This model is intended for hackathon demos and local prototyping of a hotel receptionist assistant.
Example use cases:
Greeting hotel guests
Asking follow-up questions for booking or check-in requests
Explaining that hotel staff or PMS/tool results are required before confirming sensitive information
Handling basic hospitality conversation in Vietnamese and English
Important Limitations
This is a preview demo model, not a production hotel receptionist system.
The model may produce incorrect, incomplete, repeated, or overconfident answers. It must not be used to make real booking, payment, refund, checkout, privacy, safety, legal, or operational decisions without a verified PMS/tool/RAG layer and human review.
In particular, without verified tool/PMS/RAG results, the assistant should not:
claim that it has checked the hotel system
confirm room availability
quote prices
confirm booking status
confirm airport pickup
confirm payment, refund, invoice, or checkout
confirm hotel-specific policies or facilities
claim that a housekeeping or maintenance request has already been completed
Recommended System Prompt
text
1You are a hotel receptionist assistant for CapyInn.
23Hard rules:
4- If there is no tool/PMS result, do not claim that you checked the system.
5- If there is no tool/PMS result, do not confirm room availability, booking status, airport pickup, payment, invoice, refund, checkout, or housekeeping completion.
6- If there is no hotel knowledge/RAG context, do not confirm specific policies, facilities, prices, or service availability.
7- For booking or availability requests, ask for missing details and say hotel staff/system must verify before confirmation.
8- For operational requests, acknowledge the request and say it will be passed to hotel staff; do not claim it has already been completed.
9- Never output tool calls, JSON, role labels, or the word "model".
10- Reply in the guest's language unless they request otherwise.
llama.cpp Example
bash
1llama-cli \2 -m capyinn-gemma-4-Q5_K_M.gguf \3 -ngl 99\4 -c 2048\5 -n 200\6 --temp 0\7 --top-k 1\8 -st \9 --no-display-prompt \10 --reasoning off \11 -sys 'You are a hotel receptionist assistant for CapyInn. If there is no tool/PMS/RAG result, do not claim that you checked the system and do not confirm booking, payment, pricing, policy, facility, or operational status.'\12 -p 'Khach hoi: Toi muon check-in som thi khach san co ho tro khong?'
Demo Scope
This build is optimized for a local hackathon demo on Apple Silicon using llama.cpp Metal.
It is not yet optimized for:
production-grade retrieval
live PMS integration
function calling
payment workflows
guaranteed policy accuracy
long multi-turn operational workflows
Evaluation Snapshot
Internal fixed-prompt qualitative evaluation showed useful hospitality behavior, but the model did not pass the full hard-rule production gate.
Known remaining risks include:
occasional overconfident system-check language
incomplete follow-up questions for some reservation requests
possible repeated text depending on runtime/template settings
For a production assistant, use this model behind routing, RAG/PMS context injection, and a hard-rule output validator.
