Towards sustainable AI knowledge-base assistants in computer science education: on-premise deployment and...

This study investigates the deployment of artificial intelligence (AI) knowledge-base assistants using open educational resources (OER) for computer science education, demonstrating the feasibility of providing AI-assisted technical support in resource-constrained environments. Our methodology combines structured knowledge extraction from 82 open-licensed Markdown documents, systematic benchmarking of Qwen-7B and DeepSeek-MoE on consumer-grade GPUs (NVIDIA RTX 3060, 12 GB VRAM), and a multi-criteria evaluation framework incorporating response accuracy, computational efficiency, retrieval-augmented generation (RAG)-specific metrics, and energy sustainability. Key results show that Qwen-7B (FP16) achieves 71.5% overall accuracy with 1.4 s average latency, while DeepSeek-MoE (FP16) achieves 79.8% overall accuracy and 82.3% accuracy on multi-hop reasoning tasks. NF4 4-bit quantization reduces VRAM usage by 38.7% for Qwen-7B (6.2 GB → 3.8 GB) and 37.9% for DeepSeek-MoE (5.8 GB → 3.6 GB). Our most energy-efficient configuration consumes 1.8 mWh per query (1.8 Wh per 1,000 queries), measured at the GPU device rail with idle power subtracted. This work makes four contributions: (1) systematic multi-dimensional benchmarks for on-premise, OER-aligned AI knowledge-base assistants in introductory computer science education; (2) hardware-specific deployment guidelines for consumer-grade GPUs without cloud dependency; (3) quantitative evidence that NF4 4-bit quantization-aware fine-tuning preserves educational response utility within 2 percentage points of accuracy degradation while reducing VRAM by 38.7%; and (4) a reproducible evaluation protocol serving as a reference design for future research. The system is characterized as a knowledge-base assistant rather than a validated intelligent tutor; comprehensive pedagogical validation with controlled learner studies remains a priority for future work.