Struct CrossEncoderRerank 

pub struct CrossEncoderRerank { /* private fields */ }

Cross-encoder rerank layer.

Replaces every result’s (chunk_idx, score) similarity with a fresh score from the cross-encoder, then re-sorts. The reranker joins query and document text and runs a BERT-with-classifier forward pass per pair, which is structurally higher quality than the bi-encoder’s dual-tower similarity but O(candidates) cost.

Construct via Self::new. Holds an [Arc] to a crate::rerank::Reranker so the layer chain can be cheaply cloned and the model isn’t reloaded per call.

Auto-detect via query

The layer’s score-rewrite is unconditional: if it’s in the pipeline, it runs. Auto-detect (skip rerank for symbol-shaped queries) belongs at the call site that builds the layer chain, not here. The call site already knows the query; it can decide whether to push this layer into the chain at all.

Implementations

impl CrossEncoderRerank

pub fn new(reranker: Arc<Reranker>, query: String, candidates: usize) -> Self

Build a rerank layer over reranker for query. Limits the pool to candidates (typical: 100). Uses the default blend factor of 0.7 (heavy cross-encoder, tiebroken by bi-encoder).

pub fn with_blend(self, blend: f32) -> Self

Override the bi/cross-encoder blend factor. 0.0 = pure bi-encoder (rerank is a no-op), 1.0 = pure cross-encoder (replace).

Trait Implementations

impl RankingLayer for CrossEncoderRerank

fn apply(&self, items: &mut Vec<(usize, f32)>, chunks: &[CodeChunk])

Apply this layer’s transformation.