app.py

import os
import gradio as gr
import torch # 确保导入 torch，因为 phi-2 模型需要
import logging

# LangChain 新版导入方式
from langchain_chroma import Chroma
from langchain_huggingface import HuggingFaceEmbeddings
# 注意：HuggingFacePipeline 从 0.2.x 开始推荐从 langchain_huggingface 导入
# 如果遇到问题，也可以尝试从 langchain_community.llms 导入
from langchain_huggingface.llms import HuggingFacePipeline # 或者 from langchain_community.llms import HuggingFacePipeline
from langchain.chains import RetrievalQA

# Transformers 库
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

logging.basicConfig(level=logging.INFO) # 更好的日志级别

# 设置路径
# 确保这些路径与您的预下载模型和向量库文件夹名称匹配
VECTOR_STORE_DIR = "./vector_store"
# 将模型名称指向本地预下载的路径
MODEL_NAME = "./hf_models_cache/models--uer--gpt2-chinese-cluecorpussmall"
EMBEDDING_MODEL_NAME = "./hf_models_cache/models--sentence-transformers--paraphrase-multilingual-mpnet-base-v2"


# 1. 轻量 LLM（uer/gpt2-chinese-cluecorpussmall）
print("🔧 加载生成模型...")
try:
    # 确保 tokenizer 和 model 是从正确的本地路径加载
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
        device_map="auto",
    )
    gen_pipe = pipeline(
        task="text-generation",
        model=model,
        tokenizer=tokenizer,
        max_new_tokens=256,
        temperature=0.5,
        top_p=0.9,
        do_sample=True,
    )
    llm = HuggingFacePipeline(pipeline=gen_pipe)
    print("✅ 生成模型加载成功。")
except Exception as e:
    logging.error(f"加载生成模型失败: {e}", exc_info=True)
    # 如果加载失败，可以考虑在这里退出或者给一个友好的错误提示
    llm = None # 确保 llm 为 None，避免后续报错
    print("❌ 生成模型加载失败，应用可能无法正常工作。")


# 2. 向量库和嵌入模型
print("📚 加载向量库和嵌入模型...")
try:
    embeddings = HuggingFaceEmbeddings(
        model_name=EMBEDDING_MODEL_NAME # 指向本地预下载的嵌入模型路径
    )
    vectordb = Chroma(persist_directory=VECTOR_STORE_DIR, embedding_function=embeddings)
    print("✅ 向量库加载成功。")
except Exception as e:
    logging.error(f"加载向量库失败: {e}", exc_info=True)
    vectordb = None # 确保 vectordb 为 None
    print("❌ 向量库加载失败，RAG功能将无法使用。")


# 3. RAG 问答链
qa_chain = None
if llm and vectordb: # 只有当LLM和向量库都成功加载时才构建RAG链
    try:
        retriever = vectordb.as_retriever(search_kwargs={"k": 3})
        qa_chain = RetrievalQA.from_chain_type(
            llm=llm,
            chain_type="stuff",
            retriever=retriever,
            return_source_documents=True
        )
        print("✅ RAG问答链构建成功。")
    except Exception as e:
        logging.error(f"构建RAG问答链失败: {e}", exc_info=True)
        print("❌ RAG问答链构建失败。")


# 4. 业务函数
def qa_fn(query):
    if not query.strip():
        return "❌ 请输入问题内容。"
    if not qa_chain: # 检查 qa_chain 是否已成功构建
        return "⚠️ 问答系统未完全加载，请稍后再试或检查日志。"
    try:
        result = qa_chain({"query": query})
        answer = result["result"]
        sources = result.get("source_documents", [])
        sources_text = "\n\n".join(
            [f"【片段 {i+1}】\n" + doc.page_content for i, doc in enumerate(sources)]
        )
        return f"📌 回答：{answer.strip()}\n\n📚 参考：\n{sources_text}"
    except Exception as e:
        logging.exception("问答失败：%s", e)
        return f"❌ 出现错误：{str(e)}\n请检查日志获取更多信息。"

# 5. Gradio UI
with gr.Blocks(title="数学知识问答助手", theme=gr.themes.Base()) as demo:
    gr.Markdown("## 📘 数学知识问答助手\n输入教材相关问题，例如：“什么是函数的定义域？”")
    with gr.Row():
        query_input = gr.Textbox(label="问题", placeholder="请输入你的问题", lines=2)
        output_box = gr.Textbox(label="回答", lines=15)
    submit_btn = gr.Button("提问")

    submit_btn.click(fn=qa_fn, inputs=query_input, outputs=output_box)

    gr.Markdown("---\n模型：uer/gpt2-chinese-cluecorpussmall + Chroma RAG | Powered by Hugging Face Spaces")


if __name__ == "__main__":
    demo.launch()