clementbouni-arch/hilton-yaounde-rag-concierge

A Retrieval-Augmented Generation assistant for hotel guest support — case study on the Hilton Yaounde, built with Hugging Face Transformers, sentence-transformers and pypdf # Hilton Yaounde Virtual Concierge ### A Retrieval-Augmented Generation (RAG) assistant for hotel guest support, built and documented from Yaounde, Cameroon > **Disclaimer** — This is an independent, self-directed portfolio project built to learn and > demonstrate Retrieval-Augmented Generation. It is **not affiliated with, endorsed by, or built > using any non-public data from Hilton Worldwide, Inc.** All hotel policies, rates, and documents > used here are fictional and were authored specifically for this project. --- ## Table of contents - Why this project - The problem - Architecture - Repository structure - Getting started - Walkthrough & results - Tech stack - Security notes & where this is going next - Author - Acknowledgments - License ## Why this project Hospitality in Cameroon is growing fast, and international hotels in Yaounde field an enormous volume of repetitive guest questions every day — check-in times, wifi, breakfast hours, spa prices, airport shuttle cost. During peak season, or with a short-staffed front desk, response times slip and answers get inconsistent depending on who picks up the phone. I built this project around a question I kept running into through my cybersecurity/cloud engineering work: **how do you give a language model reliable, current knowledge about a specific business, without letting it invent facts?** This is exactly the problem Retrieval-Augmented Generation (RAG) is designed to solve — and it connects directly to what I want to research at the graduate level: observability in MLOps pipelines, and the security of LLM-based agents (prompt injection, tool/context poisoning, data leakage through retrieval). Instead of a generic dataset, I picked a scenario close to home: a **virtual concierge for the Hilton Yaounde**, grounded only in self-authored documentation that mirrors what a real front-desk knowledge base looks like — PDFs written for people, not machines. ## The problem, in one sentence > A generic L …