NVIDIA's answer to confidently wrong agents: an ontology layer

What it does

Auto-Ontology is a service you run alongside your databases. It ingests schemas, profiles columns, and compiles a draft ontology: business terms mapped to columns, reviewed SQL expressions for measures, and inferred join relationships. Humans review and certify that draft in a web UI, and agents then query through it via MCP, agent skills, or a REST API. The payoff: answers come back with the SQL and the reasoning steps, and unanswerable questions get flagged instead of answered anyway.

The interesting bit

The core bet is that the usual shortcut — pasting the schema into the context window — doesn’t scale. Tokens grow with every table, joins get improvised per question, and “we have no ad spend data” becomes a confidently delivered number. Auto-Ontology instead gives the agent tools over a curated graph, including a check_answerable step that measures coverage before querying and asks for clarification when a term doesn’t resolve. It’s also compatible with Apache Ossie, the open semantic-model exchange standard, so you can import models you already own rather than starting from zero.

Key highlights

• Three-stage pipeline: generate ontology from metadata, certify it with humans, answer questions from it
• Answerability checking — the agent says “no term covers ad spend” instead of inventing one
• Every answer ships with its SQL and resolution steps, so results are auditable against the source
• Serves any agent harness through MCP, agent skills, and REST
• Companion repo auto-ontology-eval benchmarks it against schema-in-prompt baselines on BIRD, BEAVER, and others, with a deterministic execution-match verifier

Caveats

• The eval repo exists but the README doesn’t state the actual accuracy numbers — you’ll have to run it yourself
• It’s a full stack (Next.js frontend, three FastAPI services, Postgres with pgvector), not a library you drop into an existing pipeline
• Requires an NVIDIA API key by default; the README doesn’t say whether other providers work

Verdict

Worth a look if you’re building agentic analytics over enterprise databases and reliability matters more than latency. If your data fits comfortably in a context window and wrong answers are cheap, this is a lot of machinery for the problem.