HEAR: A Human-Centered AI Framework for Integrating Retrieval-Augmented Generation into Engineering Education

Abstract

The increasing reliance on ready-made responses from Large Language Models (LLMs) in education risks reducing critical thinking and engagement, especially in complex domains like engineering. This dependency is particularly problematic where learning requires deep reasoning, ethical awareness, and sensitivity to local context. In response, this paper introduces HEAR (Human Centered Engineering Education with AI Retrieval), an AI pedagogical framework that integrates Retrieval Augmented Generation (RAG) with human guidance, course-specific patterns, and ethical guidleines. The framework draws on the Copenhagen Manifesto, emphasizing transparency, fairness, and critical thinking in AI-enhanced education. HEAR was implemented in a university-level programming course and evaluated along two dimensions. First, Recall@k was used to assess the retrieval accuracy of the RAG system. Results show that Recall at 3 was 0.39, improving to 0.94 at k equals 7, beyond which gains diminished. Based on this, HEAR was configured to retrieve seven context chunks per query. Second, pedagogical effectiveness was measured using a structured survey comparing HEAR to baseline LLM response. Six participants (four educators and two students) evaluated responses across nine criteria. One sample t-tests showed statistically significant improvement in eight of nine categories, including conceptual understanding, scaffolding, and teaching quality. Effect sizes were large, and internal consistency was high (Cronbach’s alpha equals 0.82). Educators rated HEAR significantly higher than students on scaffolding, which shows strong recognition of its structured support for learning. The results indicate that HEAR offers a viable and replicable framework for integrating RAG into engineering education while preserving human oversight and aligning with curriculum goals.