"""Cross-encoder reranking: score every (query, passage) pair and reorder.""" from __future__ import annotations import time import numpy as np from sentence_transformers import CrossEncoder from src.data.loader import Corpus, passage_text Run = dict[str, dict[str, float]] class Reranker: """Wraps a CrossEncoder so it can reorder a retrieval run in place.""" def __init__( self, model_name_or_path: str = "BAAI/bge-reranker-base", max_length: int = 384, batch_size: int = 32, device: str | None = None, fp16: bool = True, ) -> None: self.model = CrossEncoder( model_name_or_path, num_labels=1, max_length=max_length, device=device, ) if fp16 and device != "cpu": self.model.model.half() self.batch_size = batch_size def score(self, pairs: list[tuple[str, str]]) -> np.ndarray: """Raw relevance logits for a batch of (query, passage) pairs.""" return np.asarray( self.model.predict( pairs, batch_size=self.batch_size, show_progress_bar=False, convert_to_numpy=True, ), dtype=np.float32, ).reshape(-1) def rerank( self, queries: dict[str, str], run: Run, corpus: Corpus, top_k: int = 50, ) -> tuple[Run, list[float]]: """Rerank each query's candidates. Returns the new run and per-query ms.""" reranked: Run = {} timings_ms: list[float] = [] for qid, candidates in run.items(): doc_ids = [d for d, _ in sorted(candidates.items(), key=lambda kv: kv[1], reverse=True)] doc_ids = doc_ids[:top_k] if not doc_ids: reranked[qid] = {} timings_ms.append(0.0) continue pairs = [(queries[qid], passage_text(corpus, d)) for d in doc_ids] start = time.perf_counter() scores = self.score(pairs) timings_ms.append((time.perf_counter() - start) * 1000.0) reranked[qid] = {d: float(s) for d, s in zip(doc_ids, scores)} return reranked, timings_ms