NVIDIA releases Auto Ontology, an open-source framework for agent data context

• NVIDIA has released Auto Ontology, an open-source Python framework under the Apache License 2.0 that builds a semantic, graph-based reasoning layer over data to enable deterministic query generation and explainable AI workflows.
• The system creates ontologies from enterprise metadata, resolves definition mismatches across systems, and exposes context to agent harnesses through MCP, CLI, and API endpoints.
• It utilizes NVIDIA NIM endpoints for inference, with default configurations for the nvidia/nemotron-3-nano-30b-a3b LLM, nvidia/nemotron-3-embed-1b embeddings, and nvidia/llama-nemotron-rerank-vl-1b-v2 reranking model.
• The repository currently has 1.1k stars and 112 forks, but NVIDIA is not accepting external contributions, issues, or pull requests at this time.

Architecture and Stack

• Frontend built with Next.js 16, React 19, and TypeScript; backend powered by FastAPI and Python 3.12+.
• Uses Postgres with pgvector for metadata, catalog, ontology, and vector store storage.
• Supports Databricks, Postgres, Snowflake, and DuckDB as data sources.

Agent Integration and Security

• Exposes an Agent API at POST /api/chat/completions using SSE streaming for stateless or multi-turn conversations.
• Provides MCP server support for agent harnesses like Cursor and Claude Desktop.
• Stores database credentials in HashiCorp Vault if configured, otherwise in Postgres; API tokens are stored as SHA-256 hashes.