frankensearch-rerank

Cross-encoder reranking for frankensearch using FlashRank and FastEmbed.

Overview

This crate provides cross-encoder reranking, which processes the query and each candidate document together through a transformer model for dramatically more accurate relevance scoring than bi-encoder (embedding) approaches. The primary implementation uses ONNX Runtime for inference with sigmoid activation on raw logits.

Cross-encoders cannot pre-compute embeddings, so they are used as a reranking step on a shortlist of candidates produced by faster retrieval methods.

(query, document) -> tokenize -> ONNX -> logit -> sigmoid -> score in [0, 1]

Key Types

• FlashRankReranker - ONNX Runtime cross-encoder reranker implementing the Reranker trait
• FastEmbedReranker - alternative reranker using the fastembed library (feature: fastembed-reranker)
• rerank_step - utility function that runs reranking on a candidate list with configurable top-k and minimum candidate thresholds
• DEFAULT_TOP_K_RERANK - default number of top results to keep after reranking
• DEFAULT_MIN_CANDIDATES - minimum candidates required before reranking is triggered

Model Layout

Required files in the model directory:

• onnx/model.onnx (preferred) or model.onnx (legacy)
• tokenizer.json

Features

Usage

use frankensearch_rerank::FlashRankReranker;

// Load a cross-encoder model
let reranker = FlashRankReranker::load("/path/to/flashrank/model")
    .expect("load reranker model");

// Rerank candidates (via the Reranker trait)
// let reranked = reranker.rerank(&cx, "search query", &documents, 10).await?;

Dependency Graph Position

frankensearch-core
  ^
  |
frankensearch-rerank
  ^
  |-- frankensearch-fusion (optional, feature: rerank)
  |-- frankensearch (root, optional, feature: rerank)

License

MIT