Abstract:

Generative AI chatbots have become a default first stop for prospective and current students seeking guidance on study programmes, admission procedures, examination regulations, and academic planning. While general-purpose large language models offer impressive fluency, they frequently hallucinate, lack institutional context, and cannot be held accountable by the universities whose policies they describe. The majority of implementations in higher education further intensify these issues by relying on proprietary, cloud-hosted models from US-based providers, raising concerns regarding GDPR compliance, data residency, and vendor lock-in.

This study examines the development and evaluation of an open-source, institutionally hosted AI chatbot for student advisory services, implemented in a collaborative project between the University of Bonn's Computing Centre (HRZ) and Interaktiv GmbH. In contrast to common approaches, the system is based entirely on open-weight language models deployed on-premise within the university’s infrastructure.

Methodologically, the system combines Retrieval-Augmented Generation (RAG) with curated institutional data sources, including programme descriptions, examination regulations, module handbooks, and official web content. A paragraph-level citation mechanism enables traceability of responses, while predefined fallback strategies ensure escalation to human advisors in cases of uncertainty or edge conditions.

The evaluation highlights four critical implementation challenges. First, data readiness emerges as a central bottleneck, as institutional documents require substantial preprocessing to become usable in RAG pipelines. Second, clearly defining and enforcing topical boundaries proves essential to maintain reliability in advisory contexts. Third, architectural decisions in the RAG pipeline, particularly regarding chunking strategies and retrieval tuning, significantly affect answer quality and trustworthiness. Fourth, ensuring safety and robustness requires systematic handling of adversarial inputs, ambiguous queries, and domain edge cases.

The findings indicate that a sovereign, open-source approach can achieve high levels of transparency, controllability, and contextual accuracy, while mitigating risks associated with proprietary systems. Rather than limiting performance, the constraints of on-premise deployment and curated data grounding contribute to more reliable and institutionally aligned AI-supported advisory services.

Keywords:

AI Chatbot, Generative AI, Higher Education.