app.py

import streamlit as st
import random
from langchain_community.llms import HuggingFaceHub
from langchain_community.embeddings import SentenceTransformerEmbeddings
from langchain_community.vectorstores import FAISS
from datasets import load_dataset
from opencc import OpenCC

# 使用 進擊的巨人 数据集
# 原数据集是是繁体中文，为了调试方便，将其转换成简体中文之后使用
if "data_list" not in st.session_state:
    st.session_state.data_list = []
    st.session_state.answer_list = []

if not st.session_state.data_list:
    try:
        with st.spinner("正在读取数据库..."):
            converter = OpenCC('tw2s') # 'tw2s.json' 表示繁体中文到简体中文的转换
            dataset = load_dataset("rorubyy/attack_on_titan_wiki_chinese")
            data_list = []
            answer_list = []
            for example in dataset["train"]:
                converted_answer = converter.convert(example["Answer"])
                converted_question = converter.convert(example["Question"])
                answer_list.append(converted_answer)
                data_list.append({"Question": converted_question, "Answer": converted_answer})
            st.session_state.answer_list = answer_list
            st.session_state.data_list = data_list
            st.success("数据库读取完成！")
            print("数据库读取完成！")
    except Exception as e:
        st.error(f"读取数据集失败：{e}")
        st.stop()

# 构建向量数据库 (如果需要，仅构建一次)
if "vector_created" not in st.session_state:
    st.session_state.vector_created = False
if not st.session_state.vector_created:
    try:
        with st.spinner("正在构建向量数据库..."):
            # all-mpnet-base-v2 是一个由 Sentence Transformers 库提供的预训练模型，
            # 专门用于生成高质量的句子嵌入（sentence embeddings）。
            # all-mpnet-base-v2 在多个自然语言处理任务上表现出色，包括语义相似度计算、
            # 文本检索、聚类等。它能够有效地捕捉句子的语义信息，并生成具有代表性的向量表示。
            st.session_state.embeddings = SentenceTransformerEmbeddings(model_name="all-mpnet-base-v2")
            st.session_state.db = FAISS.from_texts(st.session_state.answer_list, st.session_state.embeddings)
            st.success("向量数据库构建完成！")
            print("向量数据库构建完成！")
    except Exception as e:
        st.error(f"向量数据库构建失败：{e}")
        st.stop()
    st.session_state.vector_created = True

# 问答函数
if "repo_id" not in st.session_state:
    st.session_state.repo_id = ''
if "temperature" not in st.session_state:
    st.session_state.temperature = ''
if "max_length" not in st.session_state:
    st.session_state.max_length = ''
def answer_question(repo_id, temperature, max_length, question):
    # 初始化 Gemma 模型
    if repo_id != st.session_state.repo_id or temperature != st.session_state.temperature or max_length != st.session_state.max_length:
        try:
            with st.spinner("正在初始化 Gemma 模型..."):
                st.session_state.llm = HuggingFaceHub(repo_id=repo_id, model_kwargs={"temperature": temperature, "max_length": max_length})
                st.success("Gemma 模型初始化完成！")
                print("Gemma 模型初始化完成！")
                st.session_state.repo_id = repo_id
                st.session_state.temperature = temperature
                st.session_state.max_length = max_length
        except Exception as e:
            st.error(f"Gemma 模型加载失败：{e}")
            st.stop()

    # 获取答案
    try:
        with st.spinner("正在筛选本地数据集..."):
            question_embedding = st.session_state.embeddings.embed_query(question)
            question_embedding_str = " ".join(map(str, question_embedding))
            # print('question_embedding: ' + question_embedding_str)
            docs_and_scores = st.session_state.db.similarity_search_with_score(question_embedding_str)

            context = "\n".join([doc.page_content for doc, _ in docs_and_scores])
            print('context: ' + context)

            prompt = f"请根据以下知识库回答问题：\n{context}\n问题：{question}"
            print('prompt: ' + prompt)

            st.success("本地数据集筛选完成！")
            print("本地数据集筛选完成！")

        with st.spinner("正在生成答案..."):
            answer = st.session_state.llm.invoke(prompt)
            # 去掉 prompt 的内容
            answer = answer.replace(prompt, "").strip()
            st.success("答案已经生成！")
            print("答案已经生成！")
        return {"prompt": prompt, "answer": answer}
    except Exception as e:
        st.error(f"问答过程出错：{e}")
        return {"prompt": "", "answer": "An error occurred during the answering process."}

# Streamlit 界面
st.title("進擊的巨人 知识库问答系统")

col1, col2 = st.columns(2)
with col1:
    gemma = st.selectbox("repo-id", ("google/gemma-2-9b-it", "google/gemma-2-2b-it", "google/recurrentgemma-2b-it"), 2)
with col2:
    temperature = st.number_input("temperature", value=1.0)
    max_length = st.number_input("max_length", value=1024)

st.divider()

def generate_answer(repo_id, temperature, max_length, question):
    result = answer_question(repo_id, float(temperature), int(max_length), question)
    print('prompt: ' + result["prompt"])
    print('answer: ' + result["answer"])
    st.write("参考文字：")
    st.markdown(result["prompt"])
    st.write("生成答案：")
    st.write(result["answer"])

col3, col4 = st.columns(2)
with col3:
    if st.button("使用原数据集中的随机问题"):
        dataset_size = len(st.session_state.data_list)
        random_index = random.randint(0, dataset_size - 1)
        # 读取随机问题
        random_question = st.session_state.data_list[random_index]["Question"]
        origin_answer = st.session_state.data_list[random_index]["Answer"]
        print('[]' + str(random_index) + '/' + str(dataset_size) + ']random_question: ' + random_question)
        print('origin_answer: ' + origin_answer)

        st.write("随机问题：")
        st.write(random_question)
        st.write("原始答案：")
        st.write(origin_answer)
        generate_answer(gemma, float(temperature), int(max_length), random_question)

with col4:
    question = st.text_area("请输入问题", "《进击的巨人》中都有哪些主要角色？")
    if st.button("提交输入的问题"):
        if not question:
            st.warning("请输入问题！")
        else:
            generate_answer(gemma, float(temperature), int(max_length), question)