A lightweight semantic-RAG framework for personalized e-learning in architecture and spatial design education

Architecture and spatial design education requires students to navigate dense historical, conceptual, case-based, visual, and project-based knowledge. Learners must retrieve information, but they must also understand how spatial concepts, precedents, styles, typologies, materials, and pedagogical objectives relate to one another. Conventional vector-based retrieval-augmented generation (Vector-RAG) provides rapid retrieval but frequently yields fragmented evidence due to semantic blindness. Conversely, heavyweight ontology-based systems and generic GraphRAG architectures impose severe curational burdens or high computational overhead, rendering them impractical for agile studio instruction. We propose Semantic-RAG, a lightweight, instructor-auditable framework adding a 10-relation architectural schema, multimodal precedent indexing, compact knowledge graph, hybrid retrieval-reranking, and learning-path recommendation to a validated digital-human teaching system. On 120 benchmark queries, it significantly outperformed Vector-RAG, hybrid RRF, GraphRAG, and ontology baselines (BH-FDR) in retrieval, faithfulness, citation, and path coherence, at sub-second latency and < 4 KB/unit. A pilot ( N = 60) improved design transfer, mental effort, and instructional efficiency.