What is NVIDIA Auto Ontology?
Auto Ontology builds a layer of business meaning over relational databases so that an agent asking for "revenue" uses the same columns, joins and formula every time. It reads schemas, profiles columns and samples values, then drafts three kinds of definition: terms (such as Customer or Invoice Line) mapped to columns, reviewed SQL attributes such as a margin, and relationships between tables, stored in Postgres with embeddings in pgvector. People refine and certify the draft in a web UI.
Agents then ask questions through a read-only MCP server, a REST API or the built-in chat. Each answer comes back with its SQL and the steps that produced it, and a check_answerable tool tells the agent when the data cannot answer a question. It connects to ten engines, including PostgreSQL, MySQL, Snowflake, Databricks, ClickHouse, Trino and DuckDB, and imports and exports models in the Apache Ossie format.
It runs as Docker Compose or a Helm chart: a Next.js front end with Better Auth login and roles, two FastAPI services and Postgres. Models default to NVIDIA NIM endpoints on build.nvidia.com, and any OpenAI-compatible endpoint can replace them. Apache-2.0; NVIDIA does not accept outside pull requests.
Is NVIDIA Auto Ontology worth it?
Auto Ontology targets the failure that matters in agent analytics: a query that runs and returns the wrong number because "revenue" meant something else. Its approach is a draft-then-certify loop. The model drafts terms, joins and SQL attributes from metadata and query history; people correct and certify them; the agent reuses certified SQL so the same measure is computed the same way each time.
The README shows the result on Microsoft's Wide World Importers sample: one compilation produced 31 terms, and a question about one product's June 2015 revenue in Pennsylvania returns $588.80 from 3 invoice lines with the six-table join it used. NVIDIA publishes a separate evaluation repository that runs BIRD, BEAVER, FDABench-Lite and Wide World Importers twice, once with only the schema and once through the ontology, scored by execution without an LLM judge. The README itself publishes no accuracy figures, so any gain has to be measured on your own data.
The catch is weight and trust boundaries. A local install is Postgres plus pgvector, a Next.js app, two FastAPI services, optional Vault and a seeded sample database. The README warns that ports 3001 and 3002 must stay private because the backend trusts a forwarded user header and the ingestion endpoints (ingest, delete, reset, compile) have no authentication. By default schemas, column samples and questions go to NVIDIA's hosted API, and turning on KumoRFM predictions uploads connected tables to Kumo's service.
It is young and closed to contributions. Public coverage began on 1 November 2026, the repository has 6,806 stars and 52 open issues and pull requests, the version reached 1.2.0 on 5 November with no GitHub releases, and integration requests (BigQuery, Microsoft Fabric) wait on NVIDIA's team. Against dbt MetricFlow or Cube, which define metrics in code you already review, Auto Ontology's advantage is generating the first draft and answering questions directly; its disadvantage is a second system to operate.
Where it earns its place
A data platform team at a 2,000-person retailer on Snowflake
Let analysts and agents ask revenue questions without each one inventing its own definition
Connect Snowflake, compile the ontology, have the finance analysts certify revenue and margin SQL attributes, then expose the read-only MCP server to Cursor and Claude Desktop. Inference runs on NVIDIA's API or self-hosted NIM.
An ML engineer evaluating text-to-SQL for an internal assistant
Measure whether a semantic layer improves accuracy on our own schema
Use NVIDIA's ontology-sql-eval repository to run the same questions with schema only and through the ontology, scored by execution, and read the list of questions it fixed and broke.
A regulated bank's analytics group
Keep database contents off third-party model APIs
Point every model role at self-hosted NIM or an internal OpenAI-compatible endpoint, store credentials in Vault, deploy with the Helm chart and leave KumoRFM disabled.
Before you commit
- Setup
- Docker with Compose v2, uv and an NVIDIA API key; copy .env.example, set passwords and keys, seed the sample database from auto-ontology-eval, run docker compose up, then turn on Semantic Compilation in Settings. Helm chart for Kubernetes.
- Integrations
- PostgreSQL, MySQL, Databricks, Snowflake, ClickHouse, Trino, Apache Kyuubi, HeavyDB, DuckDB and SQLite; MCP server (FastMCP) with tools such as check_answerable and ask_question; REST API with streaming answers and an OpenAPI spec; three agent skills; Apache Ossie YAML import and export.
- Data and privacy
- Catalog, ontology and app data stay in your Postgres. Model calls go to integrate.api.nvidia.com by default; KumoRFM, if enabled, receives up to 20 tables. Login and roles via Better Auth; optional Vault for credentials. Apache-2.0.
- Cost at scale
- Self-hosted services plus inference: the defaults use nemotron-3.5-lightning-30b-a3b for reasoning and small Nemotron embedding and rerank models, hosted by NVIDIA or on your own GPUs.
- Limits
- Ports 3001 and 3002 must stay on a private network; 52 open issues and pull requests; no external contributions accepted; models must support structured output and the embedding model cannot change between ingest and query.
What’s good
- Drafts terms, joins and SQL attributes from metadata, then people certify them
- Every answer returns its SQL and the steps behind it
- Read-only MCP server that runs with each user's permissions
- Ten database engines and Apache Ossie import and export
- Separate, judge-free benchmark harness on BIRD and others
What’s not
- Heavy stack: Postgres, three services, optional Vault
- Internal ports 3001 and 3002 unauthenticated by design
- Defaults send schemas and samples to NVIDIA's hosted API
- No accuracy numbers in the README; no outside pull requests
- Certification needs sustained human review
Best for
Data platform teams who want agents to answer from certified business definitions on their own infrastructure and can run a multi-service stack.
NVIDIA Auto Ontology alternatives
NVIDIA Auto Ontology: common questions
- Is NVIDIA Auto Ontology worth it?
- A serious, self-hosted answer to agents that write valid SQL with the wrong business meaning, provided you can run its whole stack and spend the time certifying what it drafts. We rate it 3.9 out of 5.
- How much does NVIDIA Auto Ontology cost?
- Free under Apache-2.0. You pay for model inference (NVIDIA's hosted API catalog by default, or your own NIM or OpenAI-compatible endpoints) and for hosting Postgres and three services.
- What are the downsides of NVIDIA Auto Ontology?
- Heavy stack: Postgres, three services, optional Vault; internal ports 3001 and 3002 unauthenticated by design; defaults send schemas and samples to NVIDIA's hosted API; no accuracy numbers in the README; no outside pull requests; certification needs sustained human review.
- Who is NVIDIA Auto Ontology best for?
- Data platform teams who want agents to answer from certified business definitions on their own infrastructure and can run a multi-service stack.
- How much does NVIDIA Auto Ontology cost for a team or in production?
- Self-hosted services plus inference: the defaults use nemotron-3.5-lightning-30b-a3b for reasoning and small Nemotron embedding and rerank models, hosted by NVIDIA or on your own GPUs.
- Is NVIDIA Auto Ontology safe for company data?
- Catalog, ontology and app data stay in your Postgres. Model calls go to integrate.api.nvidia.com by default; KumoRFM, if enabled, receives up to 20 tables. Login and roles via Better Auth; optional Vault for credentials. Apache-2.0.
- What are the best NVIDIA Auto Ontology alternatives?
- The closest alternatives we have reviewed are GraphRAG, Open Ontologies and LlamaIndex.
Answers checked Nov 5, 2026.