Reciprocal Rank Fusion, or RRF, is a simple and robust algorithm for combining ranked result lists from multiple retrieval systems into a single merged ranking. It is the most common way to fuse the outputs of vector search and keyword search in hybrid retrieval.
Its elegance is that it ignores the raw scores, which are hard to compare across different systems, and uses only the rank of each result. Each item receives a contribution based on the reciprocal of its position in each list — a high rank gives a large contribution, a low rank a small one — and the contributions are summed across all lists. Items that rank well in multiple systems rise to the top of the combined list.
Because RRF depends only on ranks, it sidesteps the thorny problem that vector similarity scores and BM25 scores live on completely different scales and cannot be added directly. This makes it reliable and easy to apply without tuning, which is why it is a default choice for hybrid search. The alternative — normalising and weighting raw scores, as in relative score fusion — can perform better when tuned but requires more care.