app.py

# app.py
import os, torch
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, AutoModelForCausalLM, pipeline
from huggingface_hub import login, snapshot_download
import gradio as gr

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

DB_PATH = "./faiss_db"
if os.path.exists(DB_PATH):
    print("✅ 載入現有向量資料庫...")
    db = FAISS.load_local(DB_PATH, embeddings_model, allow_dangerous_deserialization=True)
else:
    raise ValueError("❌ 沒找到 faiss_db，請先建立向量資料庫")

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

# -------------------------------
# 1. 模型設定（中文 GPT2 + fallback）
# -------------------------------
PRIMARY_MODEL = "uer/gpt2-chinese-cluecorpusmedium"
FALLBACK_MODEL = "uer/gpt2-chinese-cluecorpussmall"

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")

def try_download_model(repo_id):
    local_dir = f"./models/{repo_id.split('/')[-1]}"
    if not os.path.exists(local_dir):
        print(f"⬇️ 嘗試下載模型 {repo_id} ...")
        try:
            snapshot_download(repo_id=repo_id, token=HF_TOKEN, local_dir=local_dir)
        except Exception as e:
            print(f"⚠️ 模型 {repo_id} 無法下載: {e}")
            return None
    return local_dir

LOCAL_MODEL_DIR = try_download_model(PRIMARY_MODEL)
if LOCAL_MODEL_DIR is None:
    print("⚠️ 切換到 fallback 模型：小型 GPT2-Chinese")
    LOCAL_MODEL_DIR = try_download_model(FALLBACK_MODEL)
    MODEL_NAME = FALLBACK_MODEL
else:
    MODEL_NAME = PRIMARY_MODEL

print(f"👉 最終使用模型：{MODEL_NAME}")

# -------------------------------
# 2. pipeline 載入
# -------------------------------
tokenizer = AutoTokenizer.from_pretrained(LOCAL_MODEL_DIR)
model = AutoModelForCausalLM.from_pretrained(LOCAL_MODEL_DIR)

# 修正 pad_token 缺失問題
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

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

def call_local_inference(prompt, max_new_tokens=256):
    try:
        if "中文" not in prompt:
            prompt += "\n（請用中文回答）"
        outputs = generator(
            prompt,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=0.7,
            pad_token_id=tokenizer.pad_token_id
        )
        return outputs[0]["generated_text"]
    except Exception as e:
        return f"（生成失敗：{e}）"

# -------------------------------
# 3. 文章生成（RAG + 即時寫入DOCX + 檢索片段 + 進度提示）
# -------------------------------
def generate_article_progress(query, segments=5):
    docx_file = "/tmp/generated_article.docx"
    doc = DocxDocument()
    doc.add_heading(query, level=1)
    doc.save(docx_file)  # 先建立空的檔案

    all_text = []

    # 🔍 RAG 檢索佛經段落
    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):
        progress_text = f"⏳ 正在生成第 {i+1}/{segments} 段..."
        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)

        # ✅ 每段生成後即時寫入 DOCX
        doc = DocxDocument(docx_file)
        doc.add_paragraph(f"第{i+1}段：\n{paragraph}")
        doc.save(docx_file)

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

    final_progress = f"✅ 已完成全部 {segments} 段生成！"
    yield "\n\n".join(all_text), docx_file, f"本次使用模型：{MODEL_NAME}", context, final_progress

# -------------------------------
# 4. 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 GPT2-Chinese + FAISS RAG，生成佛教主題文章。")

    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")
    model_used_text = gr.Textbox(label="實際使用模型", interactive=False)
    context_text = gr.Textbox(label="檢索到的佛經片段", interactive=False, lines=6)
    progress_text = gr.Textbox(label="生成進度", interactive=False)

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

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