Struct Reranker 

pub struct Reranker { /* private fields */ }

Cross-encoder reranker orchestrator.

Owns a RerankBackend (model trunk + classifier head) and the tokenizer that produced the encodings the backend expects.

Construct via Self::from_pretrained. Use score_pairs to rank candidate (query, doc) text pairs.

Implementations

impl Reranker

pub fn from_pretrained(model_repo: &str) -> Result<Self>

Load a cross-encoder by HuggingFace repo ID.

Routes through crate::backend::load_reranker_cpu for now; GPU paths slot in here as feature-gated branches when added. The tokenizer is downloaded via the same hf-hub cache, so multiple sub-agent MCP processes share weights through ~/.cache/huggingface/hub/.

Errors

Returns an error if the model can’t be downloaded, lacks a classifier head (i.e., a bi-encoder was supplied by mistake), or fails to load.

pub fn score_pairs(&self, pairs: &[(&str, &str)]) -> Result<Vec<f32>>

Score a batch of (query, document) pairs.

Returns raw logits (sentence-transformers Identity activation — the canonical public score for ms-marco cross-encoders), one per input pair, in input order. Tokenizes with a (query, doc) tuple so token_type_ids are 0 for the query side, 1 for the doc side — the convention BERT cross-encoders are trained on. The tokenizer is pre-configured with LongestFirst truncation at the model’s max_position_embeddings, so callers don’t need to clip outputs.

Errors

Propagates tokenization or forward-pass errors.

pub fn max_tokens(&self) -> usize

Max sequence length supported by the underlying model.