Module reranking 

Re-ranking with cross-encoders for improved search quality

This module implements two-stage retrieval:

1. Fast retrieval with bi-encoders (semantic search)
2. Precise re-ranking with cross-encoders

Cross-encoders jointly encode query and document for more accurate relevance scoring, but at higher computational cost. By applying them only to top-k candidates, we get both speed and accuracy.

Features

• Multiple cross-encoder backends (local models, API-based)
• Score fusion strategies (linear, rank-based, learned)
• Batch processing for efficiency
• Result caching
• Diversity-aware re-ranking

Example

use oxirs_vec::reranking::{CrossEncoderReranker, RerankingConfig};

let config = RerankingConfig::default();
let reranker = CrossEncoderReranker::new(config)?;

// First stage: fast retrieval
let candidates = vector_store.search("machine learning", 100)?;

// Second stage: precise re-ranking
let reranked = reranker.rerank("machine learning", &candidates, 10)?;

Re-exports

pub use cache::RerankingCache;

pub use cache::RerankingCacheConfig;

pub use config::FusionStrategy;

pub use config::RerankingConfig;

pub use config::RerankingMode;

pub use cross_encoder::CrossEncoder;

pub use cross_encoder::CrossEncoderBackend;

pub use diversity::DiversityReranker;

pub use diversity::DiversityStrategy;

pub use fusion::ScoreFusion;

pub use fusion::ScoreFusionConfig;

pub use models::CrossEncoderModel;

pub use models::ModelBackend;

pub use models::ModelConfig;

pub use reranker::CrossEncoderReranker;

pub use reranker::RerankingOutput;

pub use reranker::RerankingStats;

pub use types::RerankingError;

pub use types::RerankingResult as Result;

pub use types::ScoredCandidate;

Modules

cache

Re-ranking result caching

config

Configuration for re-ranking

cross_encoder

Cross-encoder model wrapper for query-document relevance scoring

diversity

Diversity-aware re-ranking strategies

fusion

Score fusion strategies

models

Cross-encoder model backends

reranker

Main cross-encoder re-ranker implementation

types

Core types for re-ranking