app.py

# -*- coding: utf-8 -*-
# 財政部財政資訊中心 江信宗

import os
from dotenv import load_dotenv
load_dotenv()
import gradio as gr
from openai import OpenAI
from langchain_community.utils import user_agent
from langchain_groq import ChatGroq
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import Chroma
from langchain.chains import RetrievalQA
from langchain_community.document_loaders import WebBaseLoader, TextLoader
from langchain.prompts import PromptTemplate
from langchain.schema import Document
import resend
import requests
import re
import time

def load_documents(sources):
    documents = []
    for source in sources:
        try:
            if isinstance(source, str):
                if source.startswith('http'):
                    loader = WebBaseLoader(source)
                else:
                    loader = TextLoader(source)
                documents.extend(loader.load())
            elif isinstance(source, dict):
                documents.append(Document(page_content=source['content'], metadata=source.get('metadata', {})))
        except Exception as e:
            print(f"Error loading source {source}: {str(e)}")
    return documents

sources = [
    "TaxQADataSet_Slim1.txt",
    "TaxQADataSet_Slim2.txt",
    "TaxQADataSet_Slim3.txt",
    "TaxQADataSet_Slim4.txt",
    "TaxQADataSet_Slim5.txt",
    "TaxQADataSet_Slim6.txt",
    "TaxQADataSet_ntpc1.txt",
    "TaxQADataSet_ntpc2.txt",
    "TaxQADataSet_kctax.txt",
    "TaxQADataSet_chutax.txt",
    "LandTaxAct1100623.txt",
    "TheEnforcementRulesoftheLandTaxAct1100923.txt",
    "HouseTaxAct1130103.txt",
    "VehicleLicenseTaxAct1101230.txt",
    "TaxCollectionAct1101217.txt",
    "AmusementTaxAct960523.txt",
    "StampTaxAct910515.txt",
    "DeedTaxAct990505.txt"
]

documents = load_documents(sources)
print(f"\n成功載入 {len(documents)} 個檔案")

text_splitter = RecursiveCharacterTextSplitter(
    chunk_size=512,
    chunk_overlap=50,
    length_function=len,
    is_separator_regex=False,
    separators=["\n\n\n","\n\n", "\n", "。"]
)

split_docs = text_splitter.split_documents(documents)
print(f"分割後的文件數量：{len(split_docs)}")

embeddings = HuggingFaceEmbeddings(model_name="intfloat/multilingual-e5-large")
print(f"\n成功初始化 Microsoft 嵌入模型")

print(f"\n開始建立向量資料庫")
vectorstore = Chroma.from_documents(split_docs, embeddings, persist_directory="./Knowledge-base")
print(f"成功建立 Chroma 向量資料庫，共有 {len(split_docs)} 個文檔")

retriever = vectorstore.as_retriever(
    search_type="mmr",
    search_kwargs={
        "k": min(4, len(split_docs)),
        "fetch_k": min(20, len(split_docs)),
        "lambda_mult": 0.8
    }
)
print(f"檢索演算法：Maximum Marginal Relevance Retrieval")
print(f"檢索文檔數量：k={min(4, len(split_docs))}, fetch_k={min(20, len(split_docs))}")

template = """Let's work this out in a step by step way to be sure we have the right answer. Must reply to me in Taiwanese Traditional Chinese.
在回答之前，請仔細分析檢索到的上下文，確保你的回答準確完整反映了上下文中的訊息，而不是依賴先前的知識，在回應的答案中絕對不要提到是根據上下文回答。
如果檢索到的多個上下文之間存在聯繫，請整合這些訊息以提供更全面的回答，但要避免過度推斷。
如果檢索到的上下文不包含足夠回答問題的訊息，請誠實的說明，不要試圖編造答案。

上下文： {context}

問題： {question}

答案："""

PROMPT = PromptTemplate(
    template=template, input_variables=["context", "question"]
)

def create_chain(llm):
    return RetrievalQA.from_chain_type(
        llm=llm,
        chain_type="stuff",
        retriever=retriever,
        return_source_documents=True,
        chain_type_kwargs={"prompt": PROMPT}
    )
print(f"成功建立 RAG Chain")

def initialize_llm(api_key):
    return ChatGroq(
        groq_api_key=api_key,
        model_name='llama-3.3-70b-versatile'
    )

def generate_insight_questions(query, api_key):
    llm = initialize_llm(api_key)
    prompt = f"""Let's work this out in a step by step way to be sure we have the right answer. Must reply to me in "Traditional Chinese".
    根據以下回答，生成3個相關的洞察問題：

    原始問題： {query}

    請提供3個簡短但有深度的問題，這些問題應該符合：
    1. 與原始問題緊密相關
    2. 準確重新描述原始問題
    3. 引導更深入的解決原始問題

    請直接列出這3個問題，每個問題一行，不要添加編號或其他文字。
    """
    try:
        response = llm.invoke(prompt)
        if hasattr(response, 'content'):
            questions = response.content.split('\n')
        else:
            questions = str(response).split('\n')
        while len(questions) < 3:
            questions.append("提供更多地方稅資訊")
        return questions[:3]
    except Exception as e:
        print(f"Error generating insight questions：{str(e)}")
        return ["提供更多地方稅資訊", "提供其他地方稅問題", "還想了解什麼地方稅目"]

def answer_question(query, api_key):
    try:
        gr.Info("檢索地方稅知識庫中......")
        llm = initialize_llm(api_key)
        chain = create_chain(llm)
        result = chain.invoke({"query": query})
        answer = result["result"]
        insight_questions = generate_insight_questions(query, api_key)
        while len(insight_questions) < 3:
            insight_questions.append("提供更多地方稅資訊")
        return answer, insight_questions[:3]
    except Exception as e:
        return f"抱歉，處理您的問題時發生錯誤：{str(e)}", []

def split_questions(query):
    questions = re.split(r'[？！。　]', query)
    return [q.strip() for q in questions if q.strip()]

def answer_multiple_questions(query, api_key):
    questions = split_questions(query)
    all_answers = []
    all_insight_questions = []
    for question in questions:
        answer, insight_questions = answer_question(question, api_key)
        if len(questions) > 1:
            all_answers.append(f"【問題】{question}\n答案：{answer}")
        else:
            all_answers.append(answer)
        all_insight_questions.extend(insight_questions)
    if len(questions) > 1:
      combined_answer = "\n\n\n".join(all_answers)
    else:
      combined_answer = "\n".join(all_answers)
    selected_insight_questions = all_insight_questions[:3]
    return combined_answer, selected_insight_questions

def get_tax_law(tax_type):
    tax_law_dict = {
        "房屋稅": "房屋稅條例",
        "地價稅": "土地稅法",
        "土地增值稅": "土地稅法",
        "增值稅": "土地稅法",
        "契稅": "契稅條例",
        "娛樂稅": "娛樂稅法",
        "印花稅": "印花稅法",
        "使用牌照稅": "使用牌照稅法",
        "牌照稅": "使用牌照稅法",
        "稅捐稽徵法": "稅捐稽徵法",
        "綜合所得稅": "所得稅法",
        "所得稅": "所得稅法",
        "遺產稅": "遺產及贈與稅法",
        "贈與稅": "遺產及贈與稅法",
        "營業稅": "營業稅法"
    }
    return tax_law_dict.get(tax_type, "無稅法")

def fetch_law_summary(query, tax_law, keywords):
    url = "https://ttc.mof.gov.tw/Api/GetData"
    headers = {
        "Content-Type": "application/x-www-form-urlencoded; charset=UTF-8",
        "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/130.0.0.0 Safari/537.36",
        "accept": "application/json, text/javascript, */*; q=0.01",
        "accept-encoding": "gzip, deflate, br, zstd",
        "accept-language": "zh-TW,zh;q=0.9,en-US;q=0.8,en;q=0.7",
        "referer": "https://ttc.mof.gov.tw/"
    }
    gr.Info("檢索法令彙編函釋中......")
    version_payload = {
        "FunctionID": "FB10001",
        "ObjParams[TaxAct]": tax_law,
        "ObjParams[TaxVer]": "請選擇",
        "ObjParams[Chapter]": "請選擇",
        "ObjParams[Article]": "請選擇",
        "ObjParams[Content]": "",
        "ObjParams[Operator01]": "0",
        "ObjParams[Content01]": "",
        "ObjParams[Operator02]": "0",
        "ObjParams[Content02]": ""
    }
    try:
        version_response = requests.post(url, data=version_payload, headers=headers)
        version_response.raise_for_status()
        version_data = version_response.json()
        if version_data["Code"] == "1" and "Table1" in version_data["Data"]:
            latest_version = "請選擇"
            for item in version_data["Data"]["Table1"]:
                if item["TaxAct"] == tax_law:
                    latest_version = item["TaxVer"]
                    break
            if latest_version == "請選擇":
                print(f"未找到 {tax_law} 的對應版本，使用預設選項。")
        else:
            gr.Warning("無法獲取稅法版本資訊，使用預設選項。")
            latest_version = "請選擇"
    except Exception as e:
        print(f"獲取稅法版本時發生錯誤：{str(e)}")
        latest_version = "請選擇"
    all_results = []
    for keyword in keywords:
        payload = {
            "FunctionID": "FB10001",
            "ObjParams[TaxAct]": tax_law,
            "ObjParams[TaxVer]": latest_version,
            "ObjParams[Chapter]": "請選擇",
            "ObjParams[Article]": "請選擇",
            "ObjParams[Content]": keyword,
            "ObjParams[Operator01]": "0",
            "ObjParams[Content01]": "",
            "ObjParams[Operator02]": "0",
            "ObjParams[Content02]": ""
        }
        try:
            response = requests.post(url, data=payload, headers=headers)
            response.raise_for_status()
            data = response.json()
            if data["Code"] == "1" and "Table" in data["Data"]:
                all_results.extend(data["Data"]["Table"])
        except Exception as e:
            print(f"檢索關鍵字 '{keyword}' 的法令彙編函釋時發生錯誤：{str(e)}")
    if all_results:
        summary = f"<h3>相關法令彙編函釋檢索結果（{tax_law} {latest_version}）：</h3>"
        unique_results = {}
        for result in all_results:
            tax_sn = result.get('TaxSN', '')
            if tax_sn not in unique_results:
                unique_results[tax_sn] = result
        for index, result in enumerate(list(unique_results.values())[:20]):  # 限制為前20個唯一結果
            client = OpenAI(
                api_key=os.environ.get("YOUR_API_TOKEN"),
                base_url="https://api.sambanova.ai/v1",
            )
            system_prompt = f"""
            請判斷以下函釋內容與user提問的內容。
            請給出一個0到100之間的相關性百分比，不要任何說明或理由。
            回答格式為：
            相關性：XX%

            函釋內容：```{result['Content']}```
            """
            prompt = f"""```{query}```"""
            max_retries = 2
            retry_delay = 4
            for attempt in range(max_retries):
                try:
                    response = client.chat.completions.create(
                    model="Meta-Llama-3.1-405B-Instruct",
                    messages=[
                            {"role": "system", "content": system_prompt},
                            {"role": "user", "content": prompt}
                        ],
                        temperature=0.7
                    )
                    relevance_percentage = response.choices[0].message.content.strip()
                    break
                except Exception as e:
                    if (attempt == 0) or (attempt == max_retries - 1):
                        try:
                            response = client.chat.completions.create(
                                model="Meta-Llama-3.1-70B-Instruct",
                                messages=[
                                    {"role": "system", "content": system_prompt},
                                    {"role": "user", "content": prompt}
                                ],
                                temperature=0.7
                            )
                            relevance_percentage = response.choices[0].message.content.strip()
                            break
                        except Exception as e2:
                            relevance_percentage = "相關性：0%"
                            break
                    else:
                        print(f"Retrying in {retry_delay} seconds...")
                        time.sleep(retry_delay)
                        retry_delay *= 2
            try:
                percentage = int(relevance_percentage.split("：")[1].strip().rstrip('%'))
            except ValueError:
                print(f"Warning: Could not parse relevance percentage from '{relevance_percentage}'")
                percentage = 0
            if percentage > 0:
                summary += f"""
                <details>
                    <summary style="cursor: pointer; color: #0066cc;">{result['Title']} （相關性：{percentage} %）</summary>
                    <p>{result['Content']}</p>
                </details>
                """
        return summary
    else:
        return "<p>未檢索到相關法令彙編函釋。</p>"

def llm_openai_api(query, answer):
    client = OpenAI(
        api_key=os.environ.get("YOUR_API_TOKEN"),
        base_url="https://api.sambanova.ai/v1",
    )
    user_prompt = f"""
    「題目：{query}
    答案：{answer}」
    請詳細分析答案內容後，依據與題目相關性最高的稅目名稱及最多3個重點關鍵字回應我，提供的3個重點關鍵字不能與稅目名稱相同，問題與答案中的稅目名稱列入TaxName，關鍵字列入KeyWord，只須根據格式回應，不要寫其他的。

    # 回應字典格式範例：
    {{"TaxName": "地價稅", "KeyWord": "宿舍用地,醫護人員"}}
    """
    try:
        response = client.chat.completions.create(
            model='Meta-Llama-3.1-405B-Instruct',
            messages=[
                {"role": "system", "content": "Must reply to user in Traditional Chinese."},
                {"role": "user", "content": user_prompt}
            ],
            temperature=0.7,
            top_p=0.9
        )
        return response.choices[0].message.content.strip()
    except Exception as e:
        print(f"檢索法令彙編函釋 API Key！Error： {str(e)}")
        gr.Warning(f"檢索法令彙編函釋 API Key 額度不足！！")
        return '{"TaxName": "", "KeyWord": ""}'

def handle_interaction(query, api_key, state):
    gr.Info("開始處理問題，請稍待片刻......")
    start_time = time.time()
    if state is None:
        state = {"history": []}
    if not api_key:
        resend.api_key = os.environ["YOUR_USE_API_KEY"]
        params: resend.Emails.SendParams = {
            "from": "Tax_KM <onboarding@resend.dev>",
            "to": ["antivir7@gmail.com"],
            "subject": "地方稅知識庫 API KEY",
            "html": f"<strong>檢索內容：<br>{query}</strong>",
        }
        try:
            email_response = resend.Emails.send(params)
            print(f"Email sent successfully. Response：{email_response}")
        except Exception as e:
            print(f"Failed to send email：{str(e)}")
        api_key = os.getenv("YOUR_API_KEY")
    query = convert_punctuation(query)
    answer, insight_questions = answer_multiple_questions(query, api_key)
    questions = split_questions(query)
    if len(questions) == 1:
        api_response = llm_openai_api(query, answer)
        tax_name = ""
        keywords = []
        print(f"LLM剖析：{api_response}")
        try:
            response_dict = eval(api_response)
            tax_name = response_dict.get("TaxName", "")
            keywords = response_dict.get("KeyWord", "").split(",")
        except:
            print("剖析相關法令彙編函釋失敗！！")
            print(f"Tax Law: {tax_law}")
            print(f"Keywords: {keywords}")
        tax_law = get_tax_law(tax_name)
        law_summary_content = fetch_law_summary(query, tax_law, keywords)
    else:
        law_summary_content = ""
        gr.Info(f"多個問題不會提供法令彙編函釋檢索結果。")
    state["history"].append((query, answer))
    while len(insight_questions) < 3:
        insight_questions.append("提供更多地方稅資訊")
    end_time = time.time()
    gr.Info(f"AI知識庫已答覆，執行時間： {(end_time - start_time):.2f} 秒。")
    return answer, insight_questions[0], insight_questions[1], insight_questions[2], state, query, law_summary_content

def convert_punctuation(text):
    return text.replace('?', '？').replace(',', '，').replace('!', '！').replace(' ', '　')

def clear_outputs():
    return "", "", gr.update(value="", visible=False)

custom_css = """
.query-input {
    background-color: #B7E0FF !important;
    padding: 15px !important;
    border-radius: 10px !important;
    margin: 0 !important;
}
.query-input textarea {
    font-size: 18px !important;
    background-color: #ffffff;
    border: 1px solid #f0f8ff;
    border-radius: 8px;
}
.answer-box {
    background-color: #FFF5CD !important;
    padding: 10px !important;
    border-radius: 10px !important;
    margin: 0 !important;
}
.answer-box textarea {
    font-size: 18px !important;
    background-color: #ffffff;
    border: 1px solid #f0f8ff;
    border-radius: 8px;
}
.center-text {
    text-align: center !important;
    color: #ff4081;
    text-shadow: 2px 2px 4px rgba(0,0,0,0.1);
    margin-bottom: 0 !important;
}
#submit-btn {
    border-radius: 10px !important;
    border: none !important;
    background-color: #ff4081 !important;
    color: white !important;
    font-weight: bold !important;
    transition: all 0.3s ease !important;
    margin: 0 !important;
}
#submit-btn:hover {
    background-color: #f50057 !important;
    transform: scale(1.05);
}
.insight-btn {
    border-radius: 10px !important;
    border: none !important;
    background-color: #4dd8e2 !important;
}
.insight-btn:hover {
    background-color: #00bcd4 !important;
}
.gr-form {
    background-color: #e8f5e9 !important;
    padding: 15px !important;
    border-radius: 10px !important;
}
.api-key-input {
    background-color: #FFCFB3 !important;
    padding: 15px !important;
    border-radius: 10px !important;
    margin: 0 !important;
}
.text-background {
    font-size: 18px !important;
    padding: 5px !important;
    border-radius: 10px !important;
    border: 2px solid #B7E0FF !important;
    margin: 0 !important;
}
.clear-button {
    color: white !important;
    background-color: #000000 !important;
    padding: 5px !important;
    border-radius: 10px !important;
    margin: 0 !important;
}
.clear-button:hover {
    background-color: #000000 !important;
    transform: scale(1.05);
}
"""

with gr.Blocks(theme=gr.themes.Soft(), css=custom_css) as iface:
    gr.Markdown("""
    # 地方稅知識庫系統 - 財政部財政資訊中心
    > ### **※ RAG-based KM 以地方稅極少知識資料作示範，僅供參考，準確資訊請依地方稅稽徵機關回覆為準。系統部署：江信宗，LLM：Llama-3.1-70B。**
    """, elem_classes="center-text")
    with gr.Row():
        query_input = gr.Textbox(label="輸入您的問題，系統將基於學習到的知識資料提供相關答案。", placeholder="支援同時輸入多個問題，例如：問題1？問題2？", autofocus=True, scale=3, max_lines=5, elem_classes="query-input")
        api_key_input = gr.Textbox(label="輸入您的 API Key", type="password", placeholder="API authentication key", scale=1, elem_classes="api-key-input")
    answer_output = gr.Textbox(label="知識庫答案", interactive=False, max_lines=40, elem_classes="answer-box")
    with gr.Row():
        insight_q1 = gr.Button("洞察問題 1", visible=False, elem_classes=["insight-btn"])
        insight_q2 = gr.Button("洞察問題 2", visible=False, elem_classes=["insight-btn"])
        insight_q3 = gr.Button("洞察問題 3", visible=False, elem_classes=["insight-btn"])
    state = gr.State()
    current_question = gr.Textbox(lines=2, label="當前問題", visible=False)
    law_summary = gr.HTML(label="法令彙編函釋檢索", elem_classes="text-background", visible=False)
    with gr.Row():
        submit_btn = gr.Button("傳送", variant="primary", scale=3, elem_id="submit-btn")
        clear_button = gr.Button("清除", variant="secondary", scale=1, elem_classes="clear-button")
    gr.HTML(
        """
        <span style="font-size: 18px; color: black;">※ 財政部各稅法令函釋檢索系統：</span><a href="https://ttc.mof.gov.tw/" title="財政部各稅法令函釋檢索系統" style="font-size: 18px; color: red;">https://ttc.mof.gov.tw/</a>
        """
    )
    def update_ui(answer, q1, q2, q3, state, current_q, law_summary):
        return [
            answer,
            gr.update(value=q1, visible=bool(q1)),
            gr.update(value=q2, visible=bool(q2)),
            gr.update(value=q3, visible=bool(q3)),
            state,
            current_q,
            gr.update(value=law_summary, visible=bool(law_summary.strip()))
        ]
    submit_btn.click(
        fn=handle_interaction,
        inputs=[query_input, api_key_input, state],
        outputs=[answer_output, insight_q1, insight_q2, insight_q3, state, current_question, law_summary]
    ).then(
        fn=update_ui,
        inputs=[answer_output, insight_q1, insight_q2, insight_q3, state, current_question, law_summary],
        outputs=[answer_output, insight_q1, insight_q2, insight_q3, state, current_question, law_summary]
    )
    for btn in [insight_q1, insight_q2, insight_q3]:
        btn.click(
            lambda x: x,
            inputs=[btn],
            outputs=[query_input]
        )
    def clear_outputs():
        return "", "", gr.update(value="", visible=False)
    clear_button.click(
        fn=clear_outputs,
        inputs=[],
        outputs=[query_input, answer_output, law_summary]
    )

if __name__ == "__main__":
    if "SPACE_ID" in os.environ:
        iface.launch()
    else:
        iface.launch(share=True, show_api=False)