Business terms and reviewed measures
Map terms to columns and attach reviewed SQL to measures whose business definitions need to be consistent.
NVIDIA Auto Ontology
NVIDIA Auto Ontology maps business terms, metrics, and table relationships to database evidence so agents can answer questions with SQL you can inspect.
Make one business metric reproducible
Choose a metric that different teams calculate differently. Use Auto Ontology’s semantic layer as a reference for tying the definition to tables, reviewed SQL and a small evaluation set.
Name the metric, candidate tables, reporting period and exclusions.
Copy the request into a new Floatboat conversation and provide the required material.
Download FloatboatYour request for Floatboat
Preview and copy the request. It is not sent automatically from this page.
Reference: NVIDIA Auto Ontology https://github.com/NVIDIA/auto-ontology Read the original documentation and check the current environment and dependencies. Propose any new installation or tool connection before proceeding. My task: Read NVIDIA Auto Ontology’s README and MCP documentation. Help me specify monthly net revenue from orders, invoices, and refunds. Return a definition, reference SQL for a sample schema, and five test questions. Flag ambiguous joins and ask before connecting to a live database. Expected deliverable: A metric contract, reference query, and cases for detecting bad joins or missing coverage.
How NVIDIA Auto Ontology grounds database questions
A question about revenue can be ambiguous even when the database is well documented. The agent needs to know which revenue definition applies, which tables can be joined, and which records to exclude. Auto Ontology compiles database metadata into a semantic layer that maps business terms to columns and relationships and can hold reviewed SQL for measures.
The project runs alongside the databases. Its chat interface can return the SQL, rows and steps behind an answer. Its read-only MCP server also exposes term lookup, answerability checks and questions to authenticated clients under the signed-in user's permissions. The companion setup, management and query Skills guide a coding agent through those operations.
When designing a Floatboat workflow for periodic reporting, use the ontology as a pattern for giving metrics an owner, a definition and a reproducible query. Start with a small test database and compare each result to an analyst-approved reference before treating the workflow as a source for decisions.
What Auto Ontology exposes to an agent
Follow a business question from the term it matched to the SQL and rows it produced. That gives an analyst separate places to check coverage, join logic and the underlying records.
Relationships and query evidence
Follow table relationships, inspect the generated SQL and compare the returned rows with a known result.
Read-only MCP access
Search terms and check whether a question is answerable before issuing it through a separately deployed, authenticated MCP server.
NVIDIA Auto Ontology use cases
Start with a question whose expected answer an analyst can reproduce. Use the SQL and rows to diagnose whether a mismatch comes from the definition, the join, the time window, or the underlying source.
Resolve two definitions of revenue
If finance reports net revenue and sales reports booked revenue, document the exclusions and approved query for each. Test the same monthly request against both measures and make the reporting agent surface the definition it selected.
Check joins before publishing a weekly KPI
Compare the generated order-to-invoice join against reference SQL using orders with multiple invoice lines. Review row counts and totals together so a plausible-looking aggregate does not hide duplicate records.
Route unsupported questions for review
Include a region or product attribute missing from the ontology in an evaluation set. Check answerability before asking for SQL, and capture the missing mapping for an analyst instead of filling the report with a guess.
Keep a reporting decision reproducible
Save the time window, approved measure, SQL and returned rows beside the result. When a source schema or definition changes, rerun the original question and give reviewers evidence for any difference.
What a test deployment needs
The upstream quickstart uses Microsoft’s Wide World Importers sample. Its Docker Compose configuration runs the supporting services, while the configured inference endpoints and credentials are managed separately.
Local runtime
Install Docker with Compose v2 and uv for the upstream sample setup.
Model endpoint access
Set the credentials for your configured inference endpoints. The documented sample quickstart uses an NVIDIA API key.
Database scope
Start with the sample database or an approved test source and restrict the credentials to the data needed for the task.
Related Software Projects
Compare the workflows, operating requirements, and source licenses of these projects.
NVIDIA Auto Ontology FAQ
What does the ontology store?
It connects business terms, columns, table relationships and SQL definitions. The underlying records stay in the connected data sources.
Which SQL has been reviewed, and which SQL do I still need to check?
A data owner can review the SQL attached to a measure. The query generated for a particular question still needs to be checked against the requested time window, joins and returned rows.
Can an MCP client change my ontology?
The upstream MCP server exposes read-only tools. Authenticate with the deployment and review its account permissions before testing.
What if a metric is missing?
Use the answerability tools to identify missing coverage and ask a data owner to define or approve the metric before using it in reporting.
What are the NVIDIA ontology Skills for?
They guide setup, semantic model management and validated querying. The query workflow depends on a running Auto Ontology deployment.
Does the sample setup need model credentials?
Yes. NVIDIA documents an API key for the sample quickstart. Check the model endpoint and credential requirements when changing the configuration.