AI for Engineering Knowledge Management
How shape-based CAD search finds parts your metadata cannot, how descriptors and nearest neighbour retrieval work, and how to test it on your own vault.
·
⏱
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
Metadata search can only find parts that somebody described well, and the parts most worth reusing are usually the ones described worst. Geometric search removes that dependency by making the model its own query, computing a compact descriptor from the geometry and retrieving nearest neighbours from an index. Getting value from it requires being precise about what similar means, because duplicate detection, family grouping, and interchangeability need different settings, and none of them can be answered by shape alone once material, standard, and lifecycle status matter. Test it against a part you know has a twin, a part you know is unique, and the same part rotated and mirrored. Those three checks, run on your own vault, will tell you more than any demonstration.
An engineer needs a mounting bracket. Something close to it almost certainly exists in the vault, modelled two years ago by somebody who has since left. Finding it requires knowing what that person called it. If they typed brkt_mtg_r3 into the filename and left the description empty, the part is invisible to every text query anyone will ever run.
Geometric search inverts that dependency. Instead of asking what a part was named, it asks what a part looks like, and it answers by comparing the geometry of your model against the geometry of everything else in the library. The part becomes its own search key, which means the quality of the result no longer depends on the discipline of whoever saved the file.
This article covers how shape-based retrieval works, what similarity has to mean before the results are worth acting on, and how to test the capability before a team relies on it.
Why Metadata Search Cannot Find a Part You Have Never Seen
Every search box in a PDM or PLM system is a bet on data entry. The query runs against filenames, part numbers, description fields, custom properties, and classification codes. All of those are things a human typed. None of them is the part.
That bet fails in four predictable ways.
Naming conventions drift. A convention written in 2014 survives until the first rush job, and after that the vault holds three generations of naming logic with no record of which era a given file belongs to.
Description fields are optional, so they are empty. An engineer under deadline fills in what the release process enforces and nothing else, which usually means a part number and a revision.
Classification taxonomies age badly. A tree built around one decade's products puts this decade's parts in whichever bucket is least wrong, and the person searching cannot guess which.
Migrations flatten meaning. Every move to a new system drops the custom properties the new schema has no column for, and those often carried the only real engineering intent.
The deeper problem is not that the metadata is bad. It is that metadata search can only retrieve what somebody remembered to describe, and the parts most worth reusing are described worst, because they were designed under pressure. The result is a vault that holds the answer and a search box that cannot reach it, a pattern covered in why PDM search is broken.
A geometry file, by contrast, is complete by construction. A solid model has to describe every face, edge, and vertex or it will not open. The information needed to recognise the part is already inside the file, fully populated, whether or not anyone documented it.
IN PRACTICE
The geometry search has been invaluable, helping me find standard parts instead of designing new ones, saving a huge amount of time and effort. The search system is smart and CAD-aware.
- eytan s., R&D Engineer
How Shape-Based Retrieval Actually Works
Comparing two solid models directly is expensive. Checking whether one boundary representation matches another face by face is heavy, and running it against a hundred thousand parts for every query is not workable. Practical systems do not compare models. They compare descriptors.
A shape descriptor is a fixed-length numeric summary computed once per part and stored in an index. Two broad families are in common use.
Global descriptors treat the part as a whole and reduce it to a statistical signature. A typical approach samples many pairs of points across the model surface and records the distribution of distances between them, producing a histogram characteristic of the overall shape and unaffected by where the part sits in space.
Structural descriptors treat the part as a graph. The boundary representation already records which faces meet which other faces, along which edges, at what angles, and with what surface types. Encoding that face-adjacency structure captures how a part is built rather than only how it looks, which is what makes it possible to recognise a specific feature such as a counterbored hole pattern.
Learned descriptors sit on top of either family. Rather than hand-designing the signature, a model is trained to place parts that engineers judge to be alike close together in the descriptor space, so similarity reflects engineering judgement rather than surface statistics alone.
Once every part has a descriptor, a query becomes a nearest neighbour lookup: compute the descriptor of the query model, find the closest entries in the index, return a shortlist. Most implementations then re-rank that shortlist with a slower and more exact comparison, because the cheap descriptor is good at ruling parts out and less reliable at ordering the few that remain.
All of this depends on reading geometry from every source system in the estate, which is why neutral formats matter. STEP, defined by ISO 10303, lets a descriptor be computed regardless of which CAD tool produced the model, so one index can span an estate that grew by acquisition. Reaching the vaults themselves is a separate problem, covered in purpose-built engineering integrations.
What Similar Has to Mean Before the Results Are Useful
Similarity is not one relationship. An engineer asking whether a part already exists is asking one of several different questions, and a tool that answers only one of them will feel broken when used for the others.
Geometrically identical, different part number. This is duplicate detection, and it is the highest confidence result a shape index can produce. The two models are the same solid, so the only open question is why the organisation is carrying both.
Same function, different size. A family of parts scaled across a range, where the engineer wants the whole family rather than the closest single member. Here strict scale invariance helps, because a descriptor that ignores absolute size will group the family together.
Same envelope, different internals. Two parts that occupy the same space and share mounting features but differ inside. This is the interchangeability question, and it needs scale sensitivity, because a part that is ten percent larger does not fit.
Same manufacturing character. Parts that share process class, wall thickness behaviour, and feature vocabulary, so that existing tooling, fixturing, or supplier setup carries across even though the shapes are not close.
Those four wants pull in opposite directions on one setting, which is why a single similarity slider is rarely enough. Scale invariance is correct for the second question and wrong for the third.
There is also a failure mode that geometry alone cannot catch. Two parts can be geometrically indistinguishable and completely non-interchangeable, because one is stainless and the other is aluminium, or one is qualified for a regulated programme and the other is not, or one is obsolete. Shape narrows the candidate set from a hundred thousand to a dozen. Material, finish, standard, and lifecycle status decide which of the dozen is usable, and a tool that cannot filter on those after ranking on geometry has done half the job. The downstream cost of getting this wrong is laid out in the real cost of duplicate parts.
Partial matching is the last piece. Often the reusable thing is not the part but a region of it, a bearing pocket or a connector cutout that has already been designed and validated. Whole-body descriptors miss these, because a small shared feature barely moves a global signature.
Where Geometric Search Changes the Engineering Workflow
The capability only pays off where a reuse decision is actually made. There are three such moments.
The first is before modelling starts. An engineer with a rough concept, a sketch, or a neighbouring part can ask whether anything close exists, at the point where the answer saves the design effort rather than merely documenting it. Asked an hour later, with the model half built, the same answer is much harder to act on.
The second is design review. A reviewer has no practical way to know whether a new part duplicates something from another programme. A geometric check turns that from an act of memory into a query, which matters most in the organisations where memory fails, the large ones with several sites and a long history.
The third is rationalization. Collapsing a sprawling part population into managed families requires knowing which parts cluster, and clustering by shape surfaces groupings no classification code records. That exercise is covered in engineering part reuse.
This is where Leo fits. Leo is an AI assistant built for mechanical engineers, trained on more than a million pages of standards, textbooks, and technical articles, and it connects to an organisation's existing knowledge base rather than replacing it. Leo offers integrations with leading PDM and PLM platforms, including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, and Arena PLM, so geometric search runs across the vault an organisation already has. It sits as an intelligence layer on top of that infrastructure, answers with citations back to the source document, and is SOC-2 certified and GDPR compliant, with no customer data used to train models.
One of our users put the practical effect plainly.
How to Evaluate Geometric Search Before You Trust It
Shape search demonstrates well. Any reasonable implementation looks impressive on a curated example, so a demo tells you little. These tests are cheap, take an afternoon on your own data, and separate a working capability from a convincing one.
Run a part you know has a near twin. Pick a case an experienced engineer can name from memory and confirm the system finds it. If it cannot recover a match you know exists, nothing else matters.
Run a part you know is unique. False positives are the more damaging error, because a plausible wrong match costs an engineer the time to open it, inspect it, and reject it. A system that always returns ten results tells you nothing by returning ten.
Test invariance directly. Take one model, translate it, rotate it, mirror it, and re-run each version. All four should return the original at the top. If rotation changes the answer, the descriptor is weak.
Test scale behaviour deliberately. Scale a part by a factor of two and check whether it is treated as the same part or a different one, then decide whether that answer is the one your reuse policy needs.
Test partial matching. Query with an assembly region or a single feature rather than a complete part, and see whether anything sensible comes back.
Check the filters. Confirm results can be narrowed by material, standard, supplier, and lifecycle status after ranking, not only before.
Check indexing scope and freshness. Ask which vaults, formats, and revision states are indexed, how long a newly released part takes to become findable, and what happens to parts never migrated cleanly.
Check where the geometry goes. Establish where models are processed, what is retained, and whether your geometry trains anything. For most engineering organisations this decides procurement, and a comparable evaluation framework is set out in how to evaluate an AI tool for PDM.
Run those eight against your own vault rather than a sample set. Geometric search is one of the few AI capabilities in engineering where the honest test is quick and the failure modes show up immediately.
FAQ
Search your vault by shape
Find the part you already have, without knowing what it was called
See how Leo runs CAD-aware geometric search across the PDM and PLM vaults you already have, so engineers find reusable parts before they model new ones.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
