LatticeLabs Toolkit

A monorepo of Python packages for CAD document processing, neural networks for 3D geometry, geometric tokenization, optical CAD recognition, and generative CAD modeling.

📚 Documentation: https://latticelabsai.github.io/ll_toolkit/ — per-package guides, tutorials, concepts, and a generated API reference. Source lives in site/.

Packages

Package	Path	Description
cadling	cadling/	CAD document processing toolkit (docling-inspired). Multi-format parsing (STEP, STL, BRep, IGES), topology analysis, RAG-ready chunking, and synthetic data generation.
ll-stepnet	ll_stepnet/	Neural network package for STEP/B-Rep CAD files. Tokenization, feature extraction, topology encoding, and task-specific models.
geotoken	geotoken/	Geometric tokenizer with adaptive quantization for CAD and mesh data. Mesh, parametric, and topology-level tokenization.
ll-ocadr	ll_ocadr/	Optical CAD Recognition. DeepSeek-OCR-inspired 3D geometry processing for LLMs (tiled chunks + global context). HF-native inference (run_ll_ocadr_hf.py); vLLM serving is experimental/future.
ll-gen	ll_gen/	Generation orchestration: neural propose, deterministic dispose in a kernel sandbox. Ships DeepCAD-trained, program-based generators (native MLX) that produce measured-valid CAD — an autoregressive command generator (validity 0.914) and a latent diffusion (sampled-z 0.934), measured through the real kernel. Includes the propose/dispose sandbox + REINFORCE alignment loop (python -m ll_gen.training.run); the legacy VAE/VQ-VAE orchestrator generators ship untrained.
ll-clouds	ll_clouds/	Point-cloud processing & analysis (NumPy/SciPy): PLY/PCD/XYZ I/O + mesh sampling, normalize/voxel/FPS/outlier preprocessing, normals & curvature, ICP registration, RANSAC/Euclidean segmentation, lazy cadling/ll_ocadr bridges.
ll-brepnet	ll_brepnet/	B-Rep face-segmentation network (UV-Net / BRepNet lineage) over coedge topology + UV-grid geometry. Trained on the Fusion 360 Segmentation split (test mIoU 0.828) with a parity-verified native-MLX port. MIT-licensed, built on cadling's B-Rep machinery (no BRepNet/UV-Net code — see ATTRIBUTION.md).

Quick Start

Prerequisites

• Python 3.9 - 3.12
• Conda (Miniconda or Miniforge recommended)

Installation

PyTorch must be installed via conda-forge (not pip) to avoid OpenMP library conflicts on macOS.

# Clone the repository
git clone https://github.com/LatticeLabsAI/ll_toolkit.git
cd ll_toolkit

# Create the conda environment (installs PyTorch, pythonocc, and all packages)
conda env create -f environment.yml
conda activate cadling

The environment installs cadling, ll_stepnet, and geotoken as editable packages. To install individual packages manually:

pip install -e ./cadling          # CAD document processing
pip install -e ./ll_stepnet       # STEP/BRep neural networks
pip install -e ./geotoken         # Geometric tokenizer
pip install -e ./ll_ocadr         # Optical CAD recognition
pip install -e ./ll_gen           # Generation orchestration

Optional dependency groups (from root pyproject.toml)

pip install -e ".[dev]"        # Testing, linting, docs
pip install -e ".[cad]"        # CAD processing (trimesh, networkx, numpy-stl)
pip install -e ".[ml]"         # ML (transformers, accelerate, einops)
pip install -e ".[vision]"     # Vision (opencv, easyocr, matplotlib)
pip install -e ".[hub]"        # HuggingFace Hub integration
pip install -e ".[drawings]"   # 2D drawings (DXF, PDF)
pip install -e ".[all]"        # Everything

Usage

cadling CLI

# Convert a CAD file to JSON or Markdown
cadling convert model.step --format json

# Chunk a CAD file for RAG
cadling chunk model.step

# Generate synthetic Q&A pairs
cadling generate-qa model.step

Python API

from cadling.backend.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("model.step")

from geotoken import GeoTokenizer

tokenizer = GeoTokenizer()
tokens = tokenizer.tokenize(mesh)

from stepnet.encoder import StepNetEncoder

encoder = StepNetEncoder()
embeddings = encoder.encode(step_data)

Architecture

Input CAD File (STEP / STL / BRep / IGES)
  |
  v
cadling: DocumentConverter
  -> Format Detection -> Backend Selection
  -> Backend (format-specific parsing)
  -> Pipeline (Build -> Assemble -> Enrich)
  -> CADlingDocument
  |
  +-> Chunking (RAG)           # cadling.chunker
  +-> SDG (Q&A pairs)          # cadling.sdg
  +-> Export (JSON / Markdown)
  |
  +-> geotoken: tokenize geometry for neural models
  +-> ll_stepnet: neural STEP/BRep processing
  +-> ll_ocadr: optical CAD recognition
  +-> ll_gen: generative CAD modeling

Development

Running Tests

# All packages (from repo root)
pytest

# Individual packages
cd cadling && pytest tests/unit/ -v
cd ll_stepnet && pytest tests/ -v
cd geotoken && pytest tests/ -v

Linting and Formatting

ruff check .
black .
mypy cadling/cadling ll_stepnet/stepnet geotoken/geotoken

Test Markers

pytest -m "not slow"                # Skip slow tests
pytest -m "not requires_gpu"        # Skip GPU tests
pytest -m "not requires_pythonocc"  # Skip pythonocc tests
pytest -n auto                      # Parallel execution

Project Structure

ll_toolkit/
  cadling/           # CAD document processing toolkit
    cadling/         #   Python package
    tests/           #   Tests
  ll_stepnet/        # Neural networks for STEP/BRep
    stepnet/         #   Python package
    tests/
  geotoken/          # Geometric tokenizer
    geotoken/        #   Python package
    tests/
  ll_ocadr/          # Optical CAD recognition
    tests/
  ll_gen/            # Generation orchestration
    ll_gen/          #   Python package
    mlx/             #   native-MLX trained generators
    tests/
  ll_clouds/         # Point-cloud processing & analysis
    ll_clouds/       #   Python package
    tests/
  ll_brepnet/        # B-Rep face-segmentation network (UV-Net/BRepNet lineage)
    ll_brepnet/      #   Python package
    mlx/             #   native-MLX trainer
    tests/
  docs/              # Research docs and plans
  pyproject.toml     # Root config (tooling, shared deps)
  environment.yml    # Conda environment definition

License

MIT

Links

• GitHub