Reciprocal Rank Fusion

reciprocalRankFusion combines ranked hit lists without requiring the scores to be on the same scale.

That matters when you want to merge:

• BM25 or TF-IDF text results
• Filtered or geo-constrained searches
• Approximate vector search results

Combine lexical and vector search

import {
  DocumentIndex,
  MatchQuery,
  TextFieldIndex,
  VectorFieldIndex,
  bigramVector,
  createSeededRandom,
  reciprocalRankFusion
} from "@tryformation/querylight-ts";

const textIndex = new DocumentIndex({ title: new TextFieldIndex() });
const vectorIndex = new VectorFieldIndex(8, 36 * 36, createSeededRandom(42));

textIndex.index({ id: "1", fields: { title: ["specialty coffee brewing"] } });
textIndex.index({ id: "2", fields: { title: ["coffee shops in berlin"] } });

vectorIndex.insert("1", [bigramVector("specialty coffee brewing")]);
vectorIndex.insert("2", [bigramVector("coffee shops in berlin")]);

const lexicalHits = textIndex.search(new MatchQuery("title", "coffee brewing"));
const vectorHits = vectorIndex.query(bigramVector("cofee bruwing"), 10);

const fusedHits = reciprocalRankFusion([lexicalHits, vectorHits]);

Combine structured search with a geo constraint

import {
  GeoPointQuery,
  MatchQuery,
  reciprocalRankFusion
} from "@tryformation/querylight-ts";

const lexicalHits = index.search(new MatchQuery("title", "specialty coffee"));
const geoHits = index.search(new GeoPointQuery("location", 52.52, 13.405));

const fusedHits = reciprocalRankFusion([lexicalHits, geoHits], {
  rankConstant: 20
});

Notes

• RRF uses rank positions, not raw scores, so lexical and vector scales do not need calibration first.
• Documents present in multiple result lists usually move up.
• rankConstant defaults to 60; smaller values reward top-ranked overlaps more aggressively.
• Use weights if one ranking should matter more than another.

Learn more

• Reciprocal rank fusion paper in the IR Anthology
• Data fusion in information retrieval on Wikipedia