Reciprocal Rank Fusion: How Hybrid Search Really Works

This post introduces Reciprocal Rank Fusion (RRF) as a method for combining ranked lists from different search systems, specifically keyword and vector search. It explains the limitations of directly adding raw scores from BM25 and vector search due to incompatible scales and drift, and how RRF, by using document ranks, sidesteps these issues. RRF's formula is presented, along with an example demonstrating how consensus across retrievers outweighs a single top rank. The post highlights RRF's unsupervised nature and its application in hybrid search, recommendation systems, and other ranking scenarios. It also stresses the importance of retrieval speed over the fusion step in production pipelines, particularly for latency-sensitive applications and AI agents. The post notes that Redis Search implements RRF for hybrid search within a single query, serving as a fast retrieval layer for systems like Redis Iris.