rerank-blend

Blend N RAG reranker score streams with weighted normalization or Reciprocal Rank Fusion.

use rerank_blend::{blend_weighted, blend_rrf, RrfOpts};

let dense = vec![("a", 0.91), ("b", 0.88), ("c", 0.55)];
let bm25  = vec![("a", 12.3), ("c", 8.1),  ("b", 3.4)];

let weighted = blend_weighted(&[(&dense[..], 0.7), (&bm25[..], 0.3)]);
let rrf      = blend_rrf(&[&dense[..], &bm25[..]], RrfOpts::default());

blend_weighted min-max normalizes each stream so scale differences don't dominate. blend_rrf is score-free (Cormack et al. 2009 RRF; default k=60). Zero deps. MIT or Apache-2.0.