NVIDIA releases Auto Ontology to ground AI agents in enterprise data

Bridging the gap between agents and data

NVIDIA has introduced Auto Ontology, an open library designed to act as a semantic layer between AI agents and enterprise data warehouses. The primary goal of this framework is to provide a context layer that tells an agent exactly where data lives and what specific fields mean. Without such a layer, AI agents often struggle to interpret complex database schemas correctly.

The developers emphasize that a fast answer is not useful if the underlying number is incorrect. Auto Ontology addresses this by allowing agents to write SQL queries against certified business definitions rather than raw tables. This ensures that the metrics and entities the AI interacts with have a consistent and trusted meaning across the organization.

The library is released under the Apache 2.0 license, making it accessible for various types of deployments. It is designed to be an open semantic layer that any agent harness can query to find reliable information. By grounding the agent in a specific context, the system reduces the risk of the AI hallucinating data relationships or using the wrong columns for a calculation.

Core services and technical architecture

The Auto Ontology stack is composed of four core services that handle different parts of the data pipeline. A Next.js 16 frontend provides a web user interface and handles authentication on port 3000. Meanwhile, a FastAPI backend powered by LangChain manages the chat functionality, natural language-to-SQL conversions, and catalog APIs on port 3001.

Data storage and ingestion are handled by a Postgres database equipped with pgvector to store the catalog, the ontology, and various vectors. An ingestion worker runs on port 3002 to write these embeddings. The system is designed to connect with several popular source databases, including Databricks, Snowflake, DuckDB, and Postgres.

For inference, the framework allows users to set models through environment triplets consisting of an endpoint, a key, and a model name. While it defaults to NVIDIA Nemotron models on NIM, it is compatible with any OpenAI-compatible endpoint. This flexibility allows developers to choose the specific reasoning, embedding, or reranking models that best suit their needs.

How agents interact with the data

One of the key features of the release is the inclusion of a Model Context Protocol (MCP) server. This server exposes nine read-only tools that allow agents to interact with the data safely. For example, a tool called ask_question returns the final answer, the generated SQL, and the resulting rows as separate fields, with a cap of 100 rows per result.

NVIDIA intentionally chose not to include a tool for browsing raw schemas within this MCP server. This design choice forces agents to reach data exclusively through the semantic layer rather than bypassing the certified definitions. By restricting the agent's view, the system ensures that the AI follows the established business logic every time it queries the database.

The system also includes a validation step to check generated queries against certified definitions and data lineage. This means the framework can verify if a query is appropriate before it is executed. Additionally, the system can learn from human interaction by folding certified corrections back into the ontology to improve future accuracy.

Testing and deployment options

NVIDIA has provided a separate evaluation harness called auto-ontology-eval to help users test the system. This harness runs two different arms in NeMo Gym: a schema-only LLM control and an ontology-grounded agent. The scoring is deterministic, meaning it checks for an exact execution match rather than using an LLM to judge the quality of the answer.

The evaluation includes several bundled datasets, such as BIRD Dev and Mini-Dev. However, as MarkTechPost reported, NVIDIA has not yet published specific text-to-SQL accuracy scores for the tool. Users will need to run the provided harness themselves to verify the performance on their own specific data types.

Deployment is flexible, offering Docker Compose for local setups and Helm charts for Kubernetes environments. The backend is also published as an NGC container for easier integration. While the tool is designed for high performance, the public evidence currently relies on NVIDIA's self-reported testing results regarding speed and efficiency.

The path toward certified AI

The output of Auto Ontology is framed as a starting point rather than a finished product. NVIDIA's developer page clarifies that business experts must still test, refine, and certify the generated ontology. This human-in-the-loop approach ensures that the AI remains aligned with the actual rules of the business while still benefiting from automated generation.

By automating the creation of the semantic layer, the tool aims to make it easier for companies to deploy reliable AI agents. It bridges the gap between the raw data stored in a warehouse and the structured knowledge required for an LLM to perform useful tasks. This could significantly lower the barrier for non-technical teams to access complex data insights.

Moving forward, the focus will be on how well these grounded agents perform in production environments. Because the tool is MCP-ready, it can be integrated into a wide variety of agent harnesses. The success of the framework will likely depend on how quickly organizations can move from raw metadata to a certified, usable ontology that agents can trust.

Next steps for developers

Developers interested in testing the framework can access the GitHub repository to find the library and the evaluation harness. The documentation for the MCP server provides specific details on the nine read-only tools and the OpenAPI spec, which lists 87 total operations. Users can also explore the NGC container for backend deployment.

To get started, users can set up a local stack using Docker Compose after configuring their environment variables. For those using NVIDIA's infrastructure, an API key from the NVIDIA build site will cover the default models. This setup allows for a quick transition from a standard LLM to an ontology-grounded agent.

The tool is particularly useful for those who need to connect to Snowflake or Databricks but want to maintain strict control over how data is interpreted. By using Auto Ontology, these users can ensure that their AI agents are always looking at the 'source of truth' as defined by the company's own experts.