How-to
Semantic layer for business data: 4-hour plan
A semantic layer for business data gives your agent one agreed definition per field, and NVIDIA's open source Auto Ontology is one way to build it. The install is the easy half. This is the four hour plan for one question on one source, what the docs actually cover, and the decision table for whether to wait.
Scattered data is a meaning problem, not a storage one. The CRM calls it a contact, the spreadsheet calls it an account, and an agent asked for active customers has to guess which one you meant. Auto Ontology is NVIDIA's open source attempt at fixing that.
What a semantic layer for business data actually does
It stores the meaning, not the rows. Instead of an agent writing a join and hoping, it reads a published definition: this is what we call an active customer, this is the attribute it maps to, this is the relationship to a quote. That is also why it is slower to adopt than it looks, which is the same lesson as bad PDF ingestion making a document chatbot give wrong answers. Garbage meaning in, confident nonsense out.
The four hour plan
The times below are my estimates from the documentation. I have not stood this up in production.
Read the project's AGENTS.md and the shape of the work is already there: a setup skill covering Helm, local Compose, or MCP against an already running instance, a management skill for glossary terms, SQL attributes, semantic relationships, model YAML import and export and governed result definitions, and a query skill. Three jobs, not one.
- Pick the one question (about 20 minutes). Something your team argues about weekly. Write down the number you expect.
- Stand up an instance with local Compose (about 50 minutes). On your own machine first, the way you would self-host team AI chat in about three hours.
- Publish about ten definitions from one system (about 90 minutes). Glossary terms and attributes only. This is the step that decides the project.
- Wire one agent to it through MCP and ask the question (about 40 minutes). Read only, nothing writes back, the same containment rule as before an agent touches your CRM.
- Write the failure log (about 40 minutes). Every wrong answer, and which missing definition caused it.
Total: about four hours. One question, one source, one agent.
Should you build a semantic layer for business data now?
| Your situation | What to do this month |
|---|---|
| One number your team argues about every week | Run the four hours on that number only |
| Several systems, nobody owns the definitions | Settle ownership first, or the layer encodes the argument |
| Metadata is a monthly manual export | Wait, start with one clean export |
How much work is a semantic layer to maintain
Arithmetic on your own inputs, not a promise. If two people spend 30 minutes a week reconciling one recurring number, that is about 52 hours a year. Four hours of setup plus, my guess, an hour a month of upkeep is about 16 hours. Whether the layer removes those 30 minutes is exactly what step 5 tells you. If you cannot measure it, give the agent an audit trail of what it actually did and read it.
Limitations
The repo has no formal releases listed, and the API site shows version 1.2.0, so treat it as moving code and pin what you install. I could not verify from primary sources that it connects to common CRM or spreadsheet tools out of the box, that it is aimed at small businesses, or that a weekend is enough for anything production ready. It looks like an engineering project, not an app. Community enthusiasm around the repo is sentiment, not evidence that it works on your data.
My take
The install is the easy half, and almost every write-up stops there. The part that pays is governance: deciding, once, what your words mean, and publishing that so an agent cannot invent its own version. Spend the four hours on one question. If the modelling step feels like a fight about definitions rather than a config task, that is the real finding, and it is worth four hours to learn.
Questions people ask
What is a semantic layer for business data?
A semantic layer stores the agreed meaning of your data: what a field means, what it maps to, and how it relates to other things. You need one when the same question gets different answers from different systems, and an agent has to guess which definition you meant.
Is NVIDIA Auto Ontology free to use?
It is an open source Python framework published on GitHub, so there is no licence fee to try it. The cost is your time: setup, data modelling and ongoing maintenance of the definitions you publish.
Can an AI agent query it directly?
Yes. The repository says the semantic layer is exposed to agent harnesses through MCP, a CLI and an API, and the project ships a separate query skill for that. Keep the first connection read only.
Have a process like this leaking time or leads?Thirty minutes, no deck, and you leave with a plan.
Book a call