Struct CrossEncoder 

pub struct CrossEncoder { /* private fields */ }

Cross-encoder that jointly scores (query, document) pairs for reranking.

Construct with CrossEncoder::new and call CrossEncoder::rerank to reorder a Vec<CandidateDoc> by cross-encoder relevance scores.

Implementations

impl CrossEncoder

pub fn new(config: CrossEncoderConfig) -> Self

Create a new cross-encoder with the given configuration.

pub fn score_pair(&self, query: &[f64], doc: &[f64]) -> f64

Compute the relevance score for a single (query, document) pair.

The embedding lengths are independently capped at CrossEncoderConfig::max_doc_length before scoring.

pub fn rerank( &mut self, query: &[f64], candidates: Vec<CandidateDoc>, ) -> Vec<RerankedDoc>

Rerank a list of candidate documents for the given query embedding.

Returns documents sorted in descending order of cross_encoder_score. Stats are updated after each call.

pub fn rerank_batch( &mut self, queries_and_candidates: Vec<(Vec<f64>, Vec<CandidateDoc>)>, ) -> Vec<Vec<RerankedDoc>>

Rerank multiple (query, candidates) pairs in a single call.

Each element of the input slice is processed independently. The method is equivalent to calling rerank for each pair sequentially.

pub fn dot_product(a: &[f64], b: &[f64]) -> f64

Raw dot product: Σ q_i · d_i.

Stops at the shorter slice length.

pub fn cosine_similarity(a: &[f64], b: &[f64]) -> f64

Cosine similarity: dot(a, b) / (|a| · |b|).

Returns 0.0 when either vector has near-zero norm to avoid division by zero; the epsilon used is 1e-10.

pub fn bilinear_score(query: &[f64], doc: &[f64], weights: &[f64]) -> f64

Diagonal bilinear form: Σ w_i · q_i · d_i.

Weights are cycled when the embedding dimension exceeds weights.len(). A zero-length weights slice returns 0.0.

pub fn linear_score( query: &[f64], doc: &[f64], weights: &[f64], bias: f64, ) -> f64

Affine linear model: dot(weights, q ⊙ d) + bias.

q ⊙ d is the element-wise product, capped at the minimum length of query and doc. Weights beyond the element-wise product length are ignored; product components without a corresponding weight are treated as having weight 0.0.

pub fn normalize_scores(docs: &mut [RerankedDoc])

Min-max normalize cross_encoder_score values to [0, 1] in-place.

When all scores are identical the method maps every score to 1.0 to avoid a degenerate zero-range normalization.

pub fn rank_changed(initial_order: &[&str], reranked: &[RerankedDoc]) -> usize

Count the number of documents whose position changed after reranking.

initial_order holds document IDs in the order returned by the first- stage retriever; reranked is the cross-encoder output. A document is counted as “changed” when its 0-indexed position in reranked differs from its position in initial_order. Documents present in only one list are not counted.

pub fn stats(&self) -> &CrossEncoderStats

Return a reference to the accumulated runtime statistics.