Reciprocal rank fusion (RRF)

Reciprocal rank fusion combines lists using only positions. Each list contributes 1 divided by k plus the rank for every document it contains, and documents are sorted by the total. A document ranked highly in both lists rises to the top; one that appears in only one list still gets credit. It was introduced in a 2009 paper by Cormack, Clarke, and Büttcher, which used k equal to 60, still the common default.

Its appeal is robustness. Vector similarity and keyword relevance scores live on different scales that shift from query to query, so adding them requires fragile normalization. RRF ignores scores entirely, needs no training data, and is a few lines of code or a short SQL expression.

The constant controls how much top positions dominate: a larger k flattens the difference between rank 1 and rank 10, a smaller one sharpens it. You can also weight lists if one retriever is consistently more reliable for your data.

One caution: RRF scores only order results for a single query. They are not relevance scores, so don't apply a fixed threshold to decide whether anything matched. Use a vector similarity floor or a reranker for that.