Back home

Auto Ontology

github+x 0.86
Visit source

Intro

Auto Ontology (NVIDIA Auto Ontology) is NVIDIA's open enterprise semantic-layer / ontology framework that gives LLMs and agents trusted context about what data means. Repo: https://github.com/NVIDIA/auto-ontology ; also part of the NVIDIA Agent Knowledge Foundation toolkit.

Core jobs (official):

  • Discover: generate ontologies from metadata, connect existing semantic models, build knowledge graphs;
  • Ground: certified business meaning for metrics, entities and relationships;
  • Validate: check generated SQL against certified definitions and lineage; return provenance;
  • Expose: read-only MCP, CLI and API for any agent harness;
  • Learn: fold human corrections and usage back into the ontology.

Deploy via Docker Compose or Kubernetes Helm; inference can use NVIDIA NIM (e.g. nemotron-3-nano). Covered by MarkTechPost around 2026-11-01; in this scan window (08:32–12:32 BJT on Nov 6) the repo hit #1 on GitHub Trending daily with roughly +2.6k stars/day.

Explain

Problem

Agents can already query warehouses and retrieve docs, but if "revenue" means different things in different systems, a fast answer can still be wrong. Auto Ontology sits a business-expert-certified definition layer between the agent and the warehouse.

Positioning

It complements rather than replaces existing semantic models (dbt semantic layer, Cube, AtScale, etc.); the pitch is MCP plug-in plus NIM integration for Agent Toolkit / enterprise agents.

Limits

  • Few published text-to-SQL accuracy numbers — run your own evals;
  • Repo currently not accepting external contributions;
  • Open-source announcement predates this window, but viral GH+X traction hit now, so we backfill.

Usage

Who it's for

  • Data/platform teams grounding enterprise agents' text-to-SQL / analytic Q&A with explainable lineage;
  • Teams with existing semantic models who want MCP exposure to Claude/Cursor-class agents;
  • NVIDIA NIM / Agent Toolkit customers.

Getting started

  1. Read https://github.com/NVIDIA/auto-ontology README;
  2. Bring up the Docker Compose stack, ingest metadata and generate a starter ontology;
  3. Have business experts certify terms in the UI, then wire MCP into your agent;
  4. Use tools like ask_question / check_answerable and inspect returned SQL + provenance.

Advertisement