Study describes the deployment and testing of CourseChat, a local study assistant for six courses
The study presents CourseChat, an assistant that allows six separate courses to share two local AI servers. The authors compared models and retained the production model with 8 billion parameters. The system offers 435 prepared questions; the study does not demonstrate improved learning outcomes.
The study authors described the deployment and evaluation of CourseChat, a study assistant for undergraduate business education. The system uses RAG, which means generating answers using information retrieved from course materials. Six separate courses share two local AI servers with a FastAPI interface, a local vector database and a language model served through Ollama. The assistant is deployed behind the university’s web gateway and is intended for integration into Moodle; both teaching materials and student conversations remain on the institution’s infrastructure.
According to the authors, several larger models failed to meet the speed requirement for teaching, while models with 12 and 7 billion parameters met it. A separately tested model with a mixture-of-experts architecture improved some corrections but introduced new factual errors and problems with continuity between responses. The authors therefore retained the production model with 8 billion parameters until an overall improvement from replacing it can be demonstrated.
According to the authors, software changes improved recognition of the topic of follow-up queries while maintaining boundaries between individual courses. The system contains 435 prepared questions for 65 modules, so practice does not have to depend on generating questions on demand. The study does not demonstrate improved learning outcomes; evaluation by instructors, capacity under peak load and full verification of access through the public web gateway remain outstanding tasks.
Why it matters
The case study provides a concrete architecture for running a study assistant on an educational institution’s infrastructure. The model comparison shows why size alone is not enough when choosing a model: speed, fidelity to source materials, continuity between responses and operational compatibility also matter for teaching.
Two audiences, two different impacts
What this means
For individuals
Students in the participating courses can practise using prepared questions for individual modules. However, the study does not provide evidence that using the assistant improves learning outcomes.
For a business
For an educational institution, this is a documented example of sharing two local AI servers across six separate courses while keeping materials and conversations on its own infrastructure. Capacity under peak load remains unverified.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.