CADSmith: Multi-Agent CAD Generation with Programmatic Geometric Validation
TLDR
A multi-agent pipeline for text-to-CAD generation using CadQuery code with iterative refinement via programmatic geometric validation from OpenCASCADE.
Reasoning
The paper introduces a novel closed-loop refinement approach combining exact geometric measurements with visual assessment, achieving strong quantitative improvements over a zero-shot baseline. However, the evaluation uses a custom benchmark and only compares against a single baseline, limiting generalizability.
Read-first score
Read-first score 64.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 94.
Field roles
Rank sensitivity
Stability: volatile; rank range: 10.
Keyword Scores
Deep Analysis
Innovations
- Multi-agent pipeline with nested correction loops: inner loop for execution errors, outer loop for programmatic geometric validation
- Programmatic geometric validation combining exact OpenCASCADE kernel measurements (bounding box, volume, solid validity) with holistic visual assessment from a vision-language model Judge
- Retrieval-augmented generation over API documentation instead of fine-tuning to adapt to evolving CAD libraries
Methodology
CADSmith is a multi-agent system that generates CadQuery code from natural language prompts and iteratively refines it through two nested loops: an inner loop resolves execution errors, and an outer loop performs programmatic geometric validation using exact measurements from the OpenCASCADE kernel and a vision-language model Judge. It uses retrieval-augmented generation over API documentation to stay current with the CAD library.
Key Results
On a 100-prompt benchmark across three difficulty tiers, CADSmith achieves 100% execution rate, median F1 0.9846, median IoU 0.9629, and mean Chamfer Distance 0.74, substantially outperforming a zero-shot baseline.