AI-Driven Knowledge Transfer in Architectural Education

The use of large language models (LLMs) in architecture is still at an early stage, but is becoming increasingly important due to the need for local, domain-specific assistance systems. This study investigates how such models can be developed and extended under local hardware conditions in order to provide architecture-specific knowledge in a context-appropriate manner. For this purpose, a methodological approach was chosen that combines lightweight fine-tuning with a two-stage retrieval augmented generation system (RAG). Two model series were tested: Mini-Spyra for dialogue-oriented knowledge retrieval and IwI-Spyra for semantic analysis of structured planning data (e.g. IFC). The results show that domain-specific training in combination with dynamic knowledge integration leads to significantly more precise, comprehensible answers. The article provides a transferable model for the use of AI in teaching and planning practice - data protection-compliant, locally executable and didactically comprehensible. The study thus contributes to the development of explainable AI tools in an architectural context. This research was conducted within a "Young Researchers" project supported by the Jade University of Applied Sciences.