Samenvatting

"While integrating Large Language Models (LLMs) can provide personalised, continuous feedback at scale in large courses, doing so responsibly requires technical accuracy
and data privacy. In this study, we present an on-premises, privacy-preserving chatbot tutor for a university course that focuses on energy and sustainability related tasks.
This tutor is built using Ollama and Docker, and is aligned to the curriculum via Retrieval-Augmented Generation (RAG). The first phase of a three-phase evaluation involves
running a full-factorial experiment (three LLMs × three RAG datasets × six prompts x five repetitions; 270 runs) under an automated Puppeteer protocol. The outputs are rated on seven dimensions, including correctness, hallucinations, pedagogical value and repeatability, that reflect the failure modes observed in external audits. The result is a practical framework for optimising LLM/RAG configurations so that energy-domain learning tasks receive reliable feedback without sending student data off-campus."