Relative score fusion is a method for combining results from multiple retrieval systems by normalising each system’s raw scores onto a common scale and then blending them, typically with weights that control each system’s influence. It is an alternative to rank-based methods like reciprocal rank fusion.
The challenge it addresses is that different retrieval methods produce scores that are not directly comparable: a vector similarity score and a BM25 keyword score occupy entirely different ranges. Relative score fusion rescales each set of scores — for instance to a common zero-to-one range — so they can be meaningfully combined, then merges them, often weighting the semantic and lexical components according to how much each should matter.
The advantage over rank-based fusion is that it preserves information about how much better one result is than another, not just their order, which can yield better rankings when tuned well. The cost is that it requires choosing a normalisation scheme and weights, making it more sensitive to configuration. Both approaches are widely used in hybrid search, with the right choice depending on the data and how much tuning effort is available.