Extracting Metadata from CAD and BIM Models: Practical Patterns for AI Pipelines

Most of the value in a CAD or BIM model is not in the visual geometry—it’s in the metadata: part numbers, materials, quantities, classifications, and relationships. The Autodesk APS Model Derivative API makes this data programmatically accessible. Here’s how to extract it reliably and feed it into AI pipelines.

What Metadata Is Available?

After translating a model via the Model Derivative API, four types of structured data are available:

• Object tree: the assembly hierarchy—which parts belong to which sub-assemblies
• Object properties: attributes for each object (name, category, material, dimensions, custom attributes set by the engineer)
• Views: which 2D and 3D views are available in the model
• Geometry: bounding boxes, centroids, and face/edge counts (for geometry-aware workflows)

Extraction Walkthrough

import requests

BASE = "https://developer.api.autodesk.com/modelderivative/v2/designdata"
HEADERS = {"Authorization": f"Bearer {access_token}"}

# 1. Get model manifest (lists available derivatives)
manifest = requests.get(f"{BASE}/{urn}/manifest", headers=HEADERS).json()
model_guid = next(d["guid"] for d in manifest["derivatives"][0]["children"] 
                  if d.get("role") == "3d")

# 2. Get object tree (assembly hierarchy)
tree_job = requests.get(f"{BASE}/{urn}/metadata/{model_guid}", headers=HEADERS).json()
# Poll until complete if large model
root_node = tree_job["data"]["objects"][0]

# 3. Get all properties for all objects
props_resp = requests.get(
    f"{BASE}/{urn}/metadata/{model_guid}/properties",
    headers=HEADERS,
    params={"forceget": "true"}  # bypass 20-object default limit
).json()

# 4. Build a flat lookup: objectid → properties dict
obj_props = {
    item["objectid"]: item.get("properties", {})
    for item in props_resp["data"]["collection"]
}

# 5. Recursive tree walker to build structured BOM
def walk_tree(node, depth=0):
    oid = node["objectid"]
    props = obj_props.get(oid, {})
    yield {
        "id": oid,
        "name": node.get("name", ""),
        "depth": depth,
        "part_number": props.get("Part Number", props.get("PN", "")),
        "material": props.get("Material", ""),
        "quantity": props.get("Quantity", 1),
    }
    for child in node.get("objects", []):
        yield from walk_tree(child, depth + 1)

bom = list(walk_tree(root_node))

Handling Large Models

For Revit models with thousands of elements, the /properties endpoint returns paginated results. Use the forceget=true parameter to get all properties in one response for smaller models, or implement pagination with the objectid filter for large ones. Production pipelines should cache extracted properties in a database (PostgreSQL or Cosmos DB) to avoid re-extracting on every request.

Feeding Metadata into AI Agents

Once you have structured BOM data, it becomes excellent context for AI agents. A few practical patterns: