scores lower than an item at rank 10 in both.

The technique requires no training data and works out of the box, making it useful for hybrid retrieval systems that combine different ranking signals (e.g., graph traversal + vector similarity).

First introduced by Cormack et al. (2009) for information retrieval, now commonly used in RAG architectures that fuse multiple retrieval approaches.

Related: 02-molecule—cascaded-retrieval-pattern, 06-atom—multi-granular-embeddings