reciprocal rank fusion
A way to merge several rankings by summing 1 divided by (k plus the rank) for each document, using ranks alone and ignoring raw scores.
In more detail
Cormack, Clarke and Büttcher proposed it in 2009 with k set to 60. Because it never compares a BM25 score with a cosine similarity, it needs no tuning to combine keyword and vector results.
Example
score(d) = Σ 1 / (60 + rank_i(d))Related terms
- hybrid search — Combining keyword and semantic search results, so exact names and paraphrases are both found.
- reranker — A model that reorders a shortlist of retrieved passages by reading each one together with the query.
Taught in
- Keyword And Hybrid Search — AI Engineering: LLM Apps, RAG & Agents
- Capstone: The Tasks Assistant — AI Engineering: LLM Apps, RAG & Agents
References
- Cormack, Clarke and Büttcher: Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods (SIGIR 2009)
- Cormack, Clarke and Buettcher: Reciprocal Rank Fusion (SIGIR 2009)