AI for Engineering Knowledge Management
What an AI really stores when it indexes your PDM vault: encrypted vectors, not files. Learn what stays put, what to ask vendors, and how Leo handles it.
·
⏱
8 min read
Dr. Maor Farid
Maor Farid is the Co-Founder and CEO of Leo AI, the first AI platform purpose-built for mechanical engineers. He holds a PhD in Mechanical Engineering and completed postdoctoral research at MIT as a Fulbright fellow. A Forbes 30 Under 30 honoree and former AI researcher and Mechanical Engineer in an elite military intelligence, Maor leads Leo AI's mission to transform how engineering teams design better products faster.
BOTTOM LINE
An AI index is a finder, not a second vault. The safest designs store only encrypted vectors that cannot be turned back into your files, leave the originals in your PDM, PLM or directories, and keep customer content out of model training. Ask each vendor to state in writing what is stored, whether it can be reversed, where it lives and whether it trains on your data. Clear one-sentence answers are a good sign. Vague ones are worth a follow-up before anything reads your drawings.
Every engineering team that considers an AI assistant for its PDM hits the same question before any demo gets scheduled: what does the AI actually keep? The honest answer is rarely in the marketing copy, and the question is a fair one. A vault holds years of released designs, supplier drawings, test reports and the decisions behind them. Nobody wants a second copy of that sitting somewhere they cannot audit.
The word that causes most of the confusion is indexing. To a security reviewer it can sound like a copy, a mirror or an export. To a search engineer it means something narrower. This article separates the two readings, walks through what an index of a vault really contains, and lists the questions worth asking in writing before any AI product reads your files.
The short version: a well-built index holds a mathematical representation of your content, not the content. The CAD stays in the vault, and the vault stays the system of record.
What Indexing Actually Produces
When a search system indexes a document, it reads the document once and builds a lookup structure that makes later questions fast. A classic keyword index stores words and the places they appear. A modern AI index does something similar for meaning: it converts a file or a passage into a long list of numbers, called a vector, so that items with similar content end up with similar numbers.
The practical effect is that an engineer can ask for "a bracket that clamps a 12 mm tube" and get back relevant parts even when no file contains those exact words. The index answers the question of which items look relevant. It then points back to the file in the vault so the engineer can open the real thing.
That pointer is the important design detail. The index is a finder, and the vault remains the place where the model, the drawing and the revision history live. Teams that already struggle to locate parts will recognise the pattern from the work described in our guide to AI search across Windchill, Teamcenter and Vault: the value comes from finding the right record quickly, not from relocating the records.
IN PRACTICE
It integrates directly with PLM and existing workflows, making past designs, standards, and calculations instantly available. The result is fewer errors, faster decision-making, and a more consistent process across teams.
- Sergey G., Board Member
What Stays in the Vault
In a layered setup, nothing about the system of record changes. Part files, assemblies, drawings, revision history, lifecycle states, check-in and check-out rules and the audit trail all remain in the PDM or PLM, or in the Windows or network directory where the files already sit. The AI layer sits on top of those systems. It does not replace them and it does not take over their job.
This matters for three reasons that engineering managers tend to care about:
Single source of truth. When the released revision lives in one place, nobody has to ask whether the AI is looking at an older copy.
Change control. Your vault already enforces who can release a drawing and when. An AI product that holds its own copy of the files sits outside that discipline.
Exit options. If a team decides to stop using an AI product, a layered design means the data was never moved, so there is nothing to retrieve, migrate or delete from the vendor's side beyond the index itself.
Leaving the files where they are also keeps the project honest about scope. The AI layer reads what it is permitted to read and builds an index. It is not a migration project, and it should not need a freeze on engineering work to deploy. For a broader view of how that layering looks across systems, see our piece on AI for PDM and PLM integration.
What an Index of Vectors Holds, and What It Cannot Rebuild
A vector is a fixed-length list of numbers produced by a model that has read a piece of content. It captures relationships such as which parts resemble each other or which paragraph answers which question. It does not store the geometry of your part or the text of your paragraph in a form that can be read back out.
Think of a library catalogue card. The card tells you the subject, the shelf and a few descriptors, and it is enough to find the book. Nobody can reprint the book from the card. A vector index plays the role of the catalogue, with a far richer description than a card, but the same one-way relationship to the original.
The security-relevant question is whether the process is reversible. If vectors are produced with a method designed so they cannot be reverse-engineered, then holding the index does not mean holding the file. If a vendor cannot say whether its index can be turned back into the source content, treat that as an open risk and ask for the answer in writing. Encryption adds a second layer: data should be protected in transit and at rest, so that the index itself is not readable by anyone who gains access to the storage.
There is one more thing worth checking: whether the answers that come back are tied to a source you can open. A citation that points to a file in your own vault is evidence that the answer was built from your records and that the record still lives where it did.
Questions to Ask Before an AI Reads Your Drawings
A security review for an AI product does not need to be long, but it does need to be specific. These are the questions that separate a clear answer from a vague one:
What exactly is stored on your side? Ask for a plain description: original files, extracted text, thumbnails, vectors, or some combination.
Can the stored index be converted back into the original content? The answer should be a direct no, with a reason.
Where is it stored, and is it encrypted in transit and at rest?
Is customer content used to train or improve any model? Get this in the contract or the security documentation, not only on a web page. Our article on why training data matters for engineering AI explains why the source of a model's knowledge deserves the same scrutiny as its storage.
Which independent attestations exist? A third-party audit report is stronger evidence than a self-declared statement. We cover how to read these in SOC-2 for engineering AI.
Is there a way to remove the index and confirm it is gone when the relationship ends?
Vendors that have done the work answer these in a sentence each. Vendors that have not tend to answer with a longer paragraph that avoids the noun. Treat the length and clarity of the answer as information.
How Leo Handles This
Leo is an AI assistant for mechanical engineers that connects to an organisation's knowledge base, including PDM, PLM, local and network directories. Leo offers integrations with leading PDM and PLM platforms such as SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter and Arena PLM, and others.
On the storage question, Leo's published security documentation is direct. Leo uses a patented method to index CAD and text files by converting them into encrypted vectors, and it stores only those vectors, never the original files. The documentation states that the vectors cannot be reverse-engineered and that CAD and text files cannot be reconstructed from them. Your data remains where it is today, in your PLM, PDM or Windows directories.
Leo does not use customer CAD files or documents to train its AI model. The model is trained on trusted engineering sources, not on a company's proprietary data. Leo is SOC-2 certified and GDPR compliant, and verified security documentation is available on request under a non-disclosure agreement.
That makes Leo an intelligence layer on top of the vault rather than a replacement for it. The engineering team keeps its PDM, its release process and its audit trail, and gains a faster way to find what is already inside. Teams evaluating the broader category can also read our overview of engineering knowledge base software.
FAQ
Find parts without moving your CAD
Leo indexes your vault and leaves every file where it is.
See how Leo connects to your PDM or PLM, stores only encrypted vectors and keeps your original CAD files in the systems you already trust. Book a demo to walk through it.
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
