Rank #02 · Daily · AI
auto-ontology: NVIDIA's Graph Ontology Layer for Deterministic Agent Analytics
Give agents a certified business meaning layer — terms, joins, and SQL attributes — so analytic answers come with SQL, steps, and a clear "unanswerable" path when coverage is missing.
What it does
Auto Ontology is NVIDIA's open library for a semantic, graph-based reasoning layer over structured data. The README frames the reliability problem directly: retrieving documents or querying a database faster is useless if the metric is wrong. Agents need to know where data lives, what it means, whether to use it, how to use it, where it came from, and what is similar.
Core workflow from the docs:
- Generate. Ingestion profiles schemas, samples values, extracts terms (business concepts with synonyms mapped to columns), infers relationships (joins, including bridge tables), and suggests SQL attributes (reviewed expressions such as margin or rolling totals). The ontology and embeddings land in Postgres + pgvector.
- Certify. People refine terms and attributes in the web UI / API, then certify names, descriptions, and SQL so answers compute measures the same way.
- Answer. The agent classifies intent, retrieves candidates by embedding search, writes SQL, validates filters against the data, repairs when checks fail, and returns the answer with SQL and the steps that produced it. When coverage is insufficient,
check_answerablecan refuse instead of hallucinating a number.
Surfaces for other agents:
- MCP server (FastMCP) with read-only tools: readiness, term search,
check_answerable,ask_question. - Agent skills for setup, ontology management, and querying (
AGENTS.md). - REST conversation API (
POST /api/chat/completions) with API tokens.
Stack sketch: Next.js frontend (auth via Better Auth), FastAPI backends (chat / NL-to-SQL, ingestion), Postgres + pgvector, optional HashiCorp Vault for credentials. Inference defaults to NVIDIA NIM endpoints (hosted catalog or self-hosted); model roles are configured as endpoint / API key / model triplets. Connectors listed include PostgreSQL, MySQL, Databricks, Snowflake, ClickHouse, Trino, Kyuubi, HeavyDB, DuckDB, and SQLite. Compatible with Apache Ossie for importing/exporting semantic models. Docker Compose and Helm are both documented.
Note from the README: the project is not currently accepting external contributions; integration requests and bug reports go through issues.
Who it's for
- Teams building agentic analytics / text-to-SQL who already own (or want) a semantic model and need explainable, provenance-bearing answers.
- Platform engineers wiring an ontology into Cursor, Claude Desktop, or other MCP clients.
- Organizations that can run a small services stack (Postgres + NIM credentials) and spend time certifying definitions — this is not a zero-config chat-over-CSV toy.
Stars / language / license
5,086 stars · Python · Apache-2.0
Data verified: 2026-11-05 09:31 Asia/Shanghai (GitHub API get_repository + README get_file_contents)
Trending note: +2,050 stars today (Trending selection captured 09:26 Asia/Shanghai)
Category tags
AI · ontology · text-to-SQL · RAG · MCP · NVIDIA · enterprise-data · explainable-AI
How it differs from similar projects
Compared with plain RAG over docs or schema dumps. Classic RAG retrieves text chunks; Auto Ontology retrieves certified terms and SQL attributes, then validates generated SQL against the ontology and data. The product goal is deterministic measure definitions, not "closest paragraph."
Compared with dbt / LookML / Cube semantic layers alone. Those define metrics for BI tools. Auto Ontology adds an agent harness (chat, MCP, answerability checks, provenance) and can import/export via Ossie converters so existing semantic models are complements, not replacements.
Compared with generic LangChain text-to-SQL demos. Those usually prompt over raw DDL. Here the SQL is grounded in a compiled ontology with certification and repair loops. NVIDIA also points BIRD-style evals to a separate ontology-sql-eval repo — TidyHow has not reproduced those numbers.
Caveats. Defaults lean on NVIDIA NIM API keys; backend ports trust identity headers from the frontend and should stay private. Predictive routing to KumoRFM (optional) uploads table data — review before enabling. Treat architectural claims above as README-sourced, not independently measured.