Fine-tuning or Prompting LLMs to accurately read complex 2D framing layouts and truss nodes?

Hey community. I am pushing the limits of multimodal models such as GPT-4o with respect to parsing technical CAD drawings for the structural framing industry. While the vision abilities to read text and recognize fundamental geometry are outstanding, LLMs often fail to preserve the true spatial relationships between connected nodes within intricate and nested structural framing systems (such as multi-residential framing with clustered plate connections). Has anyone had success in creating a vision-to-data pipeline where an AI accurately draws a geometric layout of nodes from an image or vector file, preserving the layout of the model, or any techniques of pre-processing such as segmenting or color coding individual lines to drastically increase accuracy for technical engineering vision tasks.