Score fusion is the process of combining relevance scores from multiple retrieval methods into a single ranking. It is the step that makes hybrid search work, merging the output of vector similarity search and keyword search so that the final results reflect both signals at once.
The core difficulty is that scores from different methods are not directly comparable. A cosine similarity from vector search and a BM25 score from keyword search live on entirely different scales, so you cannot simply add them. Score fusion provides principled ways to reconcile them — either by ignoring the raw values and combining ranks, or by normalising the scores onto a common scale and blending them with weights.
The two dominant approaches are reciprocal rank fusion, which uses only the position of each result and is robust without tuning, and relative score fusion, which normalises and weights the actual scores for potentially better but more configuration-sensitive results. Choosing and tuning a fusion method is a key part of building effective hybrid search, since it directly shapes how semantic and lexical relevance are balanced in the final answer.