# Retrieval Score Normalization and Fusion
Status: public
Confidence: medium (0.725) (verified)
Last verified: 2026-06-02
Generation: ai_structured


## TL;DR

Score normalization and fusion explain how RAG systems combine keyword, vector, and other retrievers without assuming their raw scores are comparable.

## Core Explanation

Retrieval pipelines often blend BM25, vector search, metadata filters, and rerankers. Raw scores can live on incompatible scales, so a system may normalize scores, combine weighted scores, or fuse rankings instead of comparing raw values directly.

Agents should record the fusion method, candidate depth, weights, and evaluation metric before tuning retrieval. A change that improves one query family can hurt another if the test set does not cover both exact-term and semantic queries.

## Source-Mapped Facts

- OpenSearch documentation says the normalization processor normalizes and combines document scores from different query clauses. ([source](https://docs.opensearch.org/docs/latest/search-plugins/search-pipelines/normalization-processor/))
- Elasticsearch documentation describes reciprocal rank fusion as combining ranked result sets from multiple retrievers. ([source](https://www.elastic.co/docs/reference/elasticsearch/rest-apis/reciprocal-rank-fusion/))
- Weaviate documentation describes hybrid search as combining vector and keyword search results by fusing result sets. ([source](https://weaviate.io/developers/weaviate/search/hybrid))

## Further Reading

- [OpenSearch Normalization Processor](https://docs.opensearch.org/docs/latest/search-plugins/search-pipelines/normalization-processor/)
- [Elasticsearch Reciprocal Rank Fusion](https://www.elastic.co/docs/reference/elasticsearch/rest-apis/reciprocal-rank-fusion/)
- [Weaviate Hybrid Search](https://weaviate.io/developers/weaviate/search/hybrid)