Nvidia releases Auto Ontology, an open-source context layer for reliable agentic analytics

What Nvidia released

Nvidia has released Auto Ontology, an open-source Python framework that builds a semantic, graph-based reasoning layer over enterprise data, designed to make analytics agents reliable rather than merely fast. The project surfaced on the front page of Hacker News on 3 November 2026 and is available on GitHub under the Apache 2.0 licence. According to the project's documentation, a fast answer to a revenue question is worthless if the figure is wrong, and the framework exists to close that gap.

The problem it targets

Agents can already retrieve documents and run database queries. What they typically lack, the repository argues, is a context layer that explains where data lives, what it means, whether and how it should be used, where it originated and what resembles it. Auto Ontology constructs that layer from existing enterprise metadata, reconciles definitions that disagree across systems, and is explicitly positioned as a complement to the semantic models an organisation already owns rather than a replacement.

Five capabilities

The framework organises its functionality into five stages:

• Discover — generate enterprise ontologies from metadata, connect ontologies and semantic models that already exist, and assemble knowledge graphs intended for AI use.
• Ground — supply agents with trusted business definitions for metrics, entities and relationships, and reconcile inconsistent definitions across systems.
• Validate — check generated queries against certified definitions and lineage, and return provenance alongside every answer.
• Expose — serve context to any agent harness through MCP, a command-line interface and an API.
• Learn — fold certified corrections and usage patterns back into the ontology so shared context improves with each question.

What is under the hood

Auto Ontology is a full-stack application, not just a library. The web front end is built with Next.js 16 (React 19, TypeScript, Tailwind CSS 4) and uses Better Auth for authentication and role-based access control. The backend is a FastAPI service written in Python 3.12+, paired with LangChain for chat and natural-language-to-SQL translation, plus a separate ingestion worker that processes tabular data and writes embeddings. A Postgres database with pgvector stores application metadata, the catalog, the ontology and vectors, while HashiCorp Vault is an optional component for securing connection credentials.

Deployment is via a Helm chart on Kubernetes or Docker Compose for a local stack, and the tool can connect to Databricks, Postgres, Snowflake and DuckDB sources.

All inference runs through NVIDIA NIM endpoints, either Nvidia's hosted API catalog or self-hosted NIMs. The defaults pair a Nemotron nano chat model for reasoning with a Nemotron embedding model and a Nemotron-based reranker, and every model role can be overridden through environment variables that fall back to shared defaults. Authentication supports SSO, and scripts can mint API tokens that are stored only as SHA-256 hashes and displayed exactly once. For evaluation, the project points to a companion repository for running text-to-SQL benchmarks such as BIRD.

Caveats worth noting

Two limitations stand out. First, the project states plainly that it is not currently accepting external contributions, so open source here means Apache-licensed code rather than a community-driven effort. Second, the README warns that connection credentials added through the UI are stored unencrypted in Postgres unless Vault is explicitly configured, a deployment detail administrators should weigh before pointing the tool at production warehouses. The default reliance on NVIDIA NIM also means teams need an Nvidia API key or self-hosted NIM infrastructure to run the stack.

Why it matters

Text-to-SQL agents usually fail on semantics, not syntax: revenue can mean three different things in three different systems, and a fluent query against the wrong definition is still wrong. Auto Ontology is Nvidia's bet that the fix is a governed semantic layer of certified definitions, lineage and provenance sitting between agents and databases. Exposing that layer over MCP matters because any compliant agent harness can consume it without bespoke integration, and provenance returned with each answer gives auditors something to verify. The deeper consequence is strategic: by open-sourcing the framework while defaulting its inference to NIM, Nvidia extends its footprint from chips and models into the data tooling layer where enterprise AI adoption is actually decided.