AI Agent Framework Research for Dynamic BIM Query via Retrieval-Augmented Code Generation for IFC

Accessing and retrieving specific information from complex Building Information Modeling (BIM) data, particularly within the Industry Foundation Classes (IFC) schema, remains a challenge for Architecture, Engineering, and Construction (AEC) professionals. Existing methods, including manual parsing, database conversion, and early-stage AI applications, often suffer from high overhead, inflexibility, or reliance on pre-processed, static data views. This study introduces a novel AI agentic framework that leverages a Large Language Model (LLM) to dynamically generate executable code for BIM data query. The core of this framework is a Retrieval-Augmented Generation (RAG) pipeline that, instead of retrieving factual data, retrieves relevant code exemplars demonstrating the use of IFC parser. The LLM-based agent analyzes a user’s natural language query, retrieves functionally relevant code snippets from a specialized vector database, and synthesizes these examples into a tailored, executable script. This “BIM code-retrieval” approach eliminates the need for IFC conversion using databases, allowing the agent to directly use the full power of the IFC parser to obtain the data it needs, and for the LLM to make appropriate inferences to the question. The proposed system demonstrates flexibility and query power compared to methods reliant on static data abstractions. It can handle complex queries that are intractable for previous RAG-based approaches. This AI agent code generation framework demonstrates the cost-effective, inference-based human–BIM interaction method in situations where IFC-to-database conversion is difficult. The prototype of study was developed to analyze the performance of the proposed BIM AI Agent framework and its strengths and weaknesses were examined.