AI-Powered Chatbot for Basic Queries

Abstract

The thesis tackles the Administrative Paradox at the University of Debrecen by developing a localized, zero-hallucination AI assistant called UniAssist. It uses a modular Retrieval-Augmented Generation (RAG) architecture to separate reasoning from institutional knowledge, ensuring high factual reliability. The system employs the Gemini 3 Flash model for fast inference. It also demonstrates strong linguistic performance in Hungarian. Overall, with a 96.6% semantic accuracy rate, the research shows that constrained generative AI can effectively and safely support university administration.