Reciprocal Rank Fusion (RRF) offers a method to combine multiple search results, such as BM25 and vector scores, into a single ranking using only document positions. By ignoring raw scores, RRF avoids issues caused by incompatible scoring scales and eliminates the need for training data. Effective hybrid search systems prioritize low-latency retrieval across parallel paths to ensure the fusion step does not become a bottleneck.
Key points
Reciprocal Rank Fusion merges multiple ranked lists by summing the inverse of document positions, effectively balancing different search strategies.
Raw search scores like BM25 and cosine similarity should not be combined directly due to incompatible scales and different measurement objectives.
RRF is unsupervised, requiring no labeled training data to function, which makes it a robust default for hybrid search implementations.
The constant k in the RRF formula serves to mitigate the influence of high-variance rankings and ensures consistency across search sources.
The performance of a hybrid search system is primarily gated by the latency of the underlying retrieval layers rather than the fusion algorithm itself.