AI for CAD Tools
A Parasolid file holds a part's full boundary shape but not its feature history. See what an AI agent can read, what the kernel keeps and where to look instead.
·
⏱
8 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
A Parasolid file gives an AI agent a complete, explicit description of a part's shape: faces, edges, vertices, the surfaces and curves behind them, and any attributes the exporting application chose to attach. It does not give the feature history, the parameters, the constraints or the release state, because those live in the CAD application and the PDM. Scope an agent for geometric and tag questions, do the arithmetic in code, and source intent and revision from the systems that own them.
Parasolid is the geometric modeling kernel underneath a large share of the CAD systems engineers use every day, so Parasolid files turn up in supplier handoffs, archives and vault folders more often than most teams realize. When someone proposes pointing an AI agent at those files, the first question is simple: what can the agent actually pull out of one? The answer is more specific than "the geometry" and less than "the model".
A Parasolid file stores the solid as a boundary representation, plus whatever tags the exporting application attached. It does not store the story of how the part was made. That gap decides which questions an agent can answer from the file alone and which ones need a different source. This post walks through what is in the file, what an agent can reliably read, and where the kernel stops.
What a Parasolid file actually contains
The native exchange format for the kernel is called a transmit file. It comes in a text flavor with the extension .x_t and a binary flavor with the extension .x_b. Both hold the same information: the entities that make up one or more bodies, the geometry those entities sit on, any attributes attached to them, and a header that records the schema version used to write the file.
The core of the file is a boundary representation, usually shortened to B-rep. A solid is described by its skin rather than its volume. Faces are bounded patches of surfaces, edges are the curves where faces meet, and vertices are the points where edges end. Between them sit a few connecting entities, such as loops that group the edges around a face, and shells and regions that group faces into closed volumes. The geometry layer then says which plane, cylinder, cone, torus or free-form surface each face lies on, and which curve each edge follows.
That is a precise, complete description of the shape. It is also all the file promises about the shape. If a hole in your part was made by a hole feature with a stored diameter and a depth rule, the file holds a cylindrical face, its radius and the edges that bound it. The rule that produced it is not part of the record.
The related comparison of neutral formats is covered in our comparison of STEP, IGES and Parasolid, which is a useful companion if you are deciding which format to request from a supplier.
IN PRACTICE
Leo uses a Large Mechanical Model trained on 1M+ technical sources. It also provides citations, so we don't have to guess whether a material property or tolerance is correct. We see 96% accuracy on technical queries.
- Dorian G., AI Engineer
What an agent can read from the topology
Topology is where a Parasolid file is strongest for an agent. Because every face, loop, edge and vertex is an explicit record with explicit links, a program can walk the structure and answer concrete questions without guessing.
How many faces, edges and vertices the body has, and how they connect.
Which faces are planar, cylindrical, conical, spherical, toroidal or free-form, with the numeric parameters of each surface, such as a cylinder's radius and axis.
Which faces share an edge, which gives adjacency for questions such as "which flat face does this bore break through?"
The lengths of straight edges and the radii of circular ones, from which hole sizes, fillet sizes and wall thicknesses can be derived.
Whether the body is a closed solid, an open sheet or a wire, and how many separate bodies the file holds.
An agent that has this structure can compute a lot. It can estimate a bounding box, count through-holes by looking for cylindrical faces with matching opposite openings, find all the fillets under a given radius, or flag a body that is not closed. It can also build a compact geometric signature of the part, which is the basis for finding similar or duplicate parts across a library.
Two cautions apply. First, the file is only as readable as the tooling around it. The text flavor can be parsed by a program, but interpreting its numeric records correctly requires the published format reference, and the binary flavor generally needs either a reader library or the kernel itself, which is licensed software. Second, the agent should do the geometry arithmetic in code and not by asking a language model to read the raw records and reason about them. Models are good at deciding which question to ask and poor at summing fillet radii from a block of numbers.
Attributes: the names and tags that travel
Beyond shape, a transmit file can carry attributes. An attribute is a small piece of data attached to an entity, such as a face, an edge or a whole body, defined by an attribute definition that the writing application declares. Applications use them for things like colors, names, layer assignments, material tags and identifiers that let them recognize their own entities again after a round trip.
This is the part of a Parasolid file that varies most from one source to the next. Two files from two different systems can hold the same shape and very different attributes. One may name its faces and carry a color on every body. Another may carry almost nothing. An agent therefore cannot assume that a particular property will be present, and it should treat a missing attribute as "not recorded" rather than as "not true".
Attributes defined by one application are also not guaranteed to mean anything to another. The definition travels with the file, so an agent can read the name and the value, but the exporting system's own conventions decide what the value stands for. That is the same kind of translation question raised in our analysis of what a STEP translator drops, and the same discipline applies: read what is there, and record where the meaning is borrowed from the source system.
What the kernel keeps and what stays behind
The most important thing to understand about Parasolid is the division of labor. The kernel builds and edits solids. The CAD application on top of it keeps the feature tree, the sketches, the constraints, the named parameters, the configurations and the equations. When the application writes a transmit file, it hands over the resulting body, and most of that application-level intent stays in the application's own file.
For an agent, that means several categories of question cannot be answered from the Parasolid file:
Why a dimension has its value, and which parameter or equation drives it.
The order in which features were created and which ones can be suppressed.
Which sketch constraints hold, and so how the part will flex when a dimension changes.
Product and manufacturing information, such as tolerances and datum schemes, which an application may or may not write out as attributes and which is not part of the shape record.
Release state, revision, approvals and where the part is used, which belong to the PDM or PLM layer and not to any geometry file.
An agent that is asked "what tolerance applies to this bore?" should not infer an answer from a cylinder's radius. The honest response is that the file holds a nominal size, and the tolerance lives in the drawing or in the model's annotations.
The same pattern shows up in other formats and in vendor interfaces. For a view from the other side, read what a lightweight visualization format keeps compared with the original and what Siemens' own CAD system exposes to an agent. In each case the exchange file or interface hands over some layer of the model, and the layer that explains the design intent usually sits one level higher.
How to scope a Parasolid agent project
Before anyone writes a connector, decide what the agent is for. Parasolid files are a strong fit for geometric questions and a weak fit for intent questions, so the scope should follow that line.
List the questions you want answered and sort them into shape questions, tag questions and intent questions. Only the first two come from the file.
Sample real files from your own sources. Check the schema versions, check how many carry attributes at all, and check whether the files are text or binary.
Decide where geometry arithmetic runs. It should run in deterministic code that returns numbers the agent can cite.
Decide where the intent and release information comes from, and connect that source separately.
Test on a few dozen parts where the right answers are already known before widening the scope.
This is where a retrieval layer on top of your PDM earns its place. Leo offers integrations with leading PDM and PLM platforms (SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, Arena PLM, and others), and it answers from the geometry, drawings and records your organization already holds, with a source attached to each answer. The file gives the shape. The PDM gives the revision, the state and the usage. Together they let an agent answer a question such as which released parts use this bore size, instead of only describing one body.
Permissions deserve the same attention as scope. An agent that reads files must see only the files its user may see, as discussed in what an AI agent can see in your PDM.
FAQ
See what Leo finds in your CAD files
Search parts and drawings with cited answers from your own data.
Leo connects to your PDM and PLM and answers engineering questions from your own released data, with a source you can check.
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
