openai-telegram-bot/main.py

import asyncio
import logging
import os
import wave
from functools import wraps
from io import BytesIO
from typing import Tuple

import openai
from aiogram import Bot, Dispatcher, types
from aiogram.contrib.middlewares.logging import LoggingMiddleware
from aiogram.types import InlineKeyboardButton, InlineKeyboardMarkup, ParseMode
from aiogram.utils import executor
from dotenv import load_dotenv
from langdetect import detect
from pydub import AudioSegment

import database

logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(message)s", level=logging.INFO
)
logger = logging.getLogger(__name__)

# Envrionment Variables Load
load_dotenv()
if os.environ.get("OPENAI_API_KEY") is None:
    print(
        "OpenAI_API_KEY is not set in.env file or OPENAI_API_KEY environment variable is not set"
    )
    exit(1)

BOT_TOKEN = os.getenv("BOT_TOKEN")

bot = Bot(token=BOT_TOKEN)
dp = Dispatcher(bot)
dp.middleware.setup(LoggingMiddleware())

ALLOWED_USERS = os.environ.get("BOT_ALLOWED_USERS").split(",")
SYSTEM_PROMPT = os.environ.get("CHATGPT_SYSTEM_PROMPT")
TEMPERATURE = os.environ.get("CHATGPT_TEMPERATURE")
MODEL = os.environ.get("OPENAI_MODEL")
WHISPER_TO_GPT = bool(int(os.environ.get("WHISPER_TO_GPT")))

# TTS Settings
ENABLE_TTS = bool(int(os.environ.get("ENABLE_TTS")))
DEFAULT_VOICE_LANGUAGE = os.environ.get("DEFAULT_VOICE_LANGUAGE")
VOICE_LANGUAGE_LIST = os.environ.get("VOICE_LANGUAGE_LIST")
LANGUAGES_22050 = os.environ.get("LANGUAGES_H_SR").split(",")

MAX_USER_CONTEXT = int(os.environ.get("CHATGPT_MAX_USER_CONTEXT"))
openai.api_key = os.environ.get("OPENAI_API_KEY")


async def getUserData(chat_id):
    user_data = database.get_user(chat_id)
    if not user_data:
        user_data = {
            "context": [],
            "usage": {"chatgpt": 0, "whisper": 0, "dalle": 0},
            "options": {
                "whisper_to_chat": WHISPER_TO_GPT,
                "assistant_voice_chat": False,
                "temperature": float(TEMPERATURE),
                "max-context": MAX_USER_CONTEXT,
            },
        }
        database.add_user(chat_id, user_data)
        user_data = database.get_user(chat_id)
    return user_data


def generate_settings_markup(chat_id: str) -> InlineKeyboardMarkup:
    keyboard = [
        [
            InlineKeyboardButton(
                "Increase Temperature", callback_data=f"setting_inc_temp_{chat_id}"
            ),
            InlineKeyboardButton(
                "Decrease Temperature", callback_data=f"setting_dec_temp_{chat_id}"
            ),
        ],
        [
            InlineKeyboardButton(
                "Enable Whisper", callback_data=f"setting_en_whisper_{chat_id}"
            ),
            InlineKeyboardButton(
                "Disable Whisper", callback_data=f"setting_dis_whisper_{chat_id}"
            ),
        ],
        [
            InlineKeyboardButton(
                "Enable assistant voice", callback_data=f"setting_en_voice_{chat_id}"
            ),
            InlineKeyboardButton(
                "Disable assistant voice", callback_data=f"setting_dis_voice_{chat_id}"
            ),
        ],
        [
            InlineKeyboardButton(
                "Increase Context", callback_data=f"setting_inc_context_{chat_id}"
            ),
            InlineKeyboardButton(
                "Decrease Context", callback_data=f"setting_dec_context_{chat_id}"
            ),
        ],
    ]
    return InlineKeyboardMarkup(inline_keyboard=keyboard)


async def send_voice_message(chat_id, assistant_message):
    await bot.send_chat_action(chat_id, action=types.ChatActions.TYPING)
    audio_data, _ = await text_to_voice(assistant_message)
    audio_data.seek(0)  # Reset the buffer's position to the beginning

    await bot.send_voice(chat_id, audio_data)


async def text_to_voice(text: str, language: str = None) -> Tuple[BytesIO, str]:
    binary_path = "./piper/piper"

    if language is None:
        language = detect(text[0:100])

    model_path = f"./piper/voices/{language}.onnx"

    text = text.replace("\n", ". ")

    cmd = [binary_path, "--model", model_path, "--output_raw"]
    proc = await asyncio.create_subprocess_exec(
        *cmd, stdin=asyncio.subprocess.PIPE, stdout=asyncio.subprocess.PIPE
    )

    stdout, _ = await proc.communicate(input=text.encode("utf-8"))

    # Create a new BytesIO object to store the WAV file.
    wav_file = BytesIO()
    with wave.open(wav_file, "wb") as wf:
        # Assuming 1 channel, 16 bits per sample, and 22050 samples per second.
        wf.setnchannels(1)
        wf.setsampwidth(2)
        wf.setframerate(22050 if language in LANGUAGES_22050 else 16000)
        wf.writeframes(stdout)

    wav_file.seek(0)

    audio = AudioSegment.from_file(wav_file, format="wav")
    audio = audio.export(format="ogg", codec="libopus", parameters=["-vbr", "on"])

    return BytesIO(audio.read()), "ogg"


def restricted(func):
    @wraps(func)
    async def wrapped(message, *args, **kwargs):
        user_id = str(message.chat.id)
        if user_id not in ALLOWED_USERS:
            if "*" != ALLOWED_USERS[0]:
                print(f"Unauthorized access denied for {user_id}.")
                return
        else:
            _ = await getUserData(user_id)
        return await func(message, *args, **kwargs)

    return wrapped


async def messageGPT(text: str, chat_id: str, user_name="User", user_data={}):
    user_data["context"].append({"role": "user", "content": text})
    if len(user_data["context"]) > user_data["options"]["max-context"]:
        user_data["context"].pop(0)

    try:
        await bot.send_chat_action(chat_id, action=types.ChatActions.TYPING)
        
        response = openai.ChatCompletion.create(
            model=MODEL,
            messages=[
                {
                    "role": "system",
                    "content": f"You are chatting with {user_name}. {SYSTEM_PROMPT}",
                }
            ]
            + user_data["context"],
            temperature=user_data["options"]["temperature"],
        )
        
        user_data["usage"]["chatgpt"] += int(
            response.get("usage", {"total_tokens": 0})["total_tokens"]
        )
    except Exception as e:
        print(e)
        return f"There was a problem with OpenAI, so I can't answer you: \n\n{e}"

    assistant_message = (
        response.get("choices", [{}])[0]
        .get("message", {"content": None})
        .get(
            "content",
            "There was a problem with OpenAI. Maybe your prompt is forbidden? They like to censor a lot!",
        )
    )

    user_data["context"].append({"role": "assistant", "content": assistant_message})
    if len(user_data["context"]) > user_data["options"]["max-context"]:
        user_data["context"].pop(0)

    database.update_user(chat_id, user_data)
    return assistant_message, user_data


@dp.message_handler(commands=["start"])
@restricted
async def start(message: types.Message):
    _ = await getUserData(message.chat.id)
    await message.reply("Hello, how can I assist you today?")


@dp.message_handler(commands=["clear"], content_types=["text"])
@restricted
async def clear(message: types.Message) -> None:
    chat_id = str(message.chat.id)
    user_data = await getUserData(chat_id)
    if user_data:
        user_data["context"] = []
        database.update_user(chat_id, user_data)
        print(f"Cleared context for {message.from_user.full_name}")
    await message.reply("Your message context history was cleared.")


@dp.message_handler(commands=["usage"])
@restricted
async def usage(message: types.Message) -> None:
    chat_id = str(message.chat.id)
    user_data = database.get_user(chat_id)
    user_usage = user_data["usage"]
    total_usage = database.get_total_usage()

    user_spent = round(
        (
            ((user_usage["chatgpt"] / 750) * 0.002)
            + (float(user_usage["dalle"]) * 0.02)
            + ((user_usage["whisper"] / 60.0) * 0.006)
        ),
        4,
    )
    total_spent = round(
        (
            ((total_usage["chatgpt"] / 750) * 0.002)
            + (float(total_usage["dalle"]) * 0.02)
            + ((total_usage["whisper"] / 60.0) * 0.006)
        ),
        4,
    )

    user_percentage = (user_spent / total_spent) * 100 if total_spent > 0 else 0

    info_message = f"""User: {message.from_user.full_name}
- Used ~{user_usage["chatgpt"]} tokens with ChatGPT.
- Generated {user_usage["dalle"]} images with DALL-E.
- Transcribed {round(float(user_usage["whisper"]) / 60.0, 2)}min with Whisper.

User total: ${user_spent} ({user_percentage:.2f}% of total)

Total usage:
- ChatGPT tokens: {total_usage["chatgpt"]}
- DALL-E images: {total_usage["dalle"]}
- Whisper transcription: {round(float(total_usage["whisper"]) / 60.0, 2)}min

Total spent: ${total_spent}"""

    await message.reply(info_message)


@dp.message_handler(
    lambda message: message.chat.type == types.ChatType.PRIVATE,
    content_types=["text"],
    regexp="^/imagine",
)
@restricted
async def imagine(message: types.Message):
    await bot.send_chat_action(message.chat.id, action=types.ChatActions.TYPING)
    user_data = await getUserData(message.chat.id)
    user_data["usage"]["dalle"] += 1
    database.update_user(message.chat.id, user_data)

    response = openai.Image.create(prompt=message.text, n=1, size="1024x1024")
    try:
        image_url = response["data"][0]["url"]
        await message.reply(image_url)
    except Exception as e:
        print(e)
        await message.reply(
            "Error generating. Your prompt may contain text that is not allowed by OpenAI safety system."
        )


@dp.message_handler(content_types=["photo", "video", "audio", "voice"])
@restricted
async def attachment(message: types.Message):
    chat_id = message.chat.id
    user_data = await getUserData(chat_id)
    await bot.send_chat_action(chat_id, action=types.ChatActions.TYPING)

    transcript = {"text": ""}
    audioMessage = False

    # Handle media types
    if message.voice:
        user_data["usage"]["whisper"] += message.voice.duration
        file_id = message.voice.file_id
        file_format = "ogg"
        audioMessage = True
    elif message.video:
        user_data["usage"]["whisper"] += message.video.duration
        file_id = message.video.file_id
        file_format = "mp4"
    elif message.audio:
        user_data["usage"]["whisper"] += message.audio.duration
        file_id = message.audio.file_id
        file_format = "mp3"
    else:
        await message.reply("Can't handle such file. Reason: unknown.")
        return

    # Download file from Telegram
    file = await bot.get_file(file_id)
    user_id = message.chat.id
    await file.download(f"{user_id}.{file_format}")

    # Convert audio to mp3 if needed, because OpenAI doesn't support ogg
    if file_format == "ogg":
        ogg_audio = AudioSegment.from_file(f"{user_id}.ogg", format="ogg")
        ogg_audio.export(f"{user_id}.mp3", format="mp3")
        os.remove(f"{user_id}.ogg")
        file_format = "mp3"

    # Transcribe audio with OpenAI API
    with open(f"{user_id}.{file_format}", "rb") as audio_file:
        try:
            await bot.send_chat_action(chat_id, action=types.ChatActions.TYPING)
            transcript = openai.Audio.transcribe("whisper-1", audio_file)
        except Exception as e:
            print(e)
            await message.reply("Transcript failed.")
            os.remove(f"{user_id}.{file_format}")
            return
        os.remove(f"{user_id}.{file_format}")

    # Handle empty transcript
    if transcript["text"] == "":
        transcript["text"] = "[Silence]"

    # If whisper_to_chat is enabled, send transcript to ChatGPT and send its response along with the transcript
    chatGPT_response = False
    if audioMessage and user_data["options"]["whisper_to_chat"]:
        chatGPT_response, user_data = await messageGPT(
            transcript["text"], str(chat_id), message.from_user.full_name, user_data
        )
        transcript["text"] = ''.join(["> ", transcript["text"], "\n\n", chatGPT_response])

    # Send transcript (and ChatGPT response if enabled)
    await message.reply(transcript["text"])
    
    # Send ChatGPT response as voice message with piper TTS if enabled
    if user_data["options"]["assistant_voice_chat"] and chatGPT_response:
        asyncio.create_task(send_voice_message(chat_id, chatGPT_response))
        
    # Update user data
    database.update_user(str(chat_id), user_data)


@restricted
@dp.message_handler(commands=["settings"])
async def settings(message: types.Message):
    chat_id = str(message.chat.id)
    settings_markup = generate_settings_markup(chat_id)
    await message.reply(text="Settings:", reply_markup=settings_markup)


async def settings_callback(callback_query: types.CallbackQuery):
    user_data = await getUserData(callback_query.message.chat.id)
    action, chat_id = callback_query.data.rsplit("_", 1)
    options = user_data["options"]

    if action.startswith("setting_inc_temp"):
        options["temperature"] = min(options["temperature"] + 0.1, 1)
    elif action.startswith("setting_dec_temp"):
        options["temperature"] = max(options["temperature"] - 0.1, 0)

    elif action.startswith("setting_en_whisper"):
        options["whisper_to_chat"] = True
    elif action.startswith("setting_dis_whisper"):
        options["whisper_to_chat"] = False

    elif action.startswith("setting_en_voice"):
        options["assistant_voice_chat"] = True
    elif action.startswith("setting_dis_voice"):
        options["assistant_voice_chat"] = False

    elif action.startswith("setting_inc_context"):
        options["max-context"] = min(options["max-context"] + 1, MAX_USER_CONTEXT)
    elif action.startswith("setting_dec_context"):
        options["max-context"] = max(options["max-context"] - 1, 1)

    settings_markup = generate_settings_markup(chat_id)
    await callback_query.message.edit_text(
        text="Choose a setting option:", reply_markup=settings_markup
    )

    database.update_user(chat_id, user_data)
    settings_txt = f"Updated settings:\n\nTemperature: {options['temperature']}\nWhisper to Chat: {options['whisper_to_chat']}\nAssistant voice: {options['assistant_voice_chat']}\nContext Length: {options['max-context']}"
    await callback_query.answer()
    await callback_query.message.reply(text=settings_txt)


@dp.message_handler(
    lambda message: message.chat.type == types.ChatType.PRIVATE
    and not message.text.startswith("/"),
    content_types=["text"],
)
async def chat(message: types.Message):
    chat_id = str(message.chat.id)
    user_data = await getUserData(chat_id)

    user_prompt = message.text
    assistant_message, user_data = await messageGPT(
        user_prompt, chat_id, message.from_user.full_name, user_data
    )

    await message.reply(assistant_message, parse_mode=ParseMode.MARKDOWN)

    if user_data["options"]["assistant_voice_chat"]:
        asyncio.ensure_future(send_voice_message(chat_id, assistant_message))


if __name__ == "__main__":
    database.init_database()

    try:
        ALLOWED_USERS = os.environ.get("BOT_ALLOWED_USERS").split(",")
    except Exception as e:
        print(e)
        ALLOWED_USERS = ALLOWED_USERS

    print(f"Allowed users: {ALLOWED_USERS}")
    print(f"System prompt: {SYSTEM_PROMPT}")
    print(f"TTS: {ENABLE_TTS}")

    # Register message handler and callback query handler for settings
    dp.register_message_handler(settings, commands=["settings"])
    dp.register_callback_query_handler(
        settings_callback, lambda c: c.data.startswith("setting_")
    )

    executor.start_polling(dp, skip_updates=True, fast=True)