Hotel AI Concierge

The reception desk answers the same questions all day: breakfast hours, the Wi-Fi password, when the pool opens, how to book a transfer. In a resort hotel in Turkey guests ask them in Russian, English, Turkish and German, and the staff do not speak all four.

A plain chatbot is risky here. An AI that invents a check-out time or a price causes real trouble at the desk, and the guest blames the hotel, not the bot.

A Telegram bot that answers only from facts the hotel wrote down, and asks a person about everything else:

• The knowledge base is one Markdown file the hotel edits. /reload applies changes without a restart; a broken file keeps the old version.
• The language model (Claude or OpenAI, chosen in the settings) returns structured JSON with a found_in_knowledge_base flag, so "do I know this?" is a yes/no value, not a guess from the model's wording.
• Unknown questions go to a staff Telegram group with an English translation. Staff simply press Reply; the bot translates the answer into the guest's language and shows the original below.
• The guest's language is detected without AI: alphabet, letters such as ı ş ğ or ä ß, and common words. Short messages like "ok" keep the language of the conversation.
• Every failure ends with a person: API down, model refusal or a cut-off answer go to staff; if the staff group cannot be reached, the guest is told to ask at the desk.

• Python 3.12
• aiogram 3
• Anthropic API
• OpenAI API
• SQLite
• pytest
• Docker

The whole knowledge base goes into the prompt instead of a vector search: a hotel FAQ is a few thousand tokens, so retrieval would only add a new way to fail. Because that prompt never changes between guests, Anthropic caches it and repeated requests cost less. An offline mock model lets the whole flow run without an API key.

The core decision, independent of Telegram, so the bot, the terminal demo and the tests all run the same code:

async def handle_guest_message(self, guest: Guest, text: str) -> GuestReply:
    language = await self.guest_language(guest, text)
    history = format_history(
        await self._repo.recent_messages(guest.chat_id, self._settings.history_messages)
    )
    await self._repo.upsert_guest(guest.chat_id, guest.name, guest.username, language)
    await self._repo.log(guest.chat_id, Role.GUEST, Kind.QUESTION, text, language)

    result: ConciergeAnswer | None = None
    try:
        result = await self._llm.answer(text, language, self._knowledge.text, history)
    except LLMError as exc:
        logger.warning("LLM failed, escalating to staff: %s", exc)

    if result and result.found_in_knowledge_base and result.answer.strip():
        answer = result.answer.strip()
        await self._repo.log(guest.chat_id, Role.BOT, Kind.ANSWER, answer, language)
        return GuestReply(language, answer)

    # Not in the knowledge base, or the AI is unavailable: never guess, ask a human.
    reason = "not in the knowledge base" if result else "AI assistant unavailable"
    question_en = result.question_in_english if result else text
    escalation_id = await self._repo.create_escalation(
        guest.chat_id, language, text, question_en, reason
    )
    ...

Guest commands were sent to the AI and to reception as questions

router.message.register(on_staff_reply, IsStaffChat(), F.reply_to_message, F.text)
router.message.register(on_guest_text, private, F.text)
router.message.register(on_guest_other, private)

router.message.register(on_staff_reply, IsStaffChat(), F.reply_to_message, F.text)
# /help or a command the guest may not use: show the welcome text, do not ask the LLM.
router.message.register(cmd_start, private, F.text.startswith("/"))
router.message.register(on_guest_text, private, F.text)
router.message.register(on_guest_other, private)

• 68 automated tests, including a run through aiogram's dispatcher with a fake Telegram connection, and tests of the Claude and OpenAI clients against a fake HTTP server.
• The offline demo answers four scripted guests in four languages, sends the fifth question to staff, and delivers the staff reply back.
• Runs without any API key in demo mode.

• Source code on GitHub↗
• Live demo→