CADVERT

Turn CAD files into something an AI can actually read.

A STEP file is a boundary representation — dense surface equations and topology built for CAD kernels, not for language models. Paste one into an LLM and it chokes; show it a screenshot and the model can see the shape but can't tell you the wall is exactly 11.000 mm thick. In engineering, the number is the point.

CADVERT reads the CAD file directly and extracts the exact geometry the engineer authored — dimensions, holes, fillets, wall thicknesses, tolerances — as clean, structured text (and other representations) any model can reason over. Geometry is analytical, straight from the B-REP — never tessellated or approximated.

pip install cadvert

Quick start

import cadvert

result = cadvert.analyze("bracket.step")

print(result.to_text())     # LLM-ready structured document (the flagship)
result.to_dict()            # JSON-safe analysis for quoting / CI / DBs
result.to_json()
result.to_graph()           # networkx face-adjacency graph  (GNN / UV-Net)
result.to_points(2048)      # (N,3) surface point cloud       (PointNet / 3D-CNN)

For an assembly, CADVERT also reports what the file says the parts are — catalog components arrive as real order codes, which describe function in a way geometry alone cannot:

DESIGN: Cylindrical face chamfering machine
AUTHORED IN: Autodesk Inventor 2021
COMPONENTS — 51 named in the file
  DIN 625 T1 - 6205 - 25 x 52 x 15
  Belt S5M-300
  ISO 4762 - M8 x 20

Example to_text() output:

PART: bracket
[GLOBAL PROPERTIES]
Bodies: 1  |  Faces: 8  |  Edges: 18
Bbox: X[0,60]  Y[0,40]  Z[0,20]
[FEATURES]
  hole_1: THROUGH HOLE  d=8.000mm  depth=20.000mm  [5/16"]
  pattern_1: PATTERN  2× through hole  d=8.000mm
[SPATIAL RELATIONSHIPS]
  Wall thickness: hole_1 bore → outer surface = 11.000 mm

Supported formats

Representations

Optional extras

pip install cadvert[graph]   # networkx  → .to_graph()
pip install cadvert[mesh]    # trimesh   → area-weighted .to_points(), OBJ
pip install cadvert[dxf]     # ezdxf     → .dxf drawings
pip install cadvert[server]  # FastAPI REST server + web UI
pip install cadvert[llm]     # OpenAI + Claude chat providers (server / desktop app)
pip install cadvert[full]    # everything

Mac, iPad and iPhone app

A native SwiftUI app with the same design and flows as the web UI lives in APPLE OS APP/. On the Mac it bundles the Python + OpenCASCADE engine and works fully offline; on iPad/iPhone it connects to a hosted cadvert-server. Questions can be answered by Apple's on-device model (macOS 26 / iOS 26, no key), OpenAI, or Claude — the server's /chat takes an X-Provider header and the web UI's model menu lists both vendors. See that folder's README for building and DISTRIBUTION.md for selling it (Mac App Store, direct DMG, iOS).

CLI

cadvert part.step -o out.hsd.txt      # convert to an HSD document
cadvert-server                        # REST API on http://localhost:8000

Building an AI CAD tool

Register analyze as a tool in your LLM framework and the model can answer questions about real parts using real dimensions:

def analyze_cad(file_path: str) -> str:
    return cadvert.analyze(file_path).to_text()

Requires Python 3.10+. Core install ships cadquery-ocp (OCCT bindings) and numpy — no other heavy dependencies.