NVIDIA Auto Ontology Is WILD: The Agent Layer That Says "I Can't Answer That" (README Tour, Not Run)

NVIDIA’s Auto Ontology is an open-source semantic layer that enhances AI agents’ reliability in querying enterprise databases by clarifying data meanings and enabling agents to admit when they lack sufficient information, thus preventing incorrect answers. The system builds and certifies an enterprise ontology from metadata, supports multiple interfaces and databases, and emphasizes transparency, security, and human validation to ensure trustworthy AI-driven business intelligence.

NVIDIA’s Auto Ontology is an open-source project designed to improve the reliability of AI agents querying enterprise databases by providing a semantic context layer that clarifies the meaning of data columns. Unlike typical agents that might confidently provide incorrect answers due to ambiguous or inconsistent data definitions, Auto Ontology enables agents to say, “I can’t answer that,” when the data lacks the necessary information. This feature is particularly valuable in complex business environments where metrics like revenue can have multiple conflicting definitions across different departments. The project has gained significant attention, boasting nearly 10,000 stars on its GitHub repository, and aims to bridge the gap between raw database schemas and meaningful business concepts.

The core idea behind Auto Ontology is to build an enterprise ontology from metadata that reconciles definitions across various systems and presents this knowledge to AI agents through multiple interfaces, including a command line, API, and MCP server. The ontology consists of three main components: terms (business concepts like customer or invoice line), SQL attributes (predefined SQL expressions for consistent metric calculations), and relationships (table joins and bridge tables inferred from keys and data). This semantic graph ensures that agents understand the true meaning of data elements and how they relate, preventing guesswork in query formulation and result interpretation.

The system operates in three stages: generate, certify, and answer. The generate stage involves ingesting metadata, profiling columns, mapping entities to columns, resolving relationships, and storing everything in a Postgres database with PG Vector embeddings. Certification is a crucial human-in-the-loop step where users validate and refine the ontology’s terms, descriptions, and SQL attributes to ensure accuracy and trustworthiness. Finally, the answer stage involves the agent classifying query intent, retrieving relevant ontology elements, constructing and validating SQL queries, and returning answers along with provenance information and the SQL used. This process enables transparency and trust in the agent’s responses.

Setting up Auto Ontology requires Docker, an Nvidia API key, and two repositories: the main Auto Ontology repo and an evaluation repo for seeding sample data. The system includes a Next.js frontend, FastAPI backend, ingestion service, MCP server, and Postgres database. It supports multiple database engines and integrates with Apache Oozie for semantic model exchange, reducing vendor lock-in concerns. The default AI models are Nvidia’s NeMo tron series, but users can configure roles with different endpoints and keys. The project emphasizes security by recommending private networks for certain services and supports role-based access control to ensure agents only access authorized data.

While the reviewer did not run the system personally, the documentation and design suggest that Auto Ontology addresses a critical pain point in enterprise AI: ensuring agents provide accurate, trustworthy answers by understanding data context rather than guessing. It is best suited for teams with complex databases, recurring metric discrepancies, and resources to certify the ontology. The project is actively maintained but does not accept external code contributions, focusing instead on issue tracking. Overall, Auto Ontology represents a serious, well-documented approach to improving AI agent reliability in business intelligence, encouraging users to identify and certify their most contentious metrics first.

Useful Links