app.py

# app.py
# -------------------------------
# 1. 套件載入
# -------------------------------
import os, glob, requests
from langchain.docstore.document import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.chains import RetrievalQA
from langchain_huggingface import HuggingFaceEmbeddings, HuggingFaceEndpoint
from docx import Document as DocxDocument
import gradio as gr
from langchain_community.vectorstores import FAISS

# -------------------------------
# 2. 環境變數與資料路徑
# -------------------------------
TXT_FOLDER = "./out_texts"
DB_PATH = "./faiss_db"
os.makedirs(DB_PATH, exist_ok=True)
os.makedirs(TXT_FOLDER, exist_ok=True)  # 避免沒有 txt 檔時錯誤

HF_TOKEN = os.environ.get("HUGGINGFACEHUB_API_TOKEN")
if not HF_TOKEN:
    raise ValueError(
        "請在 Hugging Face Space 的 Settings → Repository secrets 設定 HUGGINGFACEHUB_API_TOKEN"
    )

# -------------------------------
# 3. 建立或載入向量資料庫
# -------------------------------
EMBEDDINGS_MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
embeddings_model = HuggingFaceEmbeddings(model_name=EMBEDDINGS_MODEL_NAME)

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)}))
    splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
    split_docs = splitter.split_documents(docs)
    db = FAISS.from_documents(split_docs, embeddings_model)
    db.save_local(DB_PATH)

retriever = db.as_retriever(search_type="similarity", search_kwargs={"k": 5})

# -------------------------------
# 4. LLM 設定（Hugging Face Endpoint）
# -------------------------------
llm = HuggingFaceEndpoint(
    repo_id="google/flan-t5-large",
    task="text2text-generation",
    huggingfacehub_api_token=HF_TOKEN,
    model_kwargs={"temperature": 0.7, "max_new_tokens": 512},
)

qa_chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=retriever,
    return_source_documents=True
)

# -------------------------------
# 5. 查詢 API 剩餘額度
# -------------------------------
def get_hf_rate_limit():
    headers = {"Authorization": f"Bearer {HF_TOKEN}"}
    try:
        r = requests.get("https://huggingface.co/api/whoami", headers=headers)
        r.raise_for_status()
        data = r.json()
        remaining = data.get("rate_limit", {}).get("remaining", "未知")
        return f"本小時剩餘 API 次數：約 {remaining}"
    except Exception:
        return "無法取得 API 速率資訊"

# -------------------------------
# 6. 生成文章
# -------------------------------
def generate_article_with_rate(query, segments=5):
    docx_file = "/tmp/generated_article.docx"
    doc = DocxDocument()
    doc.add_heading(query, level=1)

    all_text = []
    prompt = f"請依據下列主題生成段落：{query}\n\n每段約150-200字。"

    for i in range(int(segments)):
        try:
            result = qa_chain({"query": prompt})
            paragraph = result.get("result", "").strip()
            if not paragraph:
                paragraph = "（本段生成失敗，請嘗試減少段落或改用較小模型。）"
        except Exception as e:
            paragraph = f"（本段生成失敗：{e}）"
        all_text.append(paragraph)
        doc.add_paragraph(paragraph)
        prompt = f"請接續上一段生成下一段：\n{paragraph}\n\n下一段："

    doc.save(docx_file)
    full_text = "\n\n".join(all_text)
    rate_info = get_hf_rate_limit()
    return f"{rate_info}\n\n{full_text}", docx_file

# -------------------------------
# 7. Gradio 介面
# -------------------------------
with gr.Blocks() as demo:
    gr.Markdown("# 佛教經論 RAG 系統 (HF API)")
    gr.Markdown("使用 Hugging Face Endpoint LLM + FAISS RAG，生成文章並提示 API 剩餘額度。")
    
    query_input = gr.Textbox(lines=2, placeholder="請輸入文章主題", label="文章主題")
    segments_input = gr.Slider(minimum=1, maximum=10, step=1, value=5, label="段落數")
    output_text = gr.Textbox(label="生成文章 + API 剩餘次數")
    output_file = gr.File(label="下載 DOCX")
    
    query_input.submit(generate_article_with_rate, [query_input, segments_input], [output_text, output_file])
    segments_input.change(generate_article_with_rate, [query_input, segments_input], [output_text, output_file])

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