Nvidia has released Auto Ontology, an open-source Python framework designed to give AI agents a structured understanding of enterprise data before they generate queries or answers. The project builds a semantic, graph-based layer over existing metadata, aiming to make data retrieval, query generation and the reasoning behind an agent's output easier to inspect.

The core problem identified by the project is reliability. An agent may be able to locate documents or connect to a database, but that does not tell it which table defines a business metric, whether two systems use the same definition, or where a figure originated. Auto Ontology is intended to reconcile those definitions and relationships rather than replace semantic models an organization already uses.

Nvidia says the resulting context can be made available to different agent environments through Model Context Protocol, a command-line interface and an API. That makes the framework potentially useful beyond a single chat interface: coding agents and other MCP-compatible software can query the same mapped layer, while service integrations can use authenticated API calls. The repository also points users to a separate evaluation project for text-to-SQL benchmarks such as BIRD.

Deployment is aimed at technical teams. The repository includes a backend, an ingestion service and a front end, with containerized setup using PostgreSQL and pgAdmin. It also documents Kubernetes deployment. Inference uses Nvidia NIM endpoints, which can be hosted through Nvidia's catalog or run by the user. Model and embedding roles are configurable through environment settings.

The project supports database connections created through its interface or supplied at deployment. Nvidia's documentation flags an important security detail: credentials entered through the interface are stored in plaintext in PostgreSQL unless a supported Vault configuration is enabled. The repository also provides single sign-on support and API tokens for scripts, notebooks and scheduled jobs.

Conversation history is scoped to an authenticated user, according to the documentation. The system can use a limited number of prior completed turns to rewrite contextual follow-up questions into standalone requests before sending them through its text-to-SQL workflow. Raw result sets, errors and internal agent thoughts are excluded from that reused context.

Auto Ontology is distributed under the Apache License 2.0. Nvidia says it is not currently accepting external contributions, although the source and third-party notices are public. The release gives enterprise AI teams a new building block for a persistent data context layer, but deploying it safely still requires careful handling of credentials, identities and the external material the software accesses.