Sovereign by Design — an open-source, on-premise chatbot for student advisory at the University of Bonn
Generative AI chatbots have become a default first stop for students asking about study programmes, admissions, and examination regulations. General-purpose LLMs hallucinate, lack institutional context, and cannot be held accountable for the policies they describe — and most higher-education implementations make this worse by running on proprietary cloud-hosted US models, bringing GDPR, data residency, and vendor lock-in into the bargain.
Details
This talk presents a collaborative project between the University of Bonn's Computing Centre (HRZ) and Interaktiv: an open-source AI chatbot for student advisory, built entirely on open-weight language models deployed on-premise inside the university's infrastructure.
Retrieval-Augmented Generation runs against curated institutional sources — programme descriptions, examination regulations, module handbooks, official web content. Paragraph-level citation makes responses traceable; predefined fallback strategies escalate to human advisors when the system is uncertain.
Four implementation challenges surfaced in the evaluation — data readiness, topical boundaries, RAG architecture, and adversarial inputs. The finding worth carrying home: the constraints of on-premise deployment and curated grounding produced more reliable, institutionally aligned advice.
What attendees take away
- A working sovereign-AI architecture, fully named: open-weight LLMs + on-premise deployment + RAG against curated institutional sources + paragraph-level citation + predefined human-escalation fallback — what each piece does and why each was chosen against the cloud-hosted default.
- The four implementation challenges that decide whether the project ships — data readiness, topical-boundary enforcement, RAG-pipeline architecture decisions, and adversarial-input handling — what each one cost in this case and how an institution should budget for them.
- A reference case for institutional buyers considering the same path — a German public university running this, with publishable evaluation findings to point procurement at.
