Text-to-CAD: Generating Editable, Parametric B-Rep CAD Models from Natural Language

1 Introduction

This article introduces Zoo's Text-to-CAD system. It also provides relevant background, outlines the underlying methodology, offers usage guidelines, and discusses the system's current limitations and future direction.

1.1 Motivation

What if you could design physical objects simply by describing what you wanted? That’s the core idea behind Text-to-CAD. It’s a system that takes your text (e.g., “Make a flange with six evenly spaced holes”) and turns it into a fully parametric CAD model, written in KCL; our in-house CAD programming language.

Crucially, this isn’t just about generating a mesh for visual appreciation. Text-to-CAD produces precise, engineering-grade B-Rep geometry. This means you receive a real, editable model ready for the next step in your development process — whether that’s tweaking dimensions, running simulations, or sending it directly to manufacturing.

It’s CAD, but faster and more intuitive.

2 Background: The Need for Accessible Parametric Modeling

Why did we build this? Aren’t CAD tools already powerful? Yes, but traditional CAD is often hampered by complexity. Most tools expect users to be fluent in specific workflows, menus, and mouse clicks. Without that fluency, users face a steep learning curve, or they might outsource their ideas to someone with the requisite skills.

We believe people should be able simply to say what they want to create.

This initiative isn’t about replacing established CAD tools; it’s about making design capabilities more accessible. Automation is central to our philosophy, and Text-to-CAD aligns perfectly with that. By turning natural language into code, and subsequently into geometry, we're lowering the barrier to entry and potentially speeding up workflows for everyone, from seasoned engineers to total beginners.

3 Technical Context: Boundary Representation (B-Rep) for Engineering Applications

Many people, particularly those from outside the physical engineering disciplines, associate the phrase "3D model" with a polygonal mesh—essentially, a collection of connected triangles. While meshes are well-suited to applications such as rendering and gaming, this representation scheme has limitations that make it a poor fit for CAD. Engineering workflows require the precision and editability provided by B-Reps¹.

There are three key characteristics of B-Rep that make it the preferred choice for CAD applications:

• Precision: B-Rep geometry represents surfaces exactly – curves are true curves, not approximations formed by tiny straight lines. This is critical for engineering applications where high precision is required.
• Editing: Need to adjust a fillet radius or resize a hole? With B-Rep, it's often as straightforward as updating a parameter. Modifying a mesh typically involves manipulating vertices directly, which can be complex and error-prone.
• Workflow compatibility: B-Rep is the standard representation for many CAD and CAM tools. For tasks like running simulations or manufacturing a part, B-Rep data is generally preferred or required.

In short, meshes are often sufficient for visuals, while B-Rep is designed for building real stuff.

4 The Text-to-CAD System: Approach and Methodology

Text-to-CAD leverages several of Zoo’s internal tools to generate an editable, water-tight solid model based on user input. The process involves several steps:

• Prompt parsing: The user's input text is analyzed to identify parameters, constraints, and geometric relationships.
• KCL code generation: A series of machine learning models generate and refine KCL code that defines the geometry for the requested object. These models utilize our KCL documentation to produce high-quality, idiomatic KCL code.
• Geometry validation: The generated KCL code is executed feature-by-feature in our geometry engine, and the resulting model is tested for validity. Any errors detected are fed back to the ML models to further refine the KCL code.
• File export: In addition to returning the generated KCL code, Text-to-CAD provides the model geometry in the standardized B-Rep or mesh format selected by the user.

5 Implementation and Usage

Whether you are a designer, developer, or simply curious about AI and 3D design, Text-to-CAD offers a straightforward interface to turn ideas into precise models, regardless of expertise level.

• In Zoo Design Studio: Our flagship modeling software is tightly integrated with Text-to-CAD. Click the green Text-to-CAD button, type your prompt (try “A flange with a 50mm diameter, 6 holes, and 10mm thickness”), and the 3D model appears, along with the generated KCL code. You can then edit the code directly or tweak the model using the app’s GUI.

• On text-to-cad.zoo.dev: For a lightweight web interface, visit text-to-cad.zoo.dev, where you can experiment with Text-to-CAD directly in your browser with no installation required.

• Via our API: If you're interested in integrating Text-to-CAD into your own tools or workflows, consider the Text-to-CAD API. Our client libraries (available in Python, Rust, Go, and JavaScript) facilitate getting started, whether automating part generation or building a plugin for another modeling tool.

# Prompt the API to generate a 3D model from text.
response = create_text_to_cad.sync(
    client=client,
    output_format=FileExportFormat.STEP,
    body=TextToCadCreateBody(
        prompt="Design a gear with 40 teeth",
    ),
)

• As a plugin for your favorite application: Text-to-CAD can potentially be integrated into various software applications. We’ve published a few open-source plugins – for example, for Blender and Discord. These can be used as-is, modified, or serve as a starting point for building integrations with other 3D modeling tools.

5.1 Prompting Guidelines

Here are a few tips for effectively using Text-to-CAD:

• Be specific: Generated models adhere closely to the dimensions and constraints provided. If important design elements are omitted, Text-to-CAD will attempt to infer the missing details based on its training.
• Be concise: Shorter prompts (typically 1-2 sentences) tend to succeed more often than very long ones. The system supports prompts up to a few thousand words, but longer inputs generally take more time to process.
• Focus on mechanical parts: Text-to-CAD is primarily fine-tuned to generate mechanical components like flanges, plates, brackets, supports, gears, and bolts. However, feel free to experiment with other prompts – for instance, try “a spatula” or “a fun mascot”.

6 Current Status, Limitations, and Future Work

It's important to understand the current state of Text-to-CAD:

• Ongoing Development: Text-to-CAD is actively being developed alongside the KCL language and our proprietary geometry engine. Its capabilities are evolving in tandem with these underlying technologies. (See our public roadmap for more details).
• Potential for Invalid Geometry: The generated KCL code is validated and executed before being returned. However, if our geometry engine fails to construct a valid model from the code, a 422 error code may be returned instead of geometry.
• Non-determinism: Submitting the same prompt multiple times may produce different results. Occasionally, a prompt that initially fails might succeed on a subsequent attempt, and vice versa.

6.1 Future Work

We view Text-to-CAD as being in its early stages and have significant plans to expand its capabilities:

• Support more CAD features and assemblies: As the Zoo ecosystem matures, Text-to-CAD will be trained to utilize more advanced modeling techniques, including assembling multiple parts into complex designs.
• Physics and simulation integration: We aim to incorporate considerations like stress analysis, material constraints, and manufacturability checks directly into the generation process.
• Multimodality: Design concepts are often conveyed using images as well as text. Future versions may incorporate visual input alongside language prompts.

Follow us for updates, or try Zoo Design Studio for free to get started with Text-to-CAD.

Footnotes