app.py

# app.py
import os
from langchain.docstore.document import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings
from docx import Document as DocxDocument
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
from huggingface_hub import login, snapshot_download
import gradio as gr

# -------------------------------
# 1. 模型設定（專門中文 T5 Pegasus）
# -------------------------------
MODEL_NAME = "imxly/t5-pegasus-small"

HF_TOKEN = os.environ.get("HUGGINGFACEHUB_API_TOKEN")
if HF_TOKEN:
    login(token=HF_TOKEN)
    print("✅ 已使用 HUGGINGFACEHUB_API_TOKEN 登入 Hugging Face")

LOCAL_MODEL_DIR = f"./models/{MODEL_NAME.split('/')[-1]}"
if not os.path.exists(LOCAL_MODEL_DIR):
    print(f"⬇️ 嘗試下載模型 {MODEL_NAME} ...")
    snapshot_download(repo_id=MODEL_NAME, token=HF_TOKEN, local_dir=LOCAL_MODEL_DIR)

print(f"👉 最終使用模型：{MODEL_NAME}")

# -------------------------------
# 2. pipeline 載入
# -------------------------------
tokenizer = AutoTokenizer.from_pretrained(
    LOCAL_MODEL_DIR,
    use_fast=False  # ✅ 避免 tiktoken 錯誤
)
model = AutoModelForSeq2SeqLM.from_pretrained(LOCAL_MODEL_DIR)

generator = pipeline(
    "text2text-generation",
    model=model,
    tokenizer=tokenizer,
    device=-1  # CPU
)

def call_local_inference(prompt, max_new_tokens=256):
    try:
        outputs = generator(
            prompt,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=0.7
        )
        return outputs[0]["generated_text"]
    except Exception as e:
        return f"（生成失敗：{e}）"

# -------------------------------
# 3. 建立或載入向量資料庫
# -------------------------------
TXT_FOLDER = "./out_texts"
DB_PATH = "./faiss_db"
os.makedirs(DB_PATH, exist_ok=True)
os.makedirs(TXT_FOLDER, exist_ok=True)

EMBEDDINGS_MODEL_NAME = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-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，建立新向量資料庫...")
    docs = []
    txt_files = [f for f in os.listdir(TXT_FOLDER) if f.endswith(".txt")]
    for filename in txt_files:
        with open(os.path.join(TXT_FOLDER, filename), "r", encoding="utf-8") as f:
            docs.append(Document(page_content=f.read(), metadata={"source": filename}))
    splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
    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. 文章生成（結合 RAG）
# -------------------------------
def generate_article_progress(query, segments=5):
    docx_file = "/tmp/generated_article.docx"
    doc = DocxDocument()
    doc.add_heading(query, level=1)

    all_text = []

    # 🔍 使用 RAG 檢索
    retrieved_docs = retriever.get_relevant_documents(query)
    context_texts = [d.page_content for d in retrieved_docs]
    context = "\n".join([f"{i+1}. {txt}" for i, txt in enumerate(context_texts[:3])])

    for i in range(segments):
        prompt = (
            f"以下是佛教經論的相關段落：\n{context}\n\n"
            f"請依據上面內容，寫一段約150-200字的中文文章，"
            f"主題：{query}。\n第{i+1}段："
        )

        paragraph = call_local_inference(prompt)
        all_text.append(paragraph)
        doc.add_paragraph(paragraph)

        yield "\n\n".join(all_text), None, f"本次使用模型：{MODEL_NAME}"

    doc.save(docx_file)
    yield "\n\n".join(all_text), docx_file, f"本次使用模型：{MODEL_NAME}"

# -------------------------------
# 5. Gradio 介面
# -------------------------------
with gr.Blocks() as demo:
    gr.Markdown("# 📺 電視弘法視頻生成文章 RAG 系統")
    gr.Markdown("使用 FAISS + 中文 T5 模型，根據資料庫生成中文文章。")

    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")
    output_file = gr.File(label="下載 DOCX")
    output_model = gr.Textbox(label="使用的模型")

    btn = gr.Button("生成文章")
    btn.click(
        generate_article_progress,
        inputs=[query_input, segments_input],
        outputs=[output_text, output_file, output_model]
    )

# -------------------------------
# 6. 啟動
# -------------------------------
if __name__ == "__main__":
    demo.launch()