NVIDIA open-sources Auto Ontology to stop AI agents from fumbling SQL

Asking an AI agent to query enterprise metrics often leads straight to disaster. You ask for last quarter's regional sales, and the model writes a syntactically flawless SQL query that produces a completely fictitious number. The real problem is not the SQL syntax, but business logic: the database rarely has a neat column called regional sales, and the agent has no idea whether sales means booked, shipped, or invoiced revenue.

NVIDIA just released Auto Ontology, an Apache 2.0 open-source framework designed to give agents that missing business context.

Giving data a unified business brain

Auto Ontology does not try to reinvent SQL generation from scratch. Instead, it builds an automated semantic reasoning layer between enterprise data warehouses and AI agents.

Once connected to databases like Snowflake, Databricks, Postgres, or DuckDB, the framework scans the metadata, extracts entities, and reconstructs relationships into a knowledge graph. It translates vague business terms like active customers or gross margin into verified queries aligned with company definitions.

When someone asks a question in plain English, the system checks whether the existing schema actually supports the calculation. If a user asks for return on ad spend but the table lacks marketing spend records, the system flags the missing dependency instead of hallucinating an answer. Every response returns deterministic SQL accompanied by full data lineage and provenance.

It also closes the feedback loop. When domain experts correct a generated query or tweak a metric definition in the web interface, the system absorbs those certified changes back into the shared ontology.

Built for modern agent workflows

On the engineering side, NVIDIA delivered a production-ready stack. The frontend is built on Next.js 16 and Better Auth, the backend runs on Python 3.12 with FastAPI and LangChain, and the metadata and vector embeddings live inside PostgreSQL with pgvector.

For inference, the framework defaults to NVIDIA NIM microservices, routing chat and query logic to models like Nemotron-3. It also supports HashiCorp Vault for teams that refuse to keep database credentials in plaintext configuration files.

Crucially, Auto Ontology exposes its context layer not just through standard APIs, but via the Model Context Protocol. That allows developers to hook their enterprise databases directly into environments like Cursor and Claude Desktop. Note that NVIDIA currently flags the repository as not accepting external contributions.

Why it matters

We think NVIDIA is solving the exact friction point that keeps agentic analytics stuck in demo mode. The industry has spent two years improving raw text-to-SQL benchmarks, yet enterprise rollouts still fail because models cannot guess internal jargon.

For enterprise teams, this frees data analysts from functioning as human glossaries. Instead of answering why two internal dashboards show conflicting revenue numbers every Monday, organizations can codify metric logic once into an auditable layer that sales, finance, and operational agents can all query safely.

For NVIDIA, this is a calculated strategic move. While competitors sell generalized LLM APIs, NVIDIA is packaging the boring, unsexy enterprise plumbing that makes those models actually work in production. By bundling the workflow with NIM endpoints and Nemotron models, it creates a sticky pipeline that pulls enterprise inference straight into its own compute ecosystem.

Keep in mind that this is not magic fairy dust for messy data. If your company's data lake is already a swamp of undocumented tables and duplicate schemas, the initial ontology will still require substantial human verification before any agent can be trusted with mission-critical numbers.