app.py

import os
from fastapi import FastAPI, HTTPException, Depends, Header
from pydantic import BaseModel
from typing import List, Optional
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
import uvicorn

app = FastAPI()

# --- 配置区域 ---
# 替换为真实的社区仓库名和 GGUF 文件名
MODEL_REPO = "mradermacher/Qwen3-Reranker-0.6B-GGUF" 
MODEL_FILE = "Qwen3-Reranker-0.6B.Q5_K_M.gguf"
# 设置你的专属 API KEY 防止别人滥用你的免费资源
MY_API_KEY = os.getenv("API_KEY", "1qazxsw2")

# --- 1. 下载并加载模型 ---
print("Downloading model from Hugging Face Hub...")
hf_token = os.getenv("HF_TOKEN")
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE, token=hf_token)

print("Loading model via llama.cpp...")
llm = Llama(
    model_path=model_path,
    n_ctx=2048,      # 上下文窗口，4B 模型在 16G 内存下可设为 2048 或 4096
    n_threads=2,     # 对应免费 Space 的 2 vCPU
    verbose=False
)

# --- 2. 定义数据结构 ---
class RerankRequest(BaseModel):
    query: str
    documents: List[str]
    top_n: Optional[int] = None

class ModelList(BaseModel):
    object: str = "list"
    data: list

# --- 3. 鉴权依赖 ---
async def verify_api_key(authorization: str = Header(None)):
    if not authorization or authorization != f"Bearer {MY_API_KEY}":
        raise HTTPException(status_code=401, detail="Unauthorized: Invalid API Key")

# --- 4. 核心接口逻辑 ---
@app.get("/v1/models")
async def list_models():
    return {
        "object": "list",
        "data": [
            {
                "id": "qwen3-reranker-0.6b", # 这里填你在 Cherry Studio 里想看到的名字
                "object": "model",
                "created": 1700000000,
                "owned_by": "huggingface"
            }
        ]
    }

@app.post("/v1/rerank", dependencies=[Depends(verify_api_key)])
async def rerank(request: RerankRequest):
    query = request.query
    documents = request.documents
    
    results = []
    for idx, doc in enumerate(documents):
        # 注意：这里需要根据具体的 Qwen Reranker prompt 格式调整。
        # 大多基于 LLM 的 Reranker 要求输出特定的 prompt，让模型打分
        # 这里使用一种通用的相关性问答 Prompt 示例：
        prompt = f"Query: {query}\nDocument: {doc}\nScore the relevance from 0 to 100:"
        
        # 让模型生成很短的回复（例如分数数字）
        response = llm(
            prompt,
            max_tokens=2,
            stop=["\n"],
            echo=False
        )
        
        try:
            # 尝试从模型输出中解析数字分数
            text_output = response['choices'][0]['text'].strip()
            score = float(text_output) if text_output.isdigit() else 0.0
        except:
            score = 0.0 # 解析失败给 0 分

        results.append({
            "index": idx,
            "document": doc,
            "relevance_score": score
        })
    
    # 按照得分从高到低排序
    results.sort(key=lambda x: x["relevance_score"], reverse=True)
    
    # 如果用户请求了 top_n，则截断
    if request.top_n is not None:
        results = results[:request.top_n]
        
    return {"results": results}

# 根路由探活
@app.get("/")
def read_root():
    return {"status": "running", "model": MODEL_FILE}

if __name__ == "__main__":
    # HF Spaces 默认公开 7860 端口
    uvicorn.run(app, host="0.0.0.0", port=7860)