Cost-effective and minimal-intervention BIM information retrieval via condensed multi-LLM agent code generation

Abstract

The Architecture, Engineering, and Construction (AEC) industry increasingly relies on Building Information Modelling (BIM) to manage complex data. However, manual information retrieval remains inefficient and widespread. Existing automatic approaches focus on Industry Foundation Classes (IFC) format with template-based support, while the potential of large language models (LLMs) to generate retrieval code across diverse domain-specific languages and libraries remains underexplored. This paper introduces Alignment-Refinement Coder for BIM (ARCBIM), an LLM-based system with a ‘prefix’ function alignment module, a ‘suffix’ code refinement module, and an ‘inflexion’ condensed agent-design to improve cost-effectiveness. Evaluation on 80 diverse queries with the Revit C# API showed that ARCBIM reduced the average error count to 1.26 per query, compared with 5.75 for the baseline Standalone Agent. Moreover, 80 % of generated code became usable within three local refinements. This system enhanced flexible BIM data retrieval with limited intervention across multiple complexities, providing a more user-centric automatic retrieval solution.