RAG-Enhanced Socratic Tutoring: Architecture for Domain-Grounded Educational AI in Engineering Simulation Games

Large language models deployed as educational tutors face a fundamental challenge: they can hallucinate plausible-sounding but incorrect information, potentially teaching learners false content. This paper presents a Retrieval-Augmented Generation (RAG) architecture designed specifically for educational AI tutoring in engineering simulation games, where factual accuracy about domain content is essential for effective learning. We describe a complete RAG pipeline that grounds AI tutor responses in verified knowledge bases, enabling accurate Socratic dialogue about complex engineering topics such as water infrastructure design. The architecture addresses key educational requirements including concept-based chunking strategies, confidence-aware retrieval with fallback mechanisms, multi-tier adaptation for diverse learner populations, and integration with real-time simulation state for contextually relevant tutoring. The proposed system maintains pedagogical effectiveness while eliminating hallucinated engineering "facts" through three innovations: (1) a hierarchical knowledge base organized around learning objectives rather than document structure, (2) a hybrid retrieval strategy combining sparse (BM25) and dense (neural embedding) methods for robust concept matching, and (3) a confidence-gated response generation that admits uncertainty rather than fabricating answers. We demonstrate the architecture through application to the AMAIG educational game platform, where learners design water storage and distribution systems using modular AME-500 components. The RAG-enhanced tutor provides accurate responses about system specifications, physics principles, and design trade-offs while maintaining Socratic questioning strategies that promote active learning. Preliminary evaluation indicates high factual accuracy (>95% alignment with knowledge base) while preserving pedagogically appropriate questioning behavior.