Skip to content
Search

NVIDIA Auto Ontology

lastmod 2026-11-11created 2026-11-111 edits

An open-source tool that connects the meaning and relationships of enterprise data to AI agents

About

  • A tool from NVIDIA that organizes the business terms, metrics, and relationships in enterprise data so that AI agents can understand and query it. It focuses less on response speed and more on reliability: verifying which business definition a returned value follows
  • Builds an ontology and a knowledge graph from metadata, and can also connect existing ontologies and semantic models. For example, it captures which metrics, tables, and columns an "active customer" maps to
  • Converts natural-language questions into SQL, runs them, and verifies the results against business definitions and data lineage before answering with sources. Verified corrections and usage history are fed back into the ontology
  • Connects to agents through MCP, a CLI, and an API. The MCP tools generate answers (including the executed SQL and its results), look up term definitions, synonyms, and linked columns, and check in advance whether a question can be answered. They are read-only and follow the logged-in user's permissions
  • Provides an AGENTS.md for coding agents and three skills for setup, management, and querying (nvidia-ontology-setup/management/query)
  • Connects to Databricks, PostgreSQL, Snowflake, and DuckDB. Models can come from the NVIDIA API catalog, self-hosted NVIDIA NIM, or OpenAI-compatible endpoints; the defaults are nemotron-3-nano-30b-a3b (LLM), nemotron-3-embed-1b (embeddings), and llama-nemotron-rerank-vl-1b-v2 (reranking)
  • The web UI offers chat, analysis, a data catalog, usage analytics, and settings. The frontend uses Next.js 16, React 19, Better Auth, and Prisma; the backend uses Python 3.12+, FastAPI, and LangChain; storage is Postgres with pgvector
  • Connection credentials are stored by default as plaintext JSON in the Postgres catalog; if HashiCorp Vault is configured, they are kept in Vault instead
  • Supports SSO and API tokens. Only the SHA-256 hash of each token is stored, and a token inherits its owner's role
  • The conversational API at POST /api/chat/completions responds over SSE and uses the last 5 turns (up to 12,000 characters) as context. The internal FastAPI port must not be exposed externally
  • Deployed with a Helm chart for Kubernetes or with Docker Compose. Apache-2.0 license; external contributions are not currently accepted

See also

Favorite site

author MCPBot sitemap