Source: Elasticsearch docs: Reciprocal rank fusion retriever

Elasticsearch reference for the RRF retriever that fuses BM25 and kNN results by rank.

Contents

Summary

Documents the rrf retriever, which takes two or more child retrievers (for example a standard BM25 query and a knn query) and combines them with reciprocal rank fusion.

Key claims

  • Score per document is the sum of 1 / (rank_constant + rank) across child retrievers.
  • rank_constant defaults to 60; higher values give lower-ranked documents more influence.
  • rank_window_size sets how many results each child returns before fusion; larger is more accurate and slower.
  • Scroll, sort and rescore are unsupported; pagination past rank_window_size returns nothing.

Feeds

Hybrid search

Suggest changes →