<![CDATA[Conventional Learning Management Systems (LMSs) mainly provide fixed materials and predefined activities, while cloud-based generative artificial intelligence may require external data transmission and produce responses that are not aligned with lecturer-approved course materials. This paper presents a local-first, course-grounded learning system using structured document processing, parent-child Retrieval-Augmented Generation (RAG), semantic retrieval, and quantized Large Language Models (LLMs). The system provides cited conversational responses, source-scoped quiz generation, lecturer moderation, image annotation, and interaction-history recording. Five quantized 7B and 14B models were evaluated with 20 course-grounded questions. The 7B models reached 57.14-58.95 tokens/s using approximately 5 GB of VRAM, whereas Qwen3 14B achieved the highest Faithfulness and Answer Relevancy scores of 0.882 and 0.906. The results demonstrate a trade-off between local inference performance and course-grounded response quality.]]>
Copyrights © 2026
