01
A CAD file contains more knowledge than its filename exposes
Most engineering workflows still treat CAD primarily as a document to be opened in a specialized application. The file may contain a rich assembly with hundreds or thousands of occurrences, but outside the CAD system a user often sees little more than a filename, revision, thumbnail, and metadata.
An AI assistant presented with only that surface information cannot answer many of the questions engineers actually care about: Which members are identical? Which parts differ only by hole pattern? What is connected to this component? Which feature changed between revisions? Which occurrence is located at this position?
Those are not fundamentally language questions. They are structured geometry and relationship queries.
02
The digital thread already points toward structured product data
Manufacturing standards work has been moving toward richer digital product definitions for years. NIST describes the digital thread as an authoritative flow of product information across design, manufacturing, and inspection, while STEP supports exchange of product and manufacturing information between engineering systems.
NIST research has also examined graphs for linking product-lifecycle data and persistent identifiers for tracing engineering characteristics across systems.
The opportunity for AI is to build on this direction rather than bypass it. A reasoning model becomes more useful when it can query structured engineering entities directly.
03
A queryable model needs identity
Geometry without identity is difficult to connect across time. If a component is exported, transformed, inspected, revised, or observed in the field, the system needs a way to determine whether the new information refers to the same engineering entity.
Stable identifiers are therefore foundational. A useful representation should distinguish a part definition from an occurrence of that part in an assembly, preserve the transform that locates the occurrence, and retain enough provenance to trace the object back to its source.
This becomes especially important when identical stock components are modified differently or when two occurrences share geometry but serve different roles.
04
Geometry becomes a graph of engineering facts
Once entities have stable identity, geometry can be organized as a graph rather than a loose collection of measurements.
Nodes can represent parts, occurrences, features, coordinate frames, documents, observations, or requirements. Edges can represent relationships such as contains, connected_to, aligned_with, supported_by, derived_from, observed_as, or changed_from.
The exact ontology will evolve with use cases. The important shift is that engineering relationships become explicit enough for deterministic tools and AI systems to traverse.
05
Queryable does not mean detached from the source
A knowledge layer becomes dangerous if it silently drifts away from the engineering artifacts it summarizes. Every extracted fact should therefore retain provenance: source file, entity identifier, transform, algorithm or rule, version, and verification state where applicable.
That allows downstream AI to answer with evidence. It also permits independent re-computation when a model changes, a geometry engine improves, or a source file is revised.
This is the core role GeometryLab is intended to serve: convert engineering geometry into a structured representation while preserving enough evidence to remain inspectable.
06
The interface can finally become conversational
Once engineering facts are queryable, language models become powerful interfaces. An engineer can ask a natural-language question, the model can translate it into structured geometry and relationship queries, and the response can be assembled from returned evidence.
The important detail is architectural: the language model does not need to memorize the CAD model. It needs permission to ask the geometry system precise questions.
That pattern scales beyond one AI vendor or one model. The engineering knowledge remains in the deterministic layer; different reasoning systems can connect through explicit tools or protocols.