import time from sentence_transformers import SentenceTransformer # 实验参数 BUDGET_BYTES = 8000 MODEL_NAME = "intfloat/multilingual-e5-small" query = "如何实现 Java ERP 采购订单审批接口?" # 模拟知识库返回顺序 documents = [ { "id": "A", "title": "Java 开发规范", "text": "Java 项目应使用清晰的类名和方法名。" * 65, "required": False, }, { "id": "B", "title": "ERP 采购基础介绍", "text": "ERP 系统用于管理采购订单及供应商信息。" * 55, "required": False, }, { "id": "C", "title": "较长的采购背景资料", "text": "企业采购业务包含多个管理流程和操作步骤。" * 85, "required": False, }, { "id": "D", "title": "采购审批核心规则", "text": ( "采购订单提交后必须经过审批," "审批通过后才能执行后续采购操作。" ), "required": True, }, ] def size_bytes(doc): return len(str(doc).encode("utf-8")) # A:原版遇到超限直接结束 def select_break(docs): selected = [] used = 0 for doc in docs: size = size_bytes(doc) if used + size > BUDGET_BYTES: break selected.append(doc) used += size return selected # B:跳过过长文档,继续检查 def select_continue(docs): selected = [] used = 0 for doc in docs: size = size_bytes(doc) if used + size > BUDGET_BYTES: continue selected.append(doc) used += size return selected # 加载你已经下载的 Embedding 模型 print("加载语义模型...") model = SentenceTransformer(MODEL_NAME) # 预先计算文档向量 passages = [ "passage: " + doc["title"] + " " + doc["text"] for doc in documents ] doc_vectors = model.encode( passages, normalize_embeddings=True, ) # C:语义重排序 + 预算 def select_semantic(docs): query_vector = model.encode( "query: " + query, normalize_embeddings=True, ) scores = doc_vectors @ query_vector ranked = sorted( zip(docs, scores), key=lambda item: float(item[1]), reverse=True, ) selected = [] used = 0 for doc, score in ranked: size = size_bytes(doc) if used + size > BUDGET_BYTES: continue selected.append(doc) used += size return selected def evaluate(name, func): start = time.perf_counter() selected = func(documents) elapsed = (time.perf_counter() - start) * 1000 used = sum(size_bytes(doc) for doc in selected) has_required = any( doc["required"] for doc in selected ) print(f"\n【{name}】") print("保留文档:", [d["id"] for d in selected]) print("使用字节:", used) print("核心审批规则是否保留:", has_required) print("处理耗时(ms):", round(elapsed, 3)) evaluate("A 原版 break", select_break) evaluate("B 改进 continue", select_continue) evaluate("C 语义重排序", select_semantic) # D:直接保留前两条检索结果 def select_top2(docs): selected = [] used = 0 for doc in docs[:2]: size = size_bytes(doc) if used + size > BUDGET_BYTES: continue selected.append(doc) used += size return selected evaluate("D Top-2 精简", select_top2)