Extracting Information from CAD Assemblies Using Large Language Models and Retrieval-Augmented Generation

Digital manufacturing systems require product-specific information retrievable from CAD models. Accessing this information is labor-intensive due to inconsistent property naming and complex assembly structures, often requiring structure-dependent and manually adapted extraction routines. This paper presents an LLM-driven method for automated information extraction from native CAD assemblies, enabling adaptive and structure-independent access to CAD information across systems, and providing results in a predefined schema. The approach employs a three-stage retrieval-augmented generation pipeline, which first plans requests using the CAD system’s API documentation. A request is then drafted based on examples from previous successful executions and refined, if necessary, using error-fix documentation. Experiments demonstrate strong performance on medium-sized assemblies, highlighting the approach’s potential to reduce manual efforts in information extraction, while deeply nested hierarchies remain challenging.