NVIDIA Adds Open-Source Auto Ontology to AI Agent Toolkit

During its GTC Berlin announcements on 21 October 2026, NVIDIA introduced Auto Ontology, an open-source library for mapping enterprise data to business concepts and relationships. According to NVIDIA's official announcement, the library is part of its new Agent Knowledge Foundation within the expanded Agent Toolkit, alongside CUDA-X libraries and models for information retrieval and structured-data analysis.

The project addresses a problem in AI-assisted analytics: a database query can execute successfully while returning an incorrect answer because the underlying business definitions are inconsistent or misunderstood. Auto Ontology is designed to provide a semantic layer that connects business terms, database fields and relationships, giving AI agents additional context when interpreting information. NVIDIA describes the library as a way to make enterprise data more understandable to AI systems without replacing existing semantic models.

According to the project's GitHub documentation, Auto Ontology works through three stages: generating an ontology, certifying its definitions and answering questions. The system examines database schemas, column values and relationships to propose business terms and SQL expressions. Human reviewers must then refine and certify these definitions before they can serve as trusted references. When processing a question, the agent identifies relevant concepts, generates SQL, checks the query against connected data and returns the answer with its SQL and processing steps.

The framework also provides a Model Context Protocol (MCP) server that allows compatible AI agents to access the semantic catalogue and submit natural-language questions. Its read-only MCP tools can retrieve business definitions, assess whether available data supports a question, and return query results. Supported relational databases include PostgreSQL, MySQL, Databricks, Snowflake, and DuckDB, among others.

Deployment involves more than installing a standalone Python library. The documented application uses a Python-based FastAPI backend, a Next.js frontend, PostgreSQL with pgvector and configured inference endpoints. NVIDIA NIM services are used by default, although compatible alternatives can be configured. The project provides Docker Compose and Kubernetes deployment options.

Auto Ontology is distributed under the Apache License 2.0 licence, but NVIDIA currently does not accept external code contributions. The maintainer team handles code changes, while integration requests and bug reports can be submitted through GitHub. The documentation describes mechanisms for checking queries and returning their provenance, but independently verified improvements in query accuracy have not been established in the reviewed material.

For the Drupal ecosystem, the development is relevant to efforts involving AI agents, structured data and external MCP services. The contributed MCP Client module enables Drupal applications to connect to external MCP servers, although it has not yet reached a supported stable release. No direct integration between Drupal and NVIDIA Auto Ontology has been documented. Its potential relevance therefore concerns how reviewed data definitions could support future AI integrations rather than any confirmed Drupal functionality.