app.py

import os
import time
import shutil
from pydub import AudioSegment
from openai import OpenAI
import gradio as gr
from fastapi import FastAPI, File, UploadFile, Form, HTTPException

# ======================================================
# 🔐 設定區
# ======================================================
PASSWORD = os.getenv("APP_PASSWORD", "defaultpass")
MAX_SIZE = 25 * 1024 * 1024
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

print("===== 🚀 啟動中 =====")
print(f"APP_PASSWORD: {'✅ 已載入' if PASSWORD else '❌ 未載入'}")
print(f"目前密碼內容：{PASSWORD}")

# ======================================================
# ⚔️ 防暴力破解
# ======================================================
MAX_FAILED_IN_WINDOW = 10
WINDOW_SECONDS = 24 * 3600
LOCK_DURATION_SECONDS = 24 * 3600
SHORT_BURST_LIMIT = 5
SHORT_BURST_SECONDS = 60

attempts = {}
locked = {}

def _now():
    return int(time.time())

def prune_old_attempts(sid):
    cutoff = _now() - WINDOW_SECONDS
    if sid in attempts:
        attempts[sid] = [t for t in attempts[sid] if t >= cutoff]
        if not attempts[sid]:
            del attempts[sid]

def check_lock(sid):
    if sid in locked:
        if _now() < locked[sid]:
            remain = locked[sid] - _now()
            return True, f"🔒 已被鎖定，請 {remain // 60} 分鐘後再試。"
        else:
            locked.pop(sid, None)
            attempts.pop(sid, None)
    prune_old_attempts(sid)
    cnt = len(attempts.get(sid, []))
    if cnt >= MAX_FAILED_IN_WINDOW:
        locked[sid] = _now() + LOCK_DURATION_SECONDS
        return True, f"🔒 嘗試過多，已鎖定 24 小時。"
    return False, ""

def record_failed_attempt(sid):
    now = _now()
    attempts.setdefault(sid, []).append(now)
    prune_old_attempts(sid)
    recent_cutoff = now - SHORT_BURST_SECONDS
    recent = [t for t in attempts[sid] if t >= recent_cutoff]
    if len(recent) >= SHORT_BURST_LIMIT:
        locked[sid] = now + 300
        return len(attempts[sid]), "⚠️ 多次快速嘗試，暫時鎖定5分鐘。"
    return len(attempts[sid]), ""

def clear_attempts(sid):
    attempts.pop(sid, None)
    locked.pop(sid, None)

# ======================================================
# 🎧 音訊轉錄核心
# ======================================================
def split_audio_if_needed(path):
    size = os.path.getsize(path)
    if size <= MAX_SIZE:
        return [path]
    audio = AudioSegment.from_file(path)
    num = int(size / MAX_SIZE) + 1
    chunk_ms = len(audio) / num
    files = []
    for i in range(num):
        start, end = int(i * chunk_ms), int((i + 1) * chunk_ms)
        chunk = audio[start:end]
        fn = f"chunk_{i+1}.wav"
        chunk.export(fn, format="wav")
        files.append(fn)
    return files

def transcribe_core(path, model):
    # 1) iPhone LINE .mp4 → 假副檔名修正
    if path and path.lower().endswith(".mp4"):
        fixed_path = path[:-4] + ".m4a"
        try:
            shutil.copy(path, fixed_path)
            path = fixed_path
            print("🔧 已自動修正 mp4 → m4a")
        except Exception as e:
            print(f"⚠️ mp4→m4a 複製失敗：{e}，改用原檔嘗試")

    # 2) Whisper 逐段轉錄（原始：可能含簡體）
    chunks = split_audio_if_needed(path)
    raw_parts = []
    for f in chunks:
        with open(f, "rb") as af:
            res = client.audio.transcriptions.create(
                model=model,                # "whisper-1" 或 "gpt-4o-mini-transcribe"
                file=af,
                response_format="text"
            )
            raw_parts.append(res)
    full_raw = "\n".join(raw_parts)  # 原始逐字稿（可能有簡體）

    # 3) 將全文「只做字形轉換」成繁體（不意譯、不刪減）
    conv_prompt = (
        "請將以下內容完整轉換為「繁體中文（台灣用語）」：\n"
        "規則：1) 僅做簡→繁字形轉換；2) 不要意譯或改寫；3) 不要添加任何前後綴。\n"
        "-----\n" + full_raw
    )
    trad_resp = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": "你是嚴格的繁體中文轉換器，只進行字形轉換。"},
            {"role": "user", "content": conv_prompt}
        ],
        temperature=0.0,
    )
    full_trad = trad_resp.choices[0].message.content.strip()  # ✅ 確保為繁體

    # 4) 以繁體全文產生摘要（也強制繁體）
    sum_prompt = (
        "請用台灣繁體中文 (避免使用簡體字), 條列式呈現精簡重點 (如果需要的話)。\n\n"
        + full_trad
    )
    sum_resp = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": "你是一位精準且嚴格使用繁體中文（台灣）的摘要助手。"},
            {"role": "user", "content": sum_prompt}
        ],
        temperature=0.2,
    )
    summ = sum_resp.choices[0].message.content.strip()

    # 依原介面回傳兩個值：完整逐字稿（已繁體）、摘要
    return full_trad, summ


# ======================================================
# 🌐 FastAPI API（iPhone 捷徑可用）
# ======================================================
app = FastAPI(title="LINE Transcription API")

@app.post("/api/transcribe")
async def api_transcribe(
    file: UploadFile = File(...),
    token: str = Form(default=None)
):
    """供捷徑上傳音訊並取得 JSON"""
    if token != PASSWORD:
        raise HTTPException(status_code=403, detail="Invalid token")

    temp = file.filename
    with open(temp, "wb") as f:
        f.write(await file.read())

    text, summary = transcribe_core(temp, "whisper-1")
    os.remove(temp)
    return {"text": text, "summary": summary}

@app.get("/health")
def health():
    """捷徑可 ping 這個確認服務運作中"""
    return {"status": "ok", "time": int(time.time())}

# ======================================================
# 💬 Gradio 介面（完整舊版）
# ======================================================
def _normalize_upload_path(file_input):
    if not file_input:
        return None
    if isinstance(file_input, str):
        return file_input
    if isinstance(file_input, list) and file_input:
        return _normalize_upload_path(file_input[0])
    path = getattr(file_input, "name", None)
    if not path and isinstance(file_input, dict):
        path = file_input.get("name") or file_input.get("path")
    return path

def transcribe_with_password(session_id, password, file_input, model_choice):
    password = password.strip().replace(" ", "").replace("\u200b", "")
    locked_flag, msg = check_lock(session_id)
    if locked_flag:
        return msg, "", ""
    if password != PASSWORD:
        cnt, msg2 = record_failed_attempt(session_id)
        return msg2 or f"密碼錯誤（第 {cnt} 次）", "", ""
    path = _normalize_upload_path(file_input)
    if not path or not os.path.exists(path):
        return "找不到上傳檔案，請重新選擇。", "", ""
    clear_attempts(session_id)
    full, summ = transcribe_core(path, model_choice)
    return "✅ 轉錄完成", full, summ

def ask_about_transcript(full_text, q):
    if not full_text.strip():
        return "⚠️ 尚未有轉錄內容"
    if not q.strip():
        return "請輸入問題"
    prompt = f"以下是轉錄內容：\n{full_text}\n\n問題：{q}\n請用繁體中文回答。"
    res = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role":"user","content":prompt}],
        temperature=0.6,
    )
    return res.choices[0].message.content.strip()

# ======================================================
# 🖥️ Gradio UI
# ======================================================
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("## 🎧 語音轉錄與摘要工具（私人API勿轉傳｜支援 iPhone LINE .mp4）")
    session_state = gr.State(value=None)
    with gr.Row():
        password_input = gr.Textbox(label="輸入密碼", placeholder="請輸入英文與數字（請切換成英文輸入法）", type="password", max_lines=1)
        model_choice = gr.Dropdown(["whisper-1", "gpt-4o-mini-transcribe"], value="whisper-1", label="選擇模型")
    file_input = gr.File(label="上傳音訊 / LINE 語音檔（支援 .m4a, .aac, .wav, .mp4）", file_count="single", file_types=["audio", ".mp4", ".m4a", ".aac", ".wav"])
    transcribe_btn = gr.Button("開始轉錄與摘要 🚀")
    status_box = gr.Textbox(label="狀態", interactive=False)
    transcript_box = gr.Textbox(label="完整轉錄文字", lines=10)
    copy_transcript = gr.Button("📋 複製轉錄文字")
    summary_box = gr.Textbox(label="摘要結果", lines=10)
    copy_summary = gr.Button("📋 複製摘要結果")

    with gr.Accordion("💬 進一步問 AI", open=False):
        user_q = gr.Textbox(label="輸入問題", lines=2)
        ask_btn = gr.Button("詢問 AI 🤔")
        ai_reply = gr.Textbox(label="AI 回覆", lines=6)
        copy_reply = gr.Button("📋 複製 AI 回覆")

    def init_session():
        import uuid
        return str(uuid.uuid4())

    demo.load(init_session, None, session_state)
    transcribe_btn.click(transcribe_with_password, [session_state, password_input, file_input, model_choice], [status_box, transcript_box, summary_box])
    ask_btn.click(ask_about_transcript, [transcript_box, user_q], [ai_reply])

    copy_js = """async (text) => {try {await navigator.clipboard.writeText(text); alert("✅ 已複製到剪貼簿！");} catch (e) {alert("❌ 複製失敗：" + e);}}"""
    copy_transcript.click(fn=None, inputs=transcript_box, outputs=None, js=copy_js)
    copy_summary.click(fn=None, inputs=summary_box, outputs=None, js=copy_js)
    copy_reply.click(fn=None, inputs=ai_reply, outputs=None, js=copy_js)

# ======================================================
# 🚀 啟動（Hugging Face）
# ======================================================
gr.mount_gradio_app(app, demo, path="/")

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)