The manual checking of building permit applications against complex building codes is a time-consuming, error-prone process, as the permit reviewer must manually match the relevant sections and their content to the provided building plans. With the emergence of Artificial Intelligence (AI), specifically Large Language Models (LLMs), it has proven to be capable of overcoming the barrier of complex documentation by converting building codes into natural language. This thesis presents a Graph Retrieval-Augmented Generation (Graph-RAG) framework to explore how an updateable knowledge graph can automate applicability checks during the permit process. The proposed three-stage pipeline is evaluated in accordance with the Ontario Building Code (OBC). In the first phase, a section of OBC was encoded into the knowledge graph. In the second stage, IFC data is parsed into the same knowledge graph with all relevant properties required by OBC. In the third stage, a LangChain ReAct agent is developed to enable dynamic multi-hop reasoning, allowing the system to answer natural-language compliance and applicability questions about various articles. Two different building models were tested, and the system was evaluated across three conditions using a self-developed 35-question survey organized into 4 levels: regulatory classifications, building properties, article selection, and sentence- and clause-level compliance queries. In condition 1, a prompt engineering baseline without access to knowledge graph retrieval, in conditions 2 and 3, the identical Graph RAG with GPT-5.4 and Claude Sonnet 4.6, respectively, were deployed. Results demonstrate that Graph RAG with Claude Sonnet achieves an overall accuracy of 82.35% and macro F1 of 0.673, substantially outperforming the LLM-only baseline and showing that graph architecture performs well when applied to complex hierarchical regulatory documents, compared to building models for review.     «

The manual checking of building permit applications against complex building codes is a time-consuming, error-prone process, as the permit reviewer must manually match the relevant sections and their content to the provided building plans. With the emergence of Artificial Intelligence (AI), specifically Large Language Models (LLMs), it has proven to be capable of overcoming the barrier of complex documentation by converting building codes into natural language. This thesis presents a Graph Retri...     »