AI for Engineering Knowledge Management
Autodesk Vault search relies on tokens and properties, so engineers still cannot find files. Here is why, and how AI search fixes it.
·
⏱
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
Autodesk Vault search is doing exactly what it was designed to do: match your query against indexed tokens and properties. The trouble is that engineering recall is about shape, intent, and prior decisions, none of which live cleanly in a property field. Tightening naming and metadata helps at the margins, but it cannot teach a text index to understand geometry. An AI intelligence layer that adds natural language and geometric search on top of Vault closes that gap, so the part you know exists becomes the part you can actually find.
You vaulted the part. You know it exists. You designed something almost identical eighteen months ago. So why does Autodesk Vault search return forty unrelated results, or worse, nothing at all? The problem is not your memory or your filing discipline. It is the way Vault search actually works under the hood. Vault indexes file names and properties into tokens, and that token model quietly decides what you can and cannot find. This article breaks down the mechanics of Vault search, where it falls short for real mechanical engineering work, and what an AI intelligence layer changes about the experience. If the symptom sounds familiar from other systems, it should, and we trace the same pattern in our piece on why engineers still cannot find anything in SolidWorks PDM.
How Autodesk Vault search actually works
Vault is a capable data management system, and understanding its search model is the first step to working around its limits. When a file is checked in, Vault records the file plus its properties (name, size, dates, custom fields) and indexes those values for fast retrieval. Search then matches your query against those indexed tokens. A token is an individual chunk of a property value. Vault breaks values into tokens at recognized separators so it can match pieces of a string, but most punctuation is not preserved as searchable text.
This has a direct consequence that surprises many engineers. If your naming convention uses hyphens, periods, or other punctuation (for example, a part number like 4471.02.A), the punctuation is generally not part of the token set, so a search on the full string may not behave the way you expect. Quick Search and Advanced Find also combine multiple criteria with an AND operator by default rather than OR, which narrows results fast and can return an empty set when you actually wanted any of several matches.
There is a second layer most teams never turn on. Searching the contents inside a file, rather than just its properties, requires the Content Indexing Service, which is enabled separately through the Vault server console and is not a default best practice. Without it, a search reaches the metadata an admin chose to expose, and nothing more.
n practice, that means a vault holding decades of valuable design work can behave like a filing cabinet that only answers to the exact label on each drawer. The data is all there, fully revision controlled and audited, yet the path to it runs entirely through whatever text someone remembered to type at check-in.
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. It was made by people who truly understand the struggles of mechanical engineers.
Eytan S., R&D Engineer, Mid-Market (G2 Review)
Why engineers still cannot find what they need
The token and property model assumes you can describe a file the way it was filed. Real engineering recall does not work that way. You remember a bracket geometry, a wall thickness, a vendor you used once, or the project a part shipped on. You rarely remember the exact property string a colleague typed three years ago. When the words in your head do not match the tokens in the index, the search comes back empty and you start scrolling folder trees.
Three failure modes show up again and again:
Vocabulary mismatch. The original engineer called it a standoff, you search for a spacer, and Vault treats the two terms as unrelated tokens.
Inconsistent metadata. Custom properties are only as good as the discipline behind them, and across years and engineers that discipline drifts.
No shape awareness. Vault search has no concept of geometry. You cannot ask it for a part that looks like the one on your screen, because the index holds text, not form.
The cost is not abstract. A 2022 survey of more than 100,000 engineers and designers found that nearly half spent at least an hour every day searching for parts, and CADENAS PartSolutions has reported that engineers lose close to two hours a day sourcing or redrawing components that already exist. When search fails, the rational move is to redraw, which is exactly how duplicate parts multiply in the vault. We cover the downstream damage in detail in our look at the real cost of bad PDM search.
Property-based search versus the way engineers think
It helps to separate two distinct problems that Vault search blends together: data management and knowledge retrieval. Vault is excellent at the first. It controls revisions, enforces check-in and check-out, and keeps a clean audit trail. That is its job, and it does it well. PLM and PDM systems were built to govern data, not to answer questions, a distinction we unpack in our comparison of PLM versus PDM.
Knowledge retrieval is a different task. It means surfacing the right prior decision, the right standard part, or the right past design from a body of work that may span decades and several file formats. Interoperability standards help move geometry between tools. ISO 10303, known as STEP, is the dominant standard for product model data exchange and now spans several hundred parts, with AP242 deployed across major CAD vendors. STEP solves moving a model from one system to another. It does not solve finding the right model in the first place.
Autodesk has acknowledged the gap. The Autodesk Assistant tech preview in Vault 2027 adds natural language querying so engineers can describe what they want instead of memorizing property names. That direction validates the core point: text and property search alone do not match how engineers actually recall their work.
How Leo adds an AI intelligence layer on top of Vault
Leo is an AI intelligence layer that sits on top of your existing data management, not a replacement for it. You keep Vault as your source of truth for revisions and governance. Leo connects to PDM, PLM, local and network directories, and ERP, then adds natural language and geometric search across your full engineering history: CAD files, specifications, and past decisions. Integrations are available for Autodesk Vault, SolidWorks PDM, PTC Windchill, Siemens Teamcenter, Arena PLM, and other systems.
Two capabilities change the day to day experience. First, natural language search lets you ask for a part the way you would describe it to a colleague, so a spacer and a standoff can both surface the same component without a token match. Second, geometric search lets you find parts by shape, so you can point at the model on your screen and ask what already looks like it. Leo prioritizes parts you already designed or bought, plus more than 120 million vendor options, before it ever generates new geometry. That ordering matters, because engineers spend roughly 35 percent of their time designing parts that already exist, and finding the right existing part can cut reported BOM costs by around 15 percent.
Leo is trained on more than one million pages of engineering standards, books, and articles, so it understands engineering language rather than treating your query as a string match. On security, Leo is SOC 2 certified and GDPR compliant, no AI is trained on your data, and your intellectual property is never shared. For teams weighing how AI layers onto an existing stack, our enterprise integration guide walks through the practical steps.
Practical steps to fix Vault search today
You do not have to wait to get more out of search. A few measures reduce the friction inside Vault itself, and an AI layer handles the cases configuration cannot reach.
Standardize naming and avoid relying on punctuation to carry meaning, since hyphens and similar characters are not reliably tokenized.
Decide deliberately whether to enable the Content Indexing Service, weighing full text search against the indexing overhead it adds.
Curate a small, enforced set of custom properties rather than many inconsistent ones, so the tokens that exist are trustworthy.
Build search folders for recurring queries so common lookups are repeatable instead of reinvented.
Add an AI search layer for the recall problems that metadata cannot solve, namely vocabulary mismatch and shape based discovery.
The first four steps make the index cleaner. They do not give Vault an understanding of geometry or engineering intent. That is the structural gap, and it is the same one teams hit with property based search in any PDM system, as we discuss in our piece on choosing PDM software for mechanical engineers.
FAQ
Autodesk Help and Knowledge Network, Searching Vault and Content Indexing documentation, supports how Vault tokenizes properties and that content indexing is enabled separately.
Autodesk Design and Manufacturing blog, Autodesk Assistant Tech Preview in Vault 2027, supports the addition of natural language querying to Vault.
NIST, Introduction to ISO 10303 (the STEP standard for product data exchange), supports the role and scope of STEP for product model data.
CADENAS PartSolutions 2022 survey of 100,000+ engineers, supports time lost per day searching for parts.
Find the part you know exists
Add AI search on top of Autodesk Vault, no rip and replace.
Leo adds natural language and geometric search across your full engineering history, on top of the Vault you already run. Book a demo.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
