rerank

Fast rank fusion algorithms for hybrid search systems.

Overview

rerank provides zero-dependency implementations of rank fusion algorithms used to combine results from multiple retrieval systems. This is essential for hybrid search in RAG (Retrieval Augmented Generation) applications.

Features

Performance

Benchmarked on M1 Mac (100 results per list):

Usage

use rerank::{fuse_rrf, fuse_combmnz, RrfConfig};

// Results from BM25 (sparse) retrieval
let sparse = vec![
    ("doc1".to_string(), 0.9),
    ("doc2".to_string(), 0.7),
];

// Results from dense (vector) retrieval
let dense = vec![
    ("doc2".to_string(), 0.85),
    ("doc3".to_string(), 0.6),
];

// RRF fusion (recommended default)
let fused = fuse_rrf(sparse.clone(), dense.clone(), RrfConfig::default());
assert_eq!(fused[0].0, "doc2"); // appears in both

// CombMNZ for overlap reward
let fused = fuse_combmnz(&sparse, &dense);

Zero-Allocation Path

For hot paths, use preallocated output:

use rerank::{fuse_rrf_into, RrfConfig};

let mut output = Vec::with_capacity(200);
fuse_rrf_into(&sparse, &dense, RrfConfig::default(), &mut output);

3+ Sources

use rerank::{fuse_rrf_multi, RrfConfig};

let lists = vec![sparse, dense, knowledge_graph_results];
let fused = fuse_rrf_multi(&lists, RrfConfig::default());

When to Use What

• RRF (k=60): Start here. Works well without tuning.
• CombMNZ: When overlap between retrievers indicates relevance.
• Weighted: When you've tuned weights on your dataset.
• Borda: Voting-style aggregation.

no_std Support

[dependencies]
rerank = { version = "0.1", default-features = false, features = ["alloc"] }

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.