# 简介 360Zhinao-search 采用自研 BERT 类模型作为基座，进行多任务微调。在 C-MTEB 榜单的 Retrieval 任务上平均得分 75.05，目前排名第一。 [C-MTEB-Retrieval 榜单](https://huggingface.co/spaces/mteb/leaderboard)共包含 8 个不同领域下的[query, passage]相似度检索子任务, 使用 NDCG@10(Norma lized Discounted Cumulative Gain @ 10)作为评判指标。 | Model | T2Retrieval | MMarcoRetrieval | DuRetrieval | CovidRetrieval | CmedqaRetrieval | EcomRetrieval | MedicalRetrieval | VideoRetrieval | Avg | | :------------------------- | :---------: | :-------------: | :---------: | :------------: | :-------------: | :-----------: | :--------------: | :------------: | :-------------: | | **360Zhinao-search** | 87.12 | 83.32 | 87.57 | 85.02 | 46.73 | 68.9 | 63.69 | 78.09 | **75.05** | | AGE_Hybrid | 86.88 | 80.65 | 89.28 | 83.66 | 47.26 | 69.28 | 65.94 | 76.79 | 74.97 | | OpenSearch-text-hybrid | 86.76 | 79.93 | 87.85 | 84.03 | 46.56 | 68.79 | 65.92 | 75.43 | 74.41 | | piccolo-large-zh-v2 | 86.14 | 79.54 | 89.14 | 86.78 | 47.58 | 67.75 | 64.88 | 73.1 | 74.36 | | stella-large-zh-v3-1792d | 85.56 | 79.14 | 87.13 | 82.44 | 46.87 | 68.62 | 65.18 | 73.89 | 73.6 | # 环境部署 **基础环境为：Ubuntu:22.04+CUDA12.1** **创建一个python环境：** ``` conda create -n 360 python=3.10 -y conda activate 360 ``` **克隆 [360zhinao](https://github.com/Qihoo360/360zhinao.git)：** ``` git clone https://github.com/Qihoo360/360zhinao.git ``` **360Zhinao-search环境安装：** ``` cd /360zhinao/Retrieval pip install -r requirements.txt ``` # 模型下载从 [huggingface](https://huggingface.co/) 下载模型，如果不能访问，使用它的镜像网站：[HF-Mirror](https://hf-mirror.com/) ``` pip install -U huggingface_hub export HF_ENDPOINT=https://hf-mirror.com # 使用镜像网站 cd /360zhinao/Retrieval hf download qihoo360/360Zhinao-search --local-dir ./model ``` # 模型推理使用推理脚本测试模型运行情况：先修改模型加载为本地目录： ``` tokenizer = AutoTokenizer.from_pretrained('/360zhinao/Retrieval/model') model = AutoModel.from_pretrained('/360zhinao/Retrieval/model') ``` ``` cd /360zhinao/Retrieval/eval python test_model.py ``` 模型可以正常使用： ![1755763627995_image.png](https://suanjiayun.oss-cn-chengdu.aliyuncs.com/communityMirror_pic/1755763627995_image.png) # 模型使用开发源码中只有测试代码，而且使用是cpu加载模型；开发要求：使用cuda加载和推理模型，fastapi作为后端开发，html作为简单的前端页面；功能：1. 输入文本获取文本文本嵌入向量；2. 计算两个文本的相似度；3. 批量计算查询文本与文档的相似度使用前安装依赖包： ``` conda activate 360 cd /360zhinao/Retrieval/eval pip install fastapi uvicorn ``` 启动服务： ``` uvicorn fast_api:app --reload --host 0.0.0.0 --port 8080 ``` 获取访问地址在浏览器打开： ![1755763744745_image.png](https://suanjiayun.oss-cn-chengdu.aliyuncs.com/communityMirror_pic/1755763744745_image.png) ![1755763754387_image.png](https://suanjiayun.oss-cn-chengdu.aliyuncs.com/communityMirror_pic/1755763754387_image.png) fastapi源码：文件保存在 `/360zhinao/Retrieval/eval/fast_api.py` ``` from fastapi import FastAPI, HTTPException from fastapi.middleware.cors import CORSMiddleware from fastapi.staticfiles import StaticFiles from fastapi.responses import FileResponse from pydantic import BaseModel from typing import List, Optional import numpy as np import torch from transformers import AutoModel, AutoTokenizer import time import logging # 配置日志 logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # 检查CUDA是否可用 device = torch.device("cuda" if torch.cuda.is_available() else "cpu") logger.info(f"使用设备: {device}") # 初始化模型和tokenizer logger.info("正在加载模型和分词器...") start_time = time.time() try: tokenizer = AutoTokenizer.from_pretrained('/360zhinao/Retrieval/model') model = AutoModel.from_pretrained('/360zhinao/Retrieval/model') # 将模型移动到GPU（如果可用） model = model.to(device) model.eval() # 设置为评估模式 logger.info(f"模型加载成功，耗时: {time.time() - start_time:.2f}秒") logger.info(f"模型已加载到: {next(model.parameters()).device}") except Exception as e: logger.error(f"模型加载失败: {str(e)}") raise e app = FastAPI(title="360Zhinao-search API", version="1.0.0") # 添加CORS中间件 app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"], ) # 请求模型 class EmbeddingRequest(BaseModel): texts: List[str] normalize: Optional[bool] = True class SimilarityRequest(BaseModel): text1: str text2: str class BatchSimilarityRequest(BaseModel): queries: List[str] documents: List[str] # 响应模型 class EmbeddingResponse(BaseModel): embeddings: List[List[float]] time_ms: float class SimilarityResponse(BaseModel): similarity: float time_ms: float class BatchSimilarityResponse(BaseModel): similarities: List[float] time_ms: float def get_embeddings(texts: List[str], normalize: bool = True) -> np.ndarray: """获取文本嵌入向量""" inputs = tokenizer( texts, padding=True, truncation=True, return_tensors='pt', max_length=512 ) # 将输入数据移动到与模型相同的设备 inputs = {key: value.to(device) for key, value in inputs.items()} with torch.no_grad(): last_hidden_state = model(**inputs, return_dict=True).last_hidden_state embeddings = last_hidden_state[:, 0] # 使用[CLS] token作为句子表示 if normalize: embeddings = torch.nn.functional.normalize(embeddings, dim=-1) embeddings = embeddings.cpu().numpy() # 将结果移回CPU并转换为numpy数组 return embeddings @app.get("/") async def read_root(): return FileResponse("static/index.html") @app.post("/embed", response_model=EmbeddingResponse) async def get_text_embeddings(request: EmbeddingRequest): """获取文本嵌入向量""" start_time = time.time() try: embeddings = get_embeddings(request.texts, request.normalize) elapsed_ms = (time.time() - start_time) * 1000 return EmbeddingResponse( embeddings=embeddings.tolist(), time_ms=elapsed_ms ) except Exception as e: logger.error(f"嵌入计算错误: {str(e)}") raise HTTPException(status_code=500, detail=str(e)) @app.post("/similarity", response_model=SimilarityResponse) async def calculate_similarity(request: SimilarityRequest): """计算两个文本的相似度""" start_time = time.time() try: embeddings = get_embeddings([request.text1, request.text2]) cos_sim = np.dot(embeddings[0], embeddings[1]) elapsed_ms = (time.time() - start_time) * 1000 return SimilarityResponse( similarity=float(cos_sim), time_ms=elapsed_ms ) except Exception as e: logger.error(f"相似度计算错误: {str(e)}") raise HTTPException(status_code=500, detail=str(e)) @app.post("/batch_similarity", response_model=BatchSimilarityResponse) async def calculate_batch_similarity(request: BatchSimilarityRequest): """批量计算查询文本与文档的相似度""" start_time = time.time() try: # 获取所有文本的嵌入 all_texts = [request.queries[0]] + request.documents embeddings = get_embeddings(all_texts) # 计算查询与每个文档的相似度 query_embedding = embeddings[0] doc_embeddings = embeddings[1:] similarities = [float(np.dot(query_embedding, doc_embedding)) for doc_embedding in doc_embeddings] elapsed_ms = (time.time() - start_time) * 1000 return BatchSimilarityResponse( similarities=similarities, time_ms=elapsed_ms ) except Exception as e: logger.error(f"批量相似度计算错误: {str(e)}") raise HTTPException(status_code=500, detail=str(e)) @app.get("/health") async def health_check(): """健康检查端点""" return { "status": "healthy", "model_loaded": True, "device": str(device), "gpu_available": torch.cuda.is_available(), "gpu_count": torch.cuda.device_count() if torch.cuda.is_available() else 0 } # 挂载静态文件 app.mount("/static", StaticFiles(directory="static"), name="static") if __name__ == "__main__": import uvicorn uvicorn.run(app, host="0.0.0.0", port=8000) ``` html测试界面：`/360zhinao/Retrieval/eval/static/index.html` ``` 360Zhinao-search 测试界面

360Zhinao-search 模型测试界面

基于在C-MTEB Retrieval任务排名第一的模型

文本相似度计算

文本 1:

文本 2:

文本嵌入生成

输入文本 (每行一个):

文档检索测试

查询语句:

文档列表 (每行一个文档):

```