The Auto Ontology interface combines a chat window with a semantic data catalog, illustrating its dual focus on interaction and context.
NVIDIA RELEASES OPEN-SOURCE AUTO-ONTOLOGY FRAMEWORK
*NVIDIA's new Auto Ontology library promises deterministic query generation and explainable AI for LLM agents. It builds a semantic graph over enterprise metadata, aiming to curb hallucinations in AI‑driven analytics.*
NVIDIA has just pushed its Auto Ontology framework to GitHub, positioning it as a core building block for trustworthy AI agents. The library promises deterministic query generation, better context retrieval and explainable outputs for large language models that currently stumble over inconsistent enterprise data. By layering a semantic graph over metadata, Auto Ontology aims to stop agents from delivering fast answers that are fundamentally wrong. The move signals NVIDIA’s intent to tighten the data‑trust chain in AI‑driven analytics, a sector where hallucinations cost revenue and credibility.
What is NVIDIA Auto Ontology?
Auto Ontology is a Python framework that creates a semantic, graph‑based reasoning layer over data. It ingests enterprise metadata, reconciles definitions across systems, and exposes the resulting ontology through MCP, CLI and API. The library supports natural‑language querying of structured data, translating questions to SQL via NVIDIA NIM. It stores the ontology and catalog in PostgreSQL with pgvector, and offers a web UI built on Next.js 16 for chat, analysis and settings. Distributed under the Apache License 2.0, the project currently lists 1,354 commits and explicitly rejects external contributions.
How does it integrate with enterprise data?
Developers can generate ontologies directly from metadata sources such as Databricks, Snowflake, Postgres or DuckDB. The system reconciles inconsistent metric definitions, attaches provenance to every answer, and folds certified corrections back into the graph. A semantic catalog layer sits alongside relational sources, enabling agents to discover entities, relationships and trusted business meanings. Authentication and role‑based access control rely on Better Auth, while the backend FastAPI services (ports 3001 for core APIs and 3002 for ingestion) write embeddings to the Postgres vector store.
Which models power its NL‑to‑SQL capability?
Auto Ontology defaults to NVIDIA NIM endpoints for inference. The main reasoning model is nvidia/nemotron-3-nano-30b-a3b, accessed via the DEFAULT_MODELS triplet. Text embeddings use nvidia/nemotron-3-embed-1b, and reranking employs nvidia/llama-nemotron-rerank-vl-1b-v2. Each role can be overridden by setting specific environment variables for endpoint, API key and model name. Unset fields fall back to the shared DEFAULT_MODELS settings, simplifying configuration for most deployments.
What are the deployment options and restrictions?
The stack runs on Docker Compose v2 for local testing, exposing ports 3000 (frontend), 3001 (backend), 3002 (ingestion) and 5432 (Postgres). Kubernetes users can deploy via the provided Helm chart. A valid NVIDIA API key (nvapi‑…) is required to call hosted NIM services or self‑hosted equivalents. Connection details for the target databases must be supplied. The repository includes a .env.example file, but the project does not accept outside contributions, limiting community extensions.
Frequently asked questions
Can developers contribute to NVIDIA Auto Ontology?
The project is currently not accepting external contributions.
What license governs NVIDIA Auto Ontology?
Auto Ontology is distributed under the Apache License 2.0.
What are the prerequisites for running Auto Ontology locally?
Docker with Docker Compose v2, a valid NVIDIA API key, and connection details for the source database are required.
Auto Ontology arrives at a moment when enterprises demand verifiable AI results. Its open‑source code, strict licensing and built‑in provenance may set a new baseline for LLM reliability. Adoption will hinge on how quickly firms can integrate the stack into existing data pipelines and secure the required NVIDIA API keys. If the framework gains traction, it could reshape the economics of AI‑augmented decision making, turning speed into a credible advantage rather than a liability.
Sources: Hacker News, GitHub NVIDIA/auto-ontology