NVIDIA Put Metric Definitions in Front of the Agent. That Is the Hard Part.

Ask three teams in the same company what revenue means and you will get three definitions that all sound reasonable and do not match. Finance counts bookings. Sales counts contracted value. The dashboard counts recognised revenue, net of refunds, on a calendar the finance team changed last quarter.
An agent that can write SQL in two seconds does not fix this. It makes it faster.
NVIDIA's Auto Ontology is an open library aimed at that specific failure. It builds an ontology from enterprise metadata, reconciles conflicting definitions across systems, and serves the result to any agent harness through MCP, CLI and API. Apache 2.0, roughly 8,900 stars, and a README that opens with the argument rather than the architecture: a revenue metric answered faster is worth nothing if the number is wrong.
Four verbs, one loop
The library is organised around what it does to context.
Discover: generate an ontology from metadata, or connect an existing semantic model. Ground: attach trusted business meaning to metrics, entities and relationships, and reconcile the definitions that disagree. Validate: check generated queries against certified definitions and lineage, and return provenance with every answer. Expose: serve that context through MCP, CLI and API.
There is a fifth, less advertised: Learn. Corrections and usage get folded back into the ontology, so the shared context improves as people use it.
That last step is what separates this from a data catalog. A catalog is a document that goes stale. A layer that ingests corrections is a system that might stay accurate, if anyone bothers to correct it.
What gets deployed
The stack is heavier than a CLI tool and the README is upfront about it.
A Next.js 16 frontend with Better Auth and Prisma on port 3000 handles the UI, authentication and the public API gateway. A FastAPI backend on 3001 runs the text-to-SQL agent and the catalog and ontology APIs. An ingestion worker on 3002 ingests tabular data and writes embeddings. A FastMCP server on 3003 exposes read-only tools. Postgres 17 with pgvector on 5432 holds the catalog, the ontology, application data and the vector store. HashiCorp Vault is optional, for connection credentials.
Inference runs through NVIDIA NIM, either the hosted API catalog or self-hosted endpoints. Deployment is a Helm chart for Kubernetes or Docker Compose for a local stack. There is a BIRD text-to-SQL evaluation path in a separate repository.
The MCP surface is the part most teams will actually touch. Nine read-only tools, split by job: `check_readiness` tells you whether the deployment can answer anything at all, `search_terms` and `get_term` and `describe_table` handle lookup, `check_answerable` asks whether the ontology covers a question before the agent wastes a call, and `ask_question` returns the answer, the SQL and the rows together.
That last triad is the design opinion. Returning SQL alongside the answer means a human can audit the query, and returning the rows means the agent can be checked against what it claimed. An agent that returns only a number is unfalsifiable.
Wiring it into a client is two lines. Point `AUTO_ONTOLOGY_API_URL` at your deployment and run `uvx --from "git+https://github.com/NVIDIA/auto-ontology.git#subdirectory=mcp" auto-ontology-mcp`. Cursor gets `{"url": "http://localhost:3003/mcp"}` in `mcp.json`. Claude Desktop bridges with `mcp-remote`.
The security notes are the interesting part
The README contains a paragraph that most open-source projects would leave out.
Keep ports 3001 and 3002 on a private network. The FastAPI backend trusts the user identity the frontend forwards in the `x-auto-ontology-user-id` header. The ingestion service's endpoints, which include ingest, delete, reset and compile, have no authentication of their own. Clients are supposed to go through the web app on port 3000.
Read that as an architecture description and a warning in the same breath. Header-based identity trust is a normal pattern behind a reverse proxy and a catastrophic one if the backend is reachable. An ingestion endpoint with no auth that can delete and reset is a loaded gun pointed at your catalog.
Demo-grade deployments documented this honestly tend to carry exactly this shape, and NVIDIA is not the first team to ship one. The value is that the project says it out loud instead of letting you find out.
What open does not mean here
The project is Apache 2.0 and explicitly not accepting external contributions. The team maintains the code, takes integration requests and bug reports as issues, and routes security reports through a separate process.
That is a deliberate posture that shows up more often now: publish the code, keep the roadmap. For a company building an analytics context layer, the ontology schema and the agent skills are the interface they want to control, and accepting patches to either would slow the part that has to stay coherent.
The practical consequence for a team evaluating it: you can read every line and run it anywhere, and you cannot get a fix merged. Plan for forking.
One more detail worth flagging. Auto Ontology is compatible with Apache Ossie, an open standard for exchanging semantic models, with import and export through the UI or `POST /api/model/import` and `/api/model/export`. If your organisation already maintains semantic models in another tool, that converter is the difference between a migration project and an afternoon.
There is also an optional predictive path through KumoRFM, a hosted service. When enabled, the agent uploads the connected tables to it, capped at 20 tables by default and unlimited rows unless you set a limit. The README tells you to review what that sends before turning it on. Take that seriously; it is the one feature in the stack whose data leaves your network.
Why this category exists now
Agentic analytics has a bad year behind it. Agents that can query a warehouse are easy to build. Agents that return the number a CFO would sign off on are not, because the failure mode is silent: a plausible number, from a plausible query, against a definition nobody checked.
The usual fix has been prompt engineering, which is to say asking the model to be careful. The durable fix is to hand the model a certified dictionary and a validator, and refuse to answer when the question falls outside it. `check_answerable` exists precisely so the agent can decline.
Whether this becomes infrastructure or a reference implementation depends on adoption, and adoption depends on how many companies have a semantic model clean enough to import. That number is smaller than the number of companies that think they do.
Still, the direction is right. The expensive part of analytics was never the SQL. It was agreeing on what the columns mean, and remembering the agreement six months later. Putting that agreement inside the agent's context, with lineage attached to every answer, is the least glamorous and most load-bearing thing anyone shipped this week.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.