app.py

import gradio as gr
import numpy as np
import os
import torch
import pandas as pd
import faiss
from huggingface_hub import hf_hub_download
import time
import sys
import json

# 创建安全缓存目录
CACHE_DIR = "/home/user/cache"
os.makedirs(CACHE_DIR, exist_ok=True)

# 减少内存占用
# os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:32"
# torch.set_num_threads(1)
os.environ["OMP_NUM_THREADS"] = "2"
os.environ["TOKENIZERS_PARALLELISM"] = "false"  # 防止tokenizer内存泄漏

# 全局变量 - 移除了clip_model和clip_processor
index = None
metadata = None

def load_resources():
    """加载所有必要资源（768维专用）"""
    # 清理残留锁文件
    lock_files = [f for f in os.listdir(CACHE_DIR) if f.endswith('.lock')]
    for lock_file in lock_files:
        try:
            os.remove(os.path.join(CACHE_DIR, lock_file))
            print(f"🧹 清理锁文件: {lock_file}")
        except: pass
            
    global index, metadata
    
    # 仅当资源未加载时才初始化
    if index is None or metadata is None:
        print("🔄 正在加载所有资源...")
        
        # 加载FAISS索引（768维）
        if index is None:
            print("📥 正在下载FAISS索引...")
            try:
                INDEX_PATH = hf_hub_download(
                    repo_id="GOGO198/GOGO_rag_index",
                    filename="faiss_index.bin",
                    cache_dir=CACHE_DIR,
                    token=os.getenv("HF_TOKEN")
                )
                index = faiss.read_index(INDEX_PATH)
                
                # 验证索引维度
                if index.d != 768:
                    raise ValueError(f"❌ 索引维度错误：预期768维，实际{index.d}维")

                # if index and not index.is_trained:
                #     print("🔧 训练量化索引...")
                #     index.train(np.random.rand(10000, 768).astype('float32'))
                #     print("✅ 索引训练完成")
                    
                print(f"✅ FAISS索引加载完成 | 维度: {index.d}")
            except Exception as e:
                print(f"❌ FAISS索引加载失败: {str(e)}")
                raise
        
        # 加载元数据
        if metadata is None:
            print("📄 正在下载元数据...")
            try:
                METADATA_PATH = hf_hub_download(
                    repo_id="GOGO198/GOGO_rag_index",
                    filename="metadata.csv",
                    cache_dir=CACHE_DIR,
                    token=os.getenv("HF_TOKEN")
                )
                metadata = pd.read_csv(METADATA_PATH)
                print(f"✅ 元数据加载完成 | 记录数: {len(metadata)}")
            except Exception as e:
                print(f"❌ 元数据加载失败: {str(e)}")
                raise

from functools import lru_cache
@lru_cache(maxsize=100)
def predict(vector):
    """处理768维向量输入并返回答案"""
    start_time = time.time()
    print(f"输入向量维度: {np.array(vector).shape}")
    
    try:
        # 验证输入格式
        if not isinstance(vector, list) or len(vector) == 0:
            error_msg = "错误：输入格式无效"
            print(error_msg)
            return error_msg

        if len(vector) != 1 or len(vector[0]) != 768:
            error_msg = f"错误：需要1x768的二维数组，收到{len(vector)}x{len(vector[0]) if vector else 0}"
            print(error_msg)
            return error_msg
        
        # 添加实际处理逻辑
        vector_array = np.array(vector, dtype=np.float32)
        D, I = index.search(vector_array, k=3)
        
        results = []
        for i in range(3):
            try:
                result = metadata.iloc[I[0][i]]
                confidence = 1/(1+D[0][i])
                results.append(f"匹配结果 {i+1}: {result['source']} | 置信度: {confidence:.2f}")
            except Exception as e:
                print(f"结果处理错误: {str(e)}")
                results.append(f"结果 {i+1}: 数据获取失败")
        
        print(f"处理完成 | 耗时: {time.time()-start_time:.2f}秒")
        return json.dumps({
            "results": results  # 确保嵌套结构合法
        })
        
    except Exception as e:
        import traceback
        error_msg = f"处理错误: {str(e)}\n{traceback.format_exc()}"
        print(error_msg)
        return "处理错误，请重试或联系管理员"

# 创建简化接口
with gr.Blocks() as demo:
    gr.Markdown("## 🛍 电商智能客服系统 (768维专用)")
    gr.Markdown("**使用CLIP-vit-large-patch14模型 | 向量维度: 768**")
    
    with gr.Row():
        vector_input = gr.Dataframe(
            headers=["向量值"], 
            type="array",
            label="输入向量 (768维)",
            value=[[0.1]*768]  # 768维默认值
        )
        output = gr.Textbox(label="智能回答", lines=5)
    
    submit_btn = gr.Button("生成回答")
    submit_btn.click(
        fn=predict, 
        inputs=vector_input, 
        outputs=output
    )

# 启动应用
if __name__ == "__main__":
    # 预加载资源
    if index is None or metadata is None:
        print("🚀 启动前预加载资源...")
        try:
            load_resources()
        except Exception as e:
            print(f"⛔ 资源加载失败: {str(e)}")
            sys.exit(1)

    # 确保缓存目录存在
    # import pathlib
    # pathlib.Path(CACHE_DIR).mkdir(parents=True, exist_ok=True)

    try:
        dummy_vector = [0.1] * 768
        predict([dummy_vector])
    except:
        pass

    print("="*50)
    print("Space启动完成 | 准备接收请求")
    print(f"索引维度: {index.d if index else '未加载'}")
    print(f"元数据记录: {len(metadata) if metadata is not None else 0}")
    print("="*50)
    
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860
    )