app.py

from openai import OpenAI
from docx import Document
import numpy as np
import faiss
from sentence_transformers import SentenceTransformer
import os
import gradio as gr

# 配置参数
WORD_DOC_PATH = "知识库.docx"
VECTOR_INDEX_PATH = "faiss_index.index"
TEXT_DATA_PATH = "text_data.npy"
EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"

# 初始化模型和客户端
client = OpenAI(
    base_url='https://api-inference.modelscope.cn/v1/',
    api_key='7ed44f86-e2c6-4b85-9c4a-26eacfc2e5ee',
)
embedder = SentenceTransformer(EMBEDDING_MODEL)


def process_word_document():
    """处理Word文档并分块"""
    doc = Document(WORD_DOC_PATH)
    chunks = []
    current_chunk = []

    for para in doc.paragraphs:
        text = para.text.strip()
        if text:
            # 检查是否是新的条款
            if text.startswith("第") and "条" in text:
                # 如果当前块有内容，则保存当前块
                if current_chunk:
                    chunks.append(" ".join(current_chunk))
                    current_chunk = []
            # 将当前段落加入当前块
            current_chunk.append(text)

    # 保存最后一个块
    if current_chunk:
        chunks.append(" ".join(current_chunk))

    return chunks

def create_vector_store():
    """创建并保存向量存储"""
    if os.path.exists(VECTOR_INDEX_PATH):
        return

    chunks = process_word_document()
    embeddings = embedder.encode(chunks, convert_to_tensor=False)
    embeddings = np.array(embeddings).astype('float32')

    dimension = embeddings.shape[1]
    index = faiss.IndexFlatL2(dimension)
    index.add(embeddings)

    faiss.write_index(index, VECTOR_INDEX_PATH)
    np.save(TEXT_DATA_PATH, np.array(chunks))

def search_knowledge(query, top_k=6):
    """知识检索"""
    index = faiss.read_index(VECTOR_INDEX_PATH)
    text_data = np.load(TEXT_DATA_PATH, allow_pickle=True)

    query_embedding = embedder.encode([query], convert_to_tensor=False)
    query_embedding = np.array(query_embedding).astype('float32')

    distances, indices = index.search(query_embedding, top_k)
    return "\n".join([text_data[i] for i in indices[0]])

def respond(message, history, max_tokens, temperature, top_p=3):
    """Gradio响应函数（包含深度思考过程）"""
    # 检索相关知识
    context = search_knowledge(message)
    
    # 构建对话消息
    messages = [
        {"role": "system", "content": f"基于以下知识回答问题，如果不知道就说不知道：\n{context}"},
        {"role": "user", "content": message}
    ]

    # 初始化响应构建
    full_response = f"🔍 知识库相关内容：\n{context}\n\n💡 深度思考：\n"
    done_reasoning = False
    yield full_response  # 首次返回知识库内容

    # 流式生成响应
    response = client.chat.completions.create(
        model='deepseek-ai/DeepSeek-R1',
        messages=messages,
        stream=True,
        max_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p
    )

    for chunk in response:
        # 获取推理内容和回答内容
        reasoning_chunk = getattr(chunk.choices[0].delta, "reasoning_content", "") or ""
        answer_chunk = chunk.choices[0].delta.content or ""

        if reasoning_chunk:
            full_response += reasoning_chunk
            yield full_response
        
        if answer_chunk and not reasoning_chunk:
            if not done_reasoning:
                full_response += "\n\n✅ 最终答案：\n"
                done_reasoning = True
            full_response += answer_chunk
            yield full_response

    # 最终返回完整响应
    yield full_response

# 初始化向量存储
create_vector_store()

# 创建Gradio界面
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Slider(512, 2048, value=2048, step=1, label="最大Token数"),
        gr.Slider(0.1, 2.0, value=0.7, step=0.1, label="温度参数"),
        gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p采样"),
    ],
    title="制度文档问答系统",
    description="输入关于广西警察学院制度的问题进行问答",
)

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