What is reciprocal rank fusion (RRF)?
Reciprocal rank fusion (RRF) is a method for merging ranked result lists into one, scoring each document by the sum of 1 divided by a constant plus its rank in each list, so raw scores never need comparing.
Checked
Hybrid search produces two lists: one from keyword search and one from vector search. Their scores live on different scales, a BM25 score of 14.2 and a cosine similarity of 0.81 cannot be added. RRF sidesteps the problem by ignoring scores and using only each document's position in each list.
The method was published by Gordon Cormack, Charles Clarke and Stefan Büttcher at SIGIR 2009 in a short paper titled Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods. It was meant as a simple baseline and turned out to beat the more elaborate fusion methods it was compared against.
How the formula works
Each document gets a score of 1 / (k + rank) from every list it appears in, and the scores are summed. With the usual k of 60, a document ranked first by BM25 and third by vector search scores 1/61 + 1/63, about 0.032. A document ranked first by only one of the two scores 1/61, about 0.016. Agreement between retrievers wins, and a single outlier cannot dominate.
The constant k damps the gap between top ranks. Cormack and colleagues fixed k = 60 in a pilot study and found the choice was not critical: in their pilot table, mean average precision moved only from 0.2072 at k = 0 to 0.2147 at k = 80.
Where RRF ships
Elasticsearch
The rrf retriever uses the same formula, with rank_constant defaulting to 60 and rank_window_size defaulting to the requested result size.
OpenAI vector stores
When hybrid search is enabled on a vector store search, embedding and text weights control how reciprocal rank fusion balances semantic matches against keyword matches.
Where RRF fails, and what it is confused with
It throws away confidence
A near-exact match and a weak match both count as rank 1 if each tops its list. When one retriever is far more reliable for a query type, weighted fusion or a reranker does better.
It depends on the window
A document missing from one list gets nothing from it. If the window is too small, a passage that one retriever ranks 30th never gets the boost it deserves.
RRF versus score blending
Score blending normalises raw scores and combines them with a weight. It keeps more information but needs calibration per corpus. RRF needs none, which is why it is the common default.
RRF versus reranking
RRF only reorders what the retrievers returned, using positions. A reranker reads the query and each passage and judges relevance itself, so it usually runs after fusion, not instead of it.
How Motiontic's engineers apply it
Motiontic's engineers start hybrid retrieval with plain RRF because it needs no tuning, then measure recall@k on the client's eval set against weighted fusion and against a reranker on top. The window size is set from those measurements, since it decides how deep each retriever's results can reach.
Ways to merge keyword and vector results
| Reciprocal rank fusion | Weighted score blending | Reranker | |
|---|---|---|---|
| Uses | Ranks only | Normalised scores | Query and passage text |
| Tuning needed | Little (k, window) | Weights per corpus | Shortlist size, latency budget |
| Keeps score confidence | No | Yes | Produces its own scores |
| Extra model call | No | No | Yes |
| Typical place | First merge | First merge | After the merge |
Common questions
Why is k set to 60 in reciprocal rank fusion?
The original 2009 paper fixed k = 60 during a pilot and found results barely changed across a wide range of values. Elasticsearch kept 60 as its default rank_constant.
Is RRF the same as a reranker?
No. RRF merges lists using only positions and costs almost nothing. A reranker is a model that reads each query and passage pair, so it is slower and usually runs on the fused shortlist.
Can reciprocal rank fusion weight one list more than another?
Many implementations allow it. OpenAI's vector store hybrid search, for example, exposes separate embedding and text weights that shift the fusion toward semantic or keyword matches.
Does RRF need score normalisation?
No. That is its main advantage: it uses positions, so BM25 scores and similarity scores never have to be put on the same scale.
Related
Hybrid search
Keyword ranking and vector similarity, fused into one result list.
BM25
The keyword ranking function behind most full-text search, and half of hybrid search.
Reranker
A second pass that rereads the shortlist and puts the right passages first.
Recall@k
Did the right passages make the shortlist? The first number to check in RAG.
AI product development
New products and major AI features, built from scratch alongside the client's team.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.