With the advance of generative pre-trained language models, new opportunities arise for domain-independent dialogue-based instruction in the context of intelligent tutoring systems (ITS). We propose an approach for combining large language models (LLMs) with domain-specific knowledge graphs to take advantage of the performance of LLMs in natural language processing tasks while still ensuring a faithful knowledge base. We present a prototypical implementation of an ITS for the structured query language SQL to explore the potential of combining LLM and domain specific knowledge graphs for dialogue based didactic intervention. While learning a programming language relies on both declarative knowledge about the language concepts and procedural knowledge for writing program code with respect to a specific task, in this paper we focus on ITS support for acquisition of declarative knowledge. In our implementation LLMs are employed for semantic parsing relating student’s
questions to the addressed knowledge elements of the domain model as well as for feedback formulation based on retrieved information from the knowledge graph. Likewise, knowledge diagnosis is realized by semantic parsing of student answers and mapping them against the LLM output. We conducted a first exploratory evaluation using the LLM Mixtral 8x7B Instruct in combination with our SQL knowledge graph. Exploring different prompting strategies, the best strategy resulted in 90% correct diagnosis of student answers.