from sentence_transformers import SentenceTransformer MODEL_NAME = "intfloat/multilingual-e5-small" class KnowledgeReranker: def __init__(self): self.model = SentenceTransformer(MODEL_NAME) def rerank(self, query, records): # 没有找到文档时,直接返回空列表 if not records: return [] # 提取现有 Knowledge Service 返回的内容 passages = [ "passage: " + r["title"] + " " + r["snippet"] for r in records ] # 将问题和文档转换为向量 query_vector = self.model.encode( "query: " + query, normalize_embeddings=True, ) document_vectors = self.model.encode( passages, normalize_embeddings=True, ) # 计算语义相似度 scores = document_vectors @ query_vector # 保留原始来源、版本、许可证等字段 ranked = sorted( zip(records, scores), key=lambda item: float(item[1]), reverse=True, ) return [ {**record, "semantic_score": float(score)} for record, score in ranked ] if __name__ == "__main__": # 模拟现有 Knowledge Service 已经授权返回的结果 records = [ { "citation_id": "WMS-001", "title": "库存管理", "snippet": "商品入库、出库和库存余额管理。", "source_url": "example://wms", }, { "citation_id": "ERP-001", "title": "采购管理", "snippet": "供应商报价、订单审批和付款管理。", "source_url": "example://erp", }, { "citation_id": "JAVA-001", "title": "Java API", "snippet": "Spring Boot REST API 接口开发。", "source_url": "example://java", }, ] reranker = KnowledgeReranker() result = reranker.rerank( "怎样管理供应商采购报价?", records, ) for item in result: print( item["citation_id"], round(item["semantic_score"], 4), item["source_url"], )