NVIDIA has open-sourced Auto Ontology, a framework designed to make AI agents more reliable when querying enterprise databases using natural-language instructions. The system creates a semantic layer between an AI agent and business data, helping agents understand what database fields mean, how business metrics are defined and which tables should be used before generating SQL queries.
Released under the Apache 2.0 licence, Auto Ontology can expose enterprise context to AI agents through the Model Context Protocol (MCP), command-line tools and APIs. Its goal is to reduce incorrect SQL queries and misleading business answers caused by AI systems misunderstanding database structures or using inconsistent definitions.
Key Takeaways
- NVIDIA has released Auto Ontology as an open-source framework under the Apache 2.0 licence.
- The technology creates a semantic layer between AI agents and enterprise databases.
- It helps agents understand business terms, metrics, relationships and database structures.
- It validates generated SQL against certified business definitions and data lineage.
- The framework supports MCP, CLI and API integrations.
- It supports Databricks, PostgreSQL, Snowflake and DuckDB as source databases.
- NVIDIA says Auto Ontology reduced ontology creation time by 90% in its internal testing.
- Developers can deploy it through Docker Compose or Kubernetes.
What Is NVIDIA Auto Ontology?
Auto Ontology is a Python-based framework that creates a semantic and graph-based reasoning layer over enterprise data.
To understand its purpose, consider how an ordinary AI agent queries a business database.
A company may have hundreds of database tables containing information about sales, customers, products, payments and operations. An employee might ask an AI agent:
“How much revenue did we generate from enterprise customers last quarter?”
The agent must identify the correct revenue field, understand what qualifies as an enterprise customer, select the appropriate reporting period and generate a SQL query.
The problem is that database structures do not always explain business meaning.
For example, a company might maintain several revenue-related fields, including gross revenue, net revenue, recognised revenue and recurring revenue. An AI agent could select the wrong field and still generate SQL that executes successfully.
Auto Ontology attempts to solve this problem by providing a layer of business context before the agent queries the database.
It helps explain what the data means, where it comes from, how different business terms relate to each other and which definitions should be used.
How Does Auto Ontology Work?
NVIDIA’s framework is designed around five important capabilities.
| Capability | Function |
|---|---|
| Discover | Generate an initial ontology using enterprise metadata and documentation |
| Ground | Provide business definitions for metrics, entities and relationships |
| Validate | Check generated queries against certified definitions and data lineage |
| Expose | Make the semantic context available through MCP, CLI and APIs |
| Learn | Incorporate approved corrections and usage information into the ontology |
The process begins by examining available enterprise metadata and existing semantic models.
The system then creates an initial ontology that describes relationships between business concepts and the underlying database structures.
Business experts can review, refine and certify these definitions before AI agents use them.
When an agent receives a natural-language question, it can retrieve the relevant business context and use that information to construct a SQL query.
The framework can also check the generated query against approved definitions and return information about the data’s origin.
This makes the process more transparent than allowing an AI model to generate SQL using database column names alone.
Why AI Agents Often Generate Incorrect SQL
Text-to-SQL technology allows users to communicate with databases using ordinary language rather than writing SQL manually.
However, translating natural language into SQL is difficult when enterprise databases contain complex structures.
There are several common problems.
1. Ambiguous business definitions
Different departments may use the same term differently.
For example, the sales team might define an active customer based on recent purchases, while the finance department uses an accounting definition.
An AI agent may choose the wrong definition without additional context.
2. Complex database relationships
A business database may contain multiple tables connected through customer IDs, product IDs, transaction IDs and other relationships.
Selecting the wrong join can produce incorrect results even when the SQL query is syntactically valid.
3. Missing historical context
A query about quarterly revenue may require information about accounting periods, refunds or reporting adjustments that are stored elsewhere.
4. Incorrect metric selection
An agent may confuse gross merchandise value with recognised revenue or total orders with completed orders.
Auto Ontology attempts to reduce these errors by giving agents access to a structured business glossary, relationships and approved metric definitions.
NVIDIA Says Ontology Creation Time Fell by 90%
NVIDIA says Auto Ontology reduced ontology creation time by 90% in its internal testing.
An ontology describes the concepts, relationships and rules that explain how information is organised within a particular business or domain.
Traditionally, creating an enterprise ontology can require significant work from data engineers, analysts and business specialists.
They must identify important entities, understand database relationships, document metrics and reconcile conflicting definitions across systems.
Auto Ontology automates the generation of an initial ontology using available enterprise metadata and documentation.
However, the generated ontology is not automatically treated as an authoritative source.
NVIDIA says business experts must test, refine and certify the resulting definitions.
This distinction is important because automated metadata analysis cannot always determine the correct business meaning of a field without human input.
The 90% figure is an NVIDIA-reported internal result, not a guarantee that every organisation will achieve the same reduction.
Integration With MCP and AI Agent Frameworks
One of Auto Ontology’s important features is its ability to provide semantic context to different AI agent environments.
The framework supports three integration methods:
- Model Context Protocol (MCP): Allows compatible AI agents to access ontology information and query enterprise data through standardised tools.
- Command-line interface: Allows developers to interact with the framework through terminal-based workflows.
- API: Enables applications and custom agent systems to retrieve semantic information programmatically.
The MCP integration is particularly relevant as more AI applications adopt standardised ways of connecting agents to external tools and data sources.
Instead of building a separate database-understanding system for every AI agent, organisations can expose a shared semantic layer.
This could allow compatible agents to use the same approved business definitions.
The framework also includes tools for searching business terms, retrieving their descriptions, examining their database mappings and checking whether a database is ready for queries.
Which Databases and Technologies Does It Support?
NVIDIA has designed Auto Ontology to work with existing enterprise data infrastructure.
| Technology | Purpose |
|---|---|
| Databricks | Enterprise data and analytics platform |
| PostgreSQL | Relational database |
| Snowflake | Cloud data warehouse |
| DuckDB | Analytical database |
| PostgreSQL with pgvector | Stores ontology information and vector data |
| NVIDIA NIM | Provides model inference endpoints |
| FastAPI | Backend APIs and query processing |
| Next.js | Web interface |
The framework uses NVIDIA NIM for its default reasoning, embedding and reranking models.
Its default configuration includes Nemotron models, but administrators can configure other OpenAI-compatible model endpoints.
This gives organisations some flexibility in selecting the models used for reasoning and retrieval.
Auto Ontology can be deployed locally through Docker Compose or on Kubernetes using a Helm chart.
Open-Source Availability and Licence
NVIDIA has released Auto Ontology under the Apache 2.0 licence.
Developers can access the source code, inspect its architecture and deploy the framework in their own environments.
The official repository includes installation instructions, configuration examples, deployment documentation and an MCP server.
However, NVIDIA says the project is currently not accepting external contributions. Developers can inspect and use the project under its licence, but outside pull requests are not currently being accepted for integration.
The open-source release also does not mean that every component is completely free of infrastructure costs. The default configuration uses NVIDIA NIM inference endpoints, and organisations must account for their model inference, database and deployment requirements.
The Bigger Picture
NVIDIA’s Auto Ontology release highlights an important challenge in enterprise AI: generating an answer is not enough if an agent does not understand the business meaning of the underlying data.
As organisations introduce AI agents for financial analysis, sales reporting, operations and business intelligence, reliable access to structured information becomes increasingly important.
A semantic layer can help bridge the gap between how employees describe business questions and how databases actually store information.
NVIDIA is also positioning Auto Ontology as one component of its broader Agent Knowledge Foundation, which includes retrieval, data processing, graph analytics and predictive technologies.
The strategy suggests that enterprise AI development is moving beyond increasingly capable language models towards complete systems that can retrieve, understand, validate and act on business information.
Looking Ahead
The next important test will be how Auto Ontology performs across real enterprise databases with inconsistent definitions, complicated relationships and large volumes of historical data. Independent text-to-SQL benchmarks and production deployments will help establish whether its semantic grounding produces measurable improvements over conventional database-querying agents.
If the framework proves effective, it could help businesses build AI assistants that answer questions about sales, finance, inventory and operations with better traceability. Its open-source availability also gives developers an opportunity to integrate a shared business knowledge layer into their existing AI agent systems.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
