NVIDIA Auto Ontology
NVIDIA Auto Ontology, styled Auto-Ontology in its repository, is an open-source service that supplies business definitions and database context to AI agents. It maps terms such as customer or revenue to source columns, SQL expressions and table relationships. People can inspect, edit and certify those definitions before agents use them to answer questions.[1]
The software is part of NVIDIA's Agent Knowledge Foundation. It can complement an existing semantic model and connect to a separate agent harness through tools and APIs. It does not require replacing the source database with a new storage platform.[1][2]
Semantic model
Auto Ontology compiles source metadata into a semantic graph. Its definitions have three main forms.[1]
| Definition | Contents | Example use |
|---|---|---|
| Term | A business concept, its description and synonyms, and its column mappings | Identify which records represent a customer |
| SQL attribute | A reviewed expression attached to a term | Reuse an agreed revenue or margin calculation |
| Relationship | A connection between tables, including paths through bridge tables | Select the joins needed to answer a question |
Ingestion extracts schemas, profiles columns and samples values. Compilation proposes terms and relationships and can suggest SQL attributes from query history. The application stores the catalog and ontology in PostgreSQL, with pgvector embeddings for retrieval. Its web interface displays the relationships as a graph.[1]
Generated definitions remain a draft. The documented workflow has users test questions, refine the terms and SQL expressions, and certify them after review. Certification records a human decision about the definition; it does not establish that every subsequent answer will be correct.[1]
The application also imports and exports semantic definitions in Apache Ossie YAML through a converter. This provides an interchange path for existing models rather than requiring every definition to originate in automatic compilation.[1]
Question answering
For a question, the agent identifies intent and entities, retrieves candidate definitions by embedding search, and writes SQL from the selected context. It checks filter values and the query against the connected database, attempts repairs when checks fail, and executes the query. The result includes a prose answer, SQL and the agent's processing steps. The web interface can also show result rows, CSV downloads and charts.[1]
The answerability check estimates whether the semantic layer covers the concepts needed for a question. For example, the documentation distinguishes knowing a product's revenue from knowing its return on advertising spend when the database contains no advertising-spend data. The check is intended to support clarification before the full question-answering process runs.[1][3]
These operations combine model-generated reasoning with retrieved business context. SQL that executes successfully can still express an incorrect interpretation, so the documented evaluation compares returned results with known answers rather than treating successful execution alone as correctness.[4]
Agent interfaces
Auto Ontology exposes a web application, a REST API and a Model Context Protocol server. Three supplied agent skills cover setup, ontology management and querying.[1]
The MCP server has nine tools. Its allowlist provides semantic inspection and question answering without exposing the application's editing operations.[3][5]
| Tool | Purpose |
|---|---|
check_readiness | Check semantic compilation and the availability of a usable database connection |
search_terms | Find business glossary terms |
get_term | Read a term's description, synonyms and related terms |
get_term_columns | Read the physical columns mapped to a term |
get_term_sql_attributes | List SQL attributes associated with a term |
get_sql_attribute | Read one SQL expression and its purpose |
describe_table | Inspect a table together with its related terms and SQL attributes |
check_answerable | Assess coverage of the entities in a question |
ask_question | Return an answer, its SQL and result rows |
The MCP interface does not supply general database or schema browsing tools. describe_table supplies the semantic context of a particular table. The full application separately permits authorized users to edit definitions and manage connections.[1][3]
ask_question calls the same text-to-SQL agent as the web interface. It can require many sequential model calls. Its MCP response caps returned rows at 100 and reports the row count and whether truncation occurred. Readiness is separate from answerability: a populated catalog can remain readable after its source connection has become unusable.[3]
The REST question endpoint is POST /api/chat/completions. It streams progress, validated SQL, results and errors as server-sent events. A client can supply a conversation ID for follow-up questions; requests sharing that ID must run sequentially.[6]
Database and model support
The connector documentation lists ten source engines.[7]
| Connection route | Engines |
|---|---|
| Web interface or connection strings | PostgreSQL, MySQL, Databricks, Snowflake, ClickHouse, Trino, Apache Kyuubi and HeavyDB |
| Connection strings only | DuckDB and SQLite |
UI-managed connections can restrict catalog extraction with schema allowlists and table allow or deny expressions. The CONNECTION_STRINGS route does not carry those filters. The documentation advises using one configuration route per deployment because ingestion and question answering prioritize the two sources differently.[7]
Four model roles handle reasoning, shorter non-reasoning steps, embedding and reranking. The published defaults use Nemotron 3.5 Lightning for the first two, Nemotron 3.5 Embed 1B for embeddings and Nemotron 3.5 Rerank 8B for reranking. Each role can instead use a configured OpenAI-compatible endpoint. The documentation requires compatible structured outputs and a consistent embedding model between ingestion and querying.[1]
An optional KumoRFM integration handles predictive questions. Enabling it uploads connected tables to the hosted Kumo service. The repository documents a default maximum of 20 tables and an unlimited row cap unless the operator sets one. This path differs from simply running the ontology service beside a local database.[1]
Deployment and access controls
The reference stack uses a Next.js frontend, Python FastAPI backend and ingestion services, a FastMCP server, and PostgreSQL 17 with pgvector. Docker Compose supports the sample deployment; NVIDIA also publishes container images and a Helm chart for Kubernetes.[1][8]
Authentication uses Better Auth with role-based authorization. Self-service signup is disabled. API tokens act with their owner's permissions; MCP clients sign in through the deployment and receive user-specific authorization. The MCP server itself holds no source-database credentials.[3][9]
The gateway is a deployment boundary. The backend trusts the identity header forwarded by the frontend, while ingestion endpoints have no independent authentication. NVIDIA's instructions therefore keep those services on a private network and route clients through the frontend.[1][6]
Connection-credential storage also depends on configuration. Without HashiCorp Vault, UI-managed connection objects, including passwords, are stored unencrypted in PostgreSQL. A complete Vault configuration moves credentials to Vault. An incomplete configuration falls back to the PostgreSQL storage path with a warning.[7]
Evaluation and limits
The companion auto-ontology-eval repository runs controlled text-to-SQL comparisons. One arm supplies a model with the question and database schema in a single prompt. The other uses Auto Ontology's agent to retrieve context and answer through the compiled semantic model. A deterministic verifier executes predicted and gold SQL and compares the results using the relevant benchmark's rule.[4]
| Dataset | Documented evaluation scope |
|---|---|
| BIRD | 1,534 development questions or a 500-question Mini-Dev subset |
| BEAVER | A 100-question data-warehouse sample |
| FDABench-Lite | 169 questions |
| Wide World Importers | 41 questions against a PostgreSQL sample database |
| BIRD-Interact | Separate multi-turn evaluation on query tasks |
The control and treatment must use matching models. Optional SQL exemplars and value anchors provide additional help only to the treatment arm, so results from that configuration measure those inputs together with ontology grounding. The BIRD evaluator uses execution accuracy and does not compute BIRD's Valid Efficiency Score.[4][10]
The harness also separates failures that produced no model output from incorrect SQL. It withholds its headline comparison when too many questions fail because of a broken or rate-limited run, or when results are incomplete. These checks help prevent endpoint failures from being reported as model-accuracy measurements.[4]
NVIDIA's foundation page reports faster ontology creation in its own testing, while the evaluation repository provides a method for measuring question-answering accuracy. Those are different measurements. Neither the ability to compile an ontology nor the existence of an evaluation harness guarantees accurate answers on a new organization's data.[2][4]
License and maintenance
The project uses the Apache License 2.0. Third-party dependencies and external models retain their applicable terms. The maintainers accept bug reports and integration requests but state that they do not accept external pull requests.[1][11]
Continue reading
- AI AgentsAn AI agent is a software system that selects and performs actions in an environment in pursuit of an objective.Linked in the introduction
- NVIDIA Agent Knowledge FoundationNVIDIA Agent Knowledge Foundation is a collection of open models and accelerated libraries for connecting enterprise data to AI agents.Links back to this article
- Nemotron 3.5 LightningNVIDIA Nemotron 3.5 Lightning is an open-weights 30 billion parameter mixture-of-experts language model with 3 billion active parameters per token, released by NVIDIA on August 11Also in Open Source AI
- NVIDIA Agent ToolkitNVIDIA Agent Toolkit is NVIDIA's umbrella name for its collection of open models, agent blueprints, libraries, skills and runtimes for building and running AI agentsAlso in Enterprise AI
- OpenClaw EnterpriseOpenClaw Enterprise (OCE) is an open-source control plane for deploying, governing and auditing persistent AI agents on an organization's own infrastructure.Also in Enterprise AI
- H CompanyH Company, also known simply as H and originally code-named Holistic, is a Paris-based artificial intelligence company that builds computer use models and AI agents: software that reads a screen and operates…Also in Enterprise AI
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15NVIDIA. Auto-Ontology repository and README. Product scope, workflow, architecture, connectors and models. Accessed November 4, 2026.
- ^1 ^2NVIDIA Developer. Agent Knowledge Foundation. Component positioning and publisher-reported testing. Accessed November 4, 2026.
- ^1 ^2 ^3 ^4 ^5NVIDIA. Auto-Ontology MCP server documentation. Tools, authentication, readiness and row limits. Accessed November 4, 2026.
- ^1 ^2 ^3 ^4 ^5NVIDIA. Auto-Ontology Eval. Controlled evaluation, dataset scope and run-health checks. Accessed November 4, 2026.
- ^NVIDIA. Curated MCP tool surface. Explicit operation allowlist. Accessed November 4, 2026.
- ^1 ^2NVIDIA. Conversation API. Streaming response, conversation semantics and gateway trust boundary. Accessed November 4, 2026.
- ^1 ^2 ^3NVIDIA. Connections and credential storage. Source engines, scoping and Vault behavior. Accessed November 4, 2026.
- ^NVIDIA. Deploying Auto-Ontology on Kubernetes. Released containers and Helm chart. Accessed November 4, 2026.
- ^NVIDIA. Authentication, SSO and API tokens. Roles and token behavior. Accessed November 4, 2026.
- ^NVIDIA. BIRD evaluation setup and scoring. Dataset splits, optional context and execution-accuracy rules. Accessed November 4, 2026.
- ^NVIDIA. Auto-Ontology license. Apache License 2.0.
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 1,579 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Cite this page: AI Wiki. "NVIDIA Auto Ontology." aiwiki.ai, updated 4 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/nvidia_auto_ontology