Source: OpenSearch docs: Normalization processor

OpenSearch search-pipeline processor that normalizes and combines BM25 and vector sub-query scores for hybrid queries.

Contents

Summary

Describes the search-phase results processor that makes the hybrid query work: it sits between the query and fetch phases and rescales then merges the scores returned by each sub-query (for example a BM25 match and a neural or k-NN query).

Key claims

  • Introduced in OpenSearch 2.10.
  • Normalization techniques: min_max (default), l2, z_score.
  • Combination via arithmetic_mean, with optional per-clause weights; without weights, clauses count equally. z_score supports only arithmetic_mean.
  • Reciprocal rank fusion is documented on a separate page rather than in this processor.

Feeds

BM25

Suggest changes →