geotoken — Overview

geotoken is a geometric tokenizer with adaptive quantization. It converts 3D geometry — CAD (STEP, IGES, B-Rep) and meshes (STL, OBJ) — into discrete token sequences suitable for transformer-based models, at three levels:

• Mesh-level — raw vertex/face geometry tokenization.
• Parametric-level — construction history (sketch-and-extrude command sequences, DeepCAD format).
• Topology-level — B-Rep graph structures with feature vectors.

The key idea: adaptive precision

Adaptive precision quantization allocates more bits to geometrically complex regions (high curvature, dense features) and fewer bits to flat/simple regions — reducing token count while preserving the features that matter.

Tier	Bits	Levels	Use case
DRAFT	6	64	Fast preview, low bandwidth
STANDARD	8	256	Balanced quality/size (default)
PRECISION	10	1024	High fidelity, lossless-adjacent

A first taste

from geotoken import GeoTokenizer, QuantizationConfig, PrecisionTier
import numpy as np

vertices = np.array([[0, 0, 0], [1, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=np.float32)
faces = np.array([[0, 1, 2], [0, 1, 3], [0, 2, 3], [1, 2, 3]], dtype=np.int64)

config = QuantizationConfig(tier=PrecisionTier.STANDARD, adaptive=True)
tokenizer = GeoTokenizer(config)

tokens = tokenizer.tokenize(vertices, faces)
reconstructed = tokenizer.detokenize(tokens)

impact = tokenizer.analyze_impact(vertices, faces)
print(f"Mean error: {impact.mean_error:.6f}")

Key classes

• GeoTokenizer — mesh tokenization.
• CommandSequenceTokenizer — CAD command sequences.
• GraphTokenizer — B-Rep topology graphs (consumes a cadling TopologyGraph).
• CADVocabulary — token → integer-ID encoding.

Status

Maturity: Stable

geotoken is the toolkit’s best-tested package (400+ tests) and is a pure NumPy/Pydantic library — trimesh is optional, only for mesh I/O. It integrates directly with cadling topology and feeds ll_stepnet.

Use the sidebar for Installation, Usage, and the API Reference.