# Fusion score normalization strategies > Source-anchored research for Fusion score normalization strategies. Run status: COMPLETED. Parent: [Atlas Hybrid Fusion and BM25 Diagnostics](https://llms-explorer.com/tree/mongodb-atlas-hybrid-fusion-bm25-frontier-20261001/) · 5 facets · 15 facts · page: https://llms-explorer.com/tree/fusion-score-normalization-strategies/ ## Definitions - RRF is not universally more relevant than score fusion. OpenSearch reports lower average NDCG@10 for RRF than its score-based hybrid baseline, while Bruch, Gai and Ingber report tuned convex combination outperforming RRF in their evaluated settings. These findings do not establish a winner for every collection. (high confidence) — [source](https://opensearch.org/blog/introducing-reciprocal-rank-fusion-hybrid-search/) ## How it works - RRF sums reciprocal rank contributions, conventionally 1/(k+rank). It avoids raw-score normalization but discards score gaps: a narrow first-place win and a large first-place win contribute equally when their ranks match. (high confidence) — [source](https://opensearch.org/blog/introducing-reciprocal-rank-fusion-hybrid-search/) - Min-max fusion depends on the scores included in the normalization population. Changing candidate limits can change extrema and fused ordering even if some documents retain the same raw scores. This follows from the formula; keep candidate selection fixed when comparing normalizers. (medium confidence) — [source](https://www.elastic.co/search-labs/blog/linear-retriever-hybrid-search) ## Problems, failure modes and limitations - Constant score lists make the usual min-max denominator zero. The fusion paper notes that zero-variance component scores cannot improve ordering and explicitly leaves the near-zero-variance setting unstudied; do not assume its normalization results settle these edge cases. (low confidence) — [source](https://arxiv.org/html/2210.11934v2) ## Comparisons and alternatives - Rank fusion uses positions rather than raw score gaps. Relative-score fusion can therefore distinguish a large lead over the runner-up from a near tie that rank fusion treats identically. (high confidence) — [source](https://weaviate.io/blog/hybrid-search-fusion-algorithms) - The fusion analysis reports that tuned convex combinations of normalized scores outperform RRF in its tested in-domain and out-of-domain settings. This is evidence for evaluation on a target workload, rather than proof that either fusion family always wins. (low confidence) — [source](https://arxiv.org/html/2210.11934v2) - MongoDB recommends per-query lexical/vector weights rather than one static setting. The fusion research separately shows why normalization-weight equivalence must distinguish individual queries from a collection of queries; assess weight policies together with normalizer choice. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/vector-search/hybrid-search/hybrid-search-overview.md) ## Facts and statements - Raw scores from different retrievers can have incompatible ranges; weighting their unnormalized sum can let the larger numerical scale dominate. Normalize before comparing score-based weights, or use rank fusion to avoid direct score-scale comparison. (high confidence) — [source](https://www.elastic.co/search-labs/blog/linear-retriever-hybrid-search) - Min-max normalization applies (score-min)/(max-min) independently to each result list. It maps nondegenerate list extrema to 0 and 1 and preserves relative score spacing before weighted combination. (high confidence) — [source](https://www.elastic.co/search-labs/blog/linear-retriever-hybrid-search) - Bruch, Gai and Ingber find RRF sensitive to its parameters and learned convex-combination fusion generally insensitive to the tested score-normalization choice. This is a study result, not permission to omit evaluation on a new target domain. (low confidence) — [source](https://arxiv.org/abs/2210.11934) - MongoDB documents input.normalization values none, sigmoid, and minMaxScaler; weights act on normalized inputs. The default combination method is avg, and expression supports custom arithmetic but is mutually exclusive with combination.weights. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/manual/reference/operator/aggregation/scoreFusion.md) - MongoDB sigmoid examples match the logistic transform 1/(1 + exp(-s)): score 1 becomes approximately 0.7311 and score 5 approximately 0.9933. Thus large positive scores compress near 1; bounded output alone does not establish calibrated relevance probabilities. (medium confidence) — [source](https://www.mongodb.com/docs/manual/reference/operator/aggregation/sigmoid.md) - For MongoDB fusion diagnostics, scoreDetails: true must be paired with explicit projection through $meta to display the metadata. Documented details include inputPipelineRawScore, pipeline weight, normalization, combination, and final value; MongoDB does not guarantee a stable scoreDetails output format. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/manual/reference/operator/aggregation/scoreFusion.md) - OpenSearch implementation planning separates normalization from combination in a search-phase processor between query and fetch. Its initial design pairs min-max normalization with arithmetic-mean combination, showing that these are distinct configuration decisions. (low confidence) — [source](https://github.com/opensearch-project/neural-search/issues/228) - The fusion analysis proves that, for an individual query with nondegenerate scores, a suitable change of convex-combination weight can make min-max and z-score normalization rank-equivalent. The paper explicitly warns that this does not automatically imply one equivalent weight across an entire query collection. (low confidence) — [source](https://arxiv.org/html/2210.11934v2)