app.py

# -------------------------------
# 1. 匯入套件
# -------------------------------
import os, glob, time
from langchain.docstore.document import Document
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import FAISS
from langchain.chat_models import ChatHuggingFaceHub
from langchain.chains import RetrievalQA
from docx import Document as DocxDocument
import gradio as gr

# -------------------------------
# 2. 設定路徑
# -------------------------------
txt_folder = "out_texts"     # 放你的 .txt 檔
db_path = "faiss_db"
os.makedirs(db_path, exist_ok=True)

# -------------------------------
# 3. 建立 embeddings
# -------------------------------
embeddings_model = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")

# -------------------------------
# 4. 建立或載入向量資料庫
# -------------------------------
if os.path.exists(os.path.join(db_path, "index.faiss")):
    print("載入現有向量資料庫...")
    db = FAISS.load_local(db_path, embeddings_model, allow_dangerous_deserialization=True)
else:
    print("沒有資料庫，開始建立新向量資料庫...")
    txt_files = glob.glob(f"{txt_folder}/*.txt")
    docs = []
    for filepath in txt_files:
        with open(filepath, "r", encoding="utf-8") as f:
            docs.append(Document(page_content=f.read(), metadata={"source": os.path.basename(filepath)}))
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
    split_docs = text_splitter.split_documents(docs)
    print("產生向量嵌入中...")
    db = FAISS.from_documents(split_docs, embeddings_model)
    db.save_local(db_path)
    print("向量資料庫建立完成。")

# -------------------------------
# 5. Hugging Face 模型設定
# -------------------------------
HUGGINGFACE_API_TOKEN = os.getenv("HF_TOKEN")  # 建議在 Spaces Secrets 設定

MODEL_DICT = {
    "google/flan-t5-large": 512,
    "tiiuae/falcon-7b-instruct": 512
}

MAX_HOURLY_REQUESTS = 50
request_count = 0
last_reset_time = time.time()

# -------------------------------
# 6. RAG 主函式
# -------------------------------
def rag_generate_hfapi(query, model_name, segments=5, max_words=1500):
    global request_count, last_reset_time
    if time.time() - last_reset_time > 3600:
        request_count = 0
        last_reset_time = time.time()

    if request_count >= MAX_HOURLY_REQUESTS:
        return f"本小時生成次數已達上限 ({MAX_HOURLY_REQUESTS})，請稍後再試。", None

    llm = ChatHuggingFaceHub(
        repo_id=model_name,
        model_kwargs={"temperature": 0.7, "max_new_tokens": MODEL_DICT[model_name]},
        huggingfacehub_api_token=HUGGINGFACE_API_TOKEN
    )

    qa_chain = RetrievalQA.from_chain_type(
        llm=llm,
        retriever=db.as_retriever(search_type="similarity", search_kwargs={"k": 5}),
        return_source_documents=True
    )

    prompt = f"""請依據下列主題生成一篇文章：
主題：{query}
需求：
- 總共 {segments} 段
- 每段約 {max_words // segments} 字
- 總字數請控制在 {max_words} 字以內
- 請自動分段輸出
"""

    try:
        result = qa_chain({"query": prompt})
        full_text = result["result"].strip()
        if not full_text:
            full_text = "(生成失敗，請改用其他模型或調整段落數)"
    except Exception as e:
        return f"(生成失敗：{str(e)})", None

    request_count += 1

    paragraphs = [p.strip() for p in full_text.split("\n") if p.strip()]

    docx_file = "generated_article.docx"
    doc = DocxDocument()
    doc.add_heading(query, level=1)
    for p in paragraphs:
        doc.add_paragraph(p)
    doc.save(docx_file)

    return "\n\n".join(paragraphs), docx_file

# -------------------------------
# 7. Gradio 介面
# -------------------------------
iface = gr.Interface(
    fn=rag_generate_hfapi,
    inputs=[
        gr.Textbox(lines=2, placeholder="請輸入文章主題"),
        gr.Dropdown(list(MODEL_DICT.keys()), value="google/flan-t5-large", label="選擇模型"),
        gr.Slider(minimum=1, maximum=10, value=5, step=1, label="段落數"),
        gr.Slider(minimum=500, maximum=3000, value=1500, step=100, label="文章字數上限")
    ],
    outputs=[
        gr.Textbox(label="生成文章"),
        gr.File(label="下載 DOCX")
    ],
    title="佛教經論 RAG 系統 (Hugging Face API)",
    description="使用 Hugging Face API 生成文章，可選大模型，分段生成並下載 DOCX，每小時生成次數有限制"
)

iface.launch()