NVIDIA's Python framework adds a semantic layer to agent queries over enterprise data, ensuring LLMs return correct answers backed by certified definitions and lineage.
NVIDIA Auto Ontology is a Python framework that builds semantic, graph-based reasoning layers over enterprise data. The core problem it addresses: agents can retrieve documents and query databases quickly, but if the answer is wrong, speed is worthless. The framework ensures agents understand what data means, whether to trust it, and where it came from.
What it does
Auto Ontology generates enterprise ontologies from existing metadata and reconciles conflicting definitions across systems. This creates a shared semantic layer that tells agents where data lives, what it means, whether to use it, and what is similar. The framework stores this ontology in Postgres alongside the catalog it describes.
The README states the framework exposes context to agents through MCP, CLI, and API. Developers can deploy it on Kubernetes using a Helm chart or run it locally with Docker Compose.
How queries work
When an agent receives a natural-language question, Auto Ontology translates it to SQL with deterministic query generation, powered by NVIDIA NIM endpoints. The default configuration uses Nemotron 3.5 Lightning 30B for reasoning, Nemotron 3 Embed 1B for embeddings, and Llama Nemotron Rerank VL 1B v2 for reranking. Each model role is configurable via environment variables, so teams can substitute self-hosted NIMs or different models.
Every query is validated against certified definitions and lineage before execution. The framework returns provenance with every answer.
Architecture and deployment
Auto Ontology runs as a multi-component stack: a Next.js 16 frontend (React 19, Tailwind CSS 4) handles the web UI and authentication via Better Auth; a FastAPI backend (Python 3.12+) manages chat, NL-to-SQL translation, and catalog APIs on port 3001; a separate FastAPI ingestion worker handles tabular data and embeddings on port 3002; and Postgres with pgvector stores metadata, the ontology, and vectors on port 5432. Optional HashiCorp Vault integration secures connection credentials.
The README notes that NVIDIA NIM endpoints require an API key and can point to either the hosted NVIDIA API catalog or self-hosted deployments. Configuration uses environment variable triplets for each model role (reasoning, non-reasoning, embeddings, reranking), with fallback to shared defaults.
Current state and limits
Auto Ontology launched on GitHub in April 2026 and reached 7,659 stars and 612 forks within months. The project has 57 open issues and is distributed under Apache 2.0. The README explicitly states the project is not currently accepting external contributions. The framework complements semantic models customers already own rather than replacing them.
Go deeper
Get the latest AI updates in the Tech AI Magazine app
iPhone · Android
