NVIDIA Open-Sources Auto Ontology: AI Queries Data 90% Faster, but the README Does Not Provide Accuracy

The boss asks the AI agent: How much revenue last quarter? Three seconds later, a clean number and a pretty chart. The problem is that it used the wrong revenue column. No one noticed.

NVIDIA's open-source Auto Ontology is aimed squarely at this scenario. It places itself between the AI agent and the database to provide a context layer: telling the agent where the data is, what each field means, and whether a given question should use a given field. The README says one thing very plainly: “A faster revenue metric is worthless if the number is wrong.”

How it works:

• Discover: reads database metadata, drafts an ontology, and maps business terms such as “active customer.”
• Ground: the agent uses these business terms instead of guessing from raw table names.
• Validate: checks generated SQL against certified definitions.
• Expose: provides context to agents through MCP, CLI, or API.
• Learn: writes certified corrections back into the ontology.

For users: ask a question, and get the answer, the SQL it ran, and the rows returned. You can inspect the process instead of just trusting the result. For builders: Apache 2.0, deployed with Docker Compose or Helm, and supports Databricks, Postgres, Snowflake, and DuckDB. It uses NVIDIA Nemotron models by default and also accepts any OpenAI-compatible endpoint. The MCP server has 9 read-only tools and deliberately has no tool for browsing raw tables; agents must go through business terms.

It sits alongside existing work. Cube and dbt MetricFlow are hand-written semantic layers; Wren AI is an open-source semantic layer for AI agents; Snowflake's Semantic View Autopilot also uses AI to draft semantic models but only serves Snowflake. Auto Ontology's difference is that it is open source, spans four databases, and is designed for agents via MCP.

More worth reading in the README are the not-so-pretty details:

• “90% faster” is NVIDIA's own test, with no published baseline.
• No accuracy score is published. NVIDIA includes an eval harness with BIRD's 1,534 questions, but there are no results.
• That harness was tested with Claude Opus 4.8, not Nemotron.
• Without HashiCorp Vault configured, database passwords are stored in plaintext in Postgres. The README itself says so.
• The project is still young; the materials mention about 750 GitHub stars, and NVIDIA does not accept external contributions.

The comments are more interesting than the launch post. Someone asks: without classes, hierarchies, logical constraints, and reasoning, is this really an ontology? Someone treats MCP as the answer; the author replies that MCP is the connection method and is not responsible for choosing the right column among your three revenue columns. Auto Ontology is on the other end of the connection, holding the business terms and certified SQL.

One commenter suggests a small step before deployment: make a glossary of term × table × business rule; revenue must distinguish net, deferred, and gross; then use known-answer validation queries to check whether the generated SQL joins to the correct fact table. Put the glossary and validation set in place first, then install Auto Ontology in front of the agent.

Another suggestion is to write field descriptions directly using the database's COMMENT ON COLUMN. The author replies that comments are the cheapest improvement, and any agent reading the schema gets them for free. But comments describe a single column. They cannot tell the agent whether “orders last quarter” should use dt or the order date. Expensive mistakes are often about which column a question should use.

Someone else suggests taking 20 known-answer questions from your own data and testing them both with and without a semantic layer. The author adds that you should also count the time the team spends certifying the ontology. After finance changes a definition once, run the same 20 questions again; that is where things most easily break. The lack of a published accuracy score also makes people want to test the eval harness first.

The bigger point: writing SQL itself is not that hard. The hard part is agreeing on what words mean. Auto Ontology puts this problem on the table and also exposes a new dependency: if the ontology is drafted wrong, the agent will consistently use the wrong columns, the results will look clean, and no one will notice. Before launch, a glossary and validation set may be worth more than one more agent tool.

GitHub: https://github.com/NVIDIA/auto-ontology
Eval harness: https://github.com/NVIDIA/auto-ontology-eval
MCP docs: https://github.com/NVIDIA/auto-ontology/blob/main/docs/mcp.md
NVIDIA page: https://developer.nvidia.com/topic/ai/agent-knowledge-foundation
MarkTechPost: https://www.marktechpost.com/2026/11/01/nvidia-auto-ontology-open-source-semantic-layer-ai-agents/