# How to add AI to Autodesk Vault on-prem using the Vault API

> Source: https://netray.co/erp-help/private-ai-layer-autodesk-vault-on-prem-architecture
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: ERP Help

## Short answer

Use the Vault web services API to read files, properties, items and change orders as the asking user, extract document text through the Job Processor or a service, and serve an open-weight model on your own hardware. Vault's roles, folder access control and lifecycle state security stay in force because every read goes through the server.

## Kind

advanced

## System

on-prem-apps

## Erp

Autodesk Vault Professional

## Module

Private AI layer / Vault API / ADMS

## Query

how to add AI to Autodesk Vault on-prem

## Alternate Queries

- Autodesk Vault AI integration
- Autodesk Vault API search with LLM
- chat with Autodesk Vault data
- Autodesk Vault private AI assistant

## Applies To

Autodesk Vault Professional (and Workgroup features where available), a recent release on Microsoft SQL Server, Autodesk Data Management Server (ADMS) on premises.

## Steps

### Reference build for a private AI layer on Vault

**Items:**

- Define the question classes: file and item lookup by property, where-used and references, lifecycle state and revision, change order status, content Q and A over PDF and document files.
- Create a tool server that uses the Vault web services client library to call services such as DocumentService, ItemService, PropertyService and ChangeOrderService with named, bounded actions.
- Use Vault Job Processor (custom jobs are supported) or a scheduled worker to extract text from newly checked-in and released files into a local index.
- Record folder path, lifecycle state, category and revision with each chunk so retrieval can filter on released data and on the caller's access.
- Serve a local open-weight model with vLLM and local embeddings. Keep the model host off the internet.
- Authenticate calls as the user with Vault credentials or Windows authentication, so Vault evaluates roles and ACLs. Log tool calls and object ids.
- Roll out read-only search first, then drafted summaries, then supported actions such as creating a change order draft after approval.

## Sections

### How data is reached in Vault

**Body:**

Vault stores metadata in a SQL Server database managed by ADMS, with file content in a filestore on the server. The database layout is internal and Autodesk does not support direct reads or writes, so use the API. The Vault web services SDK exposes services for documents and folders, items and BOMs, properties, lifecycles and change orders, in .NET libraries.

Property definitions are central. File and item properties map to database values and to CAD properties, and the property service tells you which definitions exist. Search is done with search conditions against property definitions, which suits named tool actions such as 'find latest released files where Customer equals X'.

**Code:**

// C# sketch, Vault web services (confirm signatures against the SDK for your release)
var creds = new UserPasswordCredentials("vaultserver", "EngVault", userName, password);
using (var mgr = new WebServiceManager(creds))
{
    // Build SrchCond[] using PropertyService property definitions, then:
    // mgr.DocumentService.FindFilesBySearchConditions(conds, null, folderIds,
    //     true, true, ref bookmark, out status);
}

### Reference architecture

**Body:**

Put the AI layer beside ADMS: a tool server using the Vault client libraries, a Job Processor extension or worker for extraction, a vector index, a GPU host and an audit store. Extract on check-in and on lifecycle change so the index follows the vault. The Job Processor already provides queueing and retry, so reuse it instead of building a scheduler.

For analytics (counts of files by state, release cadence), use API-built exports rather than SQL against the vault database. For air-gapped installs the whole stack, including model weights, ships by approved media and runs inside the enclave.

### Permissions and write-back

**Body:**

Vault controls access with roles, access control lists on folders and files, and lifecycle state security. When the tool server logs in as the user, the server returns only what that user may read. For the extraction worker, which runs under a broader identity, store the folder path and state on each chunk and filter at query time, then verify with an API read before quoting.

Supported write operations include changing lifecycle state and creating or updating change orders. Keep each behind explicit approval: show the user the target object, current and new state, and the reason, then run it under their identity and log the result. Never let the assistant promote states unattended.

### Sizing, limits and honest trade-offs

**Body:**

A group of 20 to 100 users with a vault of a few hundred thousand files typically fits one 24 to 48 GB GPU with a 14B to 32B class model, plus a CPU server for extraction. The index is commonly single to low tens of GB. Larger firms with several vaults scale by adding a worker per vault and a shared model pool.

Limits: Inventor and AutoCAD geometry is not language-model readable, so the assistant relies on properties, BOMs and exported documents. If the question is just 'find files by property', Vault's own search is faster and cheaper. Cloud customers on Autodesk's hosted platform follow a different path from this on-prem design.

## Pitfalls

- Reading the SQL Server database directly. It is unsupported and ignores Vault permissions.
- Running extraction as an administrator and then returning chunks without a state and folder filter.
- Treating a file's latest version as the released one. Lifecycle state must travel with the text.
- Letting the model compose search conditions freely. Offer named actions with validated property names.
- Skipping lifecycle and revision in answers, which makes quotations of old content look current.

## Ai Angle

A grounded assistant on Vault answers 'which released drawings reference this part and what changed in the last change order' by calling the property, reference and change order services as the user and returning links to the objects. Netray builds it as a private, read-only first layer with local models and an audit trail.

## Faq

### Can the Job Processor do the text extraction?

**Answer:** It can host custom jobs, so it is a good place to trigger extraction on check-in or state change. Many teams still run the heavy parsing in a separate worker the job calls, to keep the Job Processor stable and easy to restart.

### Does this work with Vault on Autodesk's cloud?

**Answer:** Not this design. The on-prem pattern assumes you run ADMS and SQL Server yourself. Cloud-hosted offerings use different APIs and a different data boundary, so assess them separately.

### How are Vault roles respected?

**Answer:** By calling the Vault server as the user for anything shown to that user. For indexed text, each chunk carries folder, state and category, and results are re-checked through the API. Unmapped rules lead to exclusion from the index.

### What is the first useful deliverable?

**Answer:** Property-aware search plus where-used and change order status in natural language, over released data only. It is read-only, quick to validate against known answers, and builds trust before drafts or actions.

## Search terms

Autodesk Vault AI, Autodesk Vault API, Vault Professional on-prem AI, chat with Autodesk Vault, Vault Job Processor extraction, Vault web services search, Autodesk Vault LLM, Vault RAG

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json