Recognition, Retrieval, and Response: The AITEE Framework for Socratic Tutoring in Electrical Engineering

Intelligent tutoring systems combined with large language models offer a promising approach to addressing the teacher bandwidth problem in electrical engineering education. However, standard text-based retrieval fails in symbolic domains where topology dictates solution methodology. In this paper, we present AITEE, an agentic tutoring framework that addresses three fundamental challenges: robust perception of hand-drawn circuits, structure-aware retrieval in symbolic domains, and pedagogically sound dialogue generation. Our circuit recognition pipeline processes informal sketches with high fidelity through a novel line-loss metric for connection validation. To overcome the limitations of text-based similarity in circuit analysis, we propose a structure-aware Multi-Representation Indexing (MRI), which uses Graph Neural Networks to generate topology-based embeddings that retrieve methodologically relevant content based on structural identity rather than textual overlap. The system enforces a strict separation between conversational reasoning and numerical computation, delegating arithmetic verification to SPICE simulation to eliminate calculation hallucinations. Experimental evaluations demonstrate that MRI enables even medium-sized open-source models (70B parameters) to achieve competitive performance on complex circuit topologies where advanced text-based retrieval methods fail, reaching 85% accuracy in methodology application. Qualitative analysis reveals that while instruction-prompted models successfully maintain Socratic dialogue, rigid inquiry-based approaches can impose excessive cognitive load on novices, suggesting that effective AI tutoring requires adaptive scaffolding to balance learner autonomy with necessary cognitive support.

IEEE Access, vol. 14, pp. 50109-50126