Awesome Search KG

Reciprocal Rank Fusion

3 min read

Reciprocal Rank Fusion

A score-based result merging algorithm that combines ranked lists from multiple retrieval systems without requiring calibrated scores.

Formula

For each document d across k ranked lists:

RRF(d) = Σ 1 / (rank(d, list_i) + c)

Where c is a constant (typically 60) that dampens the impact of high-ranked documents.

Why RRF

  • Score-agnostic: BM25 and dense vectors use incomparable scales; RRF only uses rank positions
  • Simple: No learned weights, no calibration
  • Robust: Consistently beats more complex fusion approaches in practice

Limitations

  • Treats all systems equally — no way to weight one retrieval path more than another
  • Loses fine-grained score signal; two items at rank 1 contribute identically even if scores differ greatly
  • Does not account for query-type variance (some queries may benefit more from lexical vs. semantic)
  • Criticized in RRF is Not Enough for losing signal in hybrid search contexts

Weighting, and why the constant matters

The constant is not decoration. Because it dominates the denominator at low ranks, it also sets the scale of the output, which is why raw RRF scores come out as small, hard-to-read numbers that barely separate. Weights fix both problems at once.

In MongoDB’s $rankFusion, where k = 60 is a fixed built-in, a weight of 30 on each of two pipelines rescales the summed score into a comfortable ~0.0–1.0 range: each pipeline contributes at most ~0.5, and a document ranked first in both lists lands at ~1.0.

Equal weights also produce structural ties: a document at rank 1 in list A and a document at rank 1 in list B score identically, since rank is all RRF sees. Assigning different weights per pipeline is the way to break them — and doubles as the knob for expressing that one retrieval path should count for more than another, which plain unweighted RRF cannot do.

Worked end-to-end with numbers in Reciprocal Rank Fusion and Relative Score Fusion.

Explainability

Because the contribution of each list is just w × 1/(k + rank), RRF is unusually easy to explain. Engines can expose the per-pipeline rank, weight, and resulting term — including a marker for lists the document did not appear in — making a fused score fully reconstructible. See Search Results Explainability.

RRF is also implementable directly in SQL — see Search using PostgreSQL for a PostgreSQL hybrid query fusing FTS and pgvector results.

RRF is the default fusion method in Hybrid Search pipelines combining:

Alternatives

  • Linear combination: α * dense_score + (1-α) * sparse_score — requires score normalization
  • Learned fusion: train a model to optimally weight retrieval paths per query type (e.g. LambdaMART via Metarank)
  • Interleaving: zip-merge ranked lists; a simple cold-start baseline and online-eval method
  • Wormhole Vectors: bridge across representation spaces at embedding level

Articles

Graph View

Backlinks