from sentence_transformers import SentenceTransformer, util # 加载支持中英文的 Embedding 模型 MODEL_NAME = "intfloat/multilingual-e5-small" print("正在加载语义模型(第一次运行需要下载)...") model = SentenceTransformer(MODEL_NAME) # 模拟 ZKCoding 的 Java / ERP / WMS 知识库 documents = [ "库存管理系统记录商品入库、出库和剩余库存数量。", "ERP 采购订单系统负责供应商报价、采购审批和付款流程。", "Java Spring Boot 可以通过 REST Controller 实现 HTTP API。", "设备运维系统负责设备巡检、维修工单及故障记录。", ] # E5 模型使用 passage: 和 query: 前缀 passages = ["passage: " + doc for doc in documents] document_embeddings = model.encode( passages, convert_to_tensor=True, normalize_embeddings=True, ) questions = [ "如何追踪商品的进出及剩余数量?", "怎样管理供应商的报价和采购流程?", "Spring Boot 如何编写接口?", ] for question in questions: query_embedding = model.encode( "query: " + question, convert_to_tensor=True, normalize_embeddings=True, ) scores = util.cos_sim( query_embedding, document_embeddings, )[0] best_index = int(scores.argmax()) print("\n问题:", question) print("最相关文档:", documents[best_index]) print("相似度:", round(float(scores[best_index]), 4))