Reciprocal Rank Fusion Combines Vector and Keyword Search for Better RAG Retrieval

Retrieval-Augmented Generation (RAG) systems often struggle because vector search excels at semantic similarity but fails on exact tokens like error codes, while keyword search does the opposite. Reciprocal Rank Fusion (RRF), a technique introduced in a 2009 paper by Cormack, Clarke, and Buettcher, solves this by ignoring raw scores entirely and combining results based solely on document rankings from each search method. The core formula assigns each document a fused score by summing the inverse of a constant k plus its rank across all result lists, making scores from incompatible systems directly comparable. A damping constant of k=60 is the widely accepted default, ensuring top-ranked results carry more weight without completely overshadowing lower-ranked ones. Because RRF only requires ordered lists rather than raw scores, it can be applied to any combination of rankers, including dense retrievers, lexical indexes, and rerankers simultaneously.