RRF Reciprocal Rank Fusion

Reciprocal Rank Fusion (RRF) is an unsupervised, calibration-free multi-retriever rank aggregation algorithm; when merging heterogeneous search branches (e.g., BM25 scores $[0, 30+]$ and Dense cosine similarities $[0, 1]$), RRF discards incommensurable raw score scales, computing unified relevance strictly from ordinal rank positions: $\text{RRF}(d) = \sum_{m \in M} \frac{1}{k + r_m(d)}$, where $r_m(d)$ represents the rank position of document $d$ in retriever $m$ (1-indexed), and $k$ is a smoothing constant (standard empirical default $k = 60$).

Aggregating keyword sparse search and dense vector results in Hybrid RAG, and merging multi-query parallel search outputs.

Linear score weighting $\alpha S_{\text{dense}} + (1-\alpha) S_{\text{sparse}}$ demands fragile Min-Max / Z-score normalization and easily collapses when one branch produces outlier scores; RRF evaluates pure relative ranks, delivering unmatched cross-retriever robustness.

Derive the RRF scoring formula and explain how constant $k=60$ smoothens aggressive score gaps between top-ranked ranks?

Compare RRF scores between a document ranking #1 on a single branch vs ranking #5 across both branches?

How does Weighted RRF apply branch weights $w_m$ when domain requirements favor lexical or dense semantics?

Weigh RRF against supervised Learning to Rank (LTR / RankNet) regarding cold-start feasibility and annotation costs?

Analyze native RRF implementations and execution parameters within Elasticsearch and Qdrant Hybrid Search APIs?