🔍 Sparse Lexical BM25 • Dense Vector Cosine • Reciprocal Rank Fusion (RRF)
Hybrid Search & RRF Fusion Studio
Eliminate the blind spots of pure vector search. Blend exact keyword matches (SKUs, IDs, acronyms) with semantic vector recall using Reciprocal Rank Fusion.
RRF Hyperparameters
k = 60
Standard industry default is k = 60. Lower k penalizes lower ranks aggressively; higher k flattens rank score distribution.
1.0x
1.0x
RRF Score Top-1
0.0328
Fused confidence
Recall Robustness
98.4%
Captures OOD & Exact IDs
Zero Rank Inversion
100%
Scale-independent math
Live Candidate Document Fusion
Rank-Based Score Merging| Fused Rank | Document Title / Snippet | BM25 Rank | Dense Rank | RRF Score |
|---|
📐 Why RRF Outperforms Score Normalization:
BM25 outputs unbounded positive floats $[0, \infty)$, while dense embeddings yield cosine angles $[-1, 1]$. Attempting linear min-max score weighting fails when score distributions shift. RRF uses ordinal rank positions, making it completely immune to score magnitude drift.