CAD_AI

CAD_AI is an experimental FeatureScript generation system for Onshape-style CAD workflows. It combines prompt parsing, retrieval context, local heuristics, and a deterministic fallback path so the pipeline can still produce a usable FeatureScript file even when the AI path is uncertain or partially fails.

Status: this project is still a work in progress. It is not complete, it has known rough edges, and it should be treated as a research and prototyping codebase rather than a production-ready CAD platform.

What this project tries to do

• Turn a natural-language CAD request into an editable FeatureScript file.
• Use retrieved knowledge, examples, and local rules to guide generation.
• Preserve a strong default path that never returns an empty or omitted result.
• Provide a contributor-friendly starting point for further work on validation, repair, and model orchestration.

Current guarantees

The current implementation is designed to avoid the worst failure mode: returning no usable FeatureScript.

• The generator now uses a deterministic guaranteed fallback when the main AI path is weak, invalid, or incomplete.
• That fallback always produces a FeatureScript payload with a valid header, a defineFeature block, and editable parameters.
• The system is intentionally conservative in that mode so it can always deliver something usable instead of silently omitting output.

Repository layout

• ai.js — core generation, validation, repair, and deterministic fallback logic.
• server.js — HTTP API entry point for generation and debugging.
• Auth.js — authentication and Supabase-related helpers.
• learning.js — learning and memory orchestration helpers.
• scripts/ — data import, training, pruning, and smoke-test utilities.
• data/ — knowledge, memory, and dataset assets.
• docs/ — architecture notes and research summaries.

Getting started

1. Install dependencies:

   npm install

2. Create a local environment file if needed with your Groq API keys.
3. Start the server:

   npm start

4. Send a request to the generation endpoint or run the built-in smoke test:

   npm run test:guaranteed

Important notes for contributors

• The project is still incomplete and there are many open issues.
• Some parts are experimental and may change frequently.
• The generator has been intentionally biased toward producing a usable FeatureScript rather than failing silently.
• If you change generation behavior, keep that guarantee in mind and test it carefully.

Known areas that still need work

• Better shape-specific FeatureScript generation for more CAD forms.
• Stronger validation and repair quality.
• More robust prompt interpretation and dimension extraction.
• Cleaner separation of runtime logic, knowledge handling, and UI/API concerns.
• More thorough testing and regression coverage.

Development guidance

If you want to help improve this project:

• Keep the core generation path understandable and modular.
• Prefer small, testable changes over broad rewrites.
• Preserve the guarantee that a FeatureScript result is always emitted when possible.
• Document new behavior, assumptions, and limitations clearly.

Summary

This system is a promising starting point for AI-assisted CAD generation, but it is still rough and evolving. The main design principle now is simple: do not fail silently. When the model is uncertain, the system should still return a concrete FeatureScript rather than omitting the result.