NVIDIA Open-Sources Auto Ontology as a Semantic Layer for AI Agents

X postGul Saeed KhattakAnnouncement

NVIDIA has open-sourced Auto Ontology, a semantic layer designed to stop AI agents from hallucinating or picking incorrect database columns during SQL queries.

Key points

  1. Auto Ontology acts as a semantic layer between AI agents and underlying databases to prevent guesswork around column selection.
  2. Standard protocols such as MCP establish connections but do not resolve business-level ambiguity between similar metrics.
  3. Database schema comments help describe fields but fail to define operational business rules and date logic.
  4. Unlike basic RAG setups that may conflate net and gross terms via embeddings, Auto Ontology incorporates business terms, a certified state, and a learn step.
  5. Provenance can be audited through execution logs and term attributes, though human certification remains a requirement.

Why it matters

AI-generated database queries frequently produce clean charts based on faulty columns or incorrect joins, creating hidden risks in executive reporting. A semantic layer enforces business definitions before queries execute.

Analysis

Khattak points out that semantic SQL access is not a novel concept, but the primary bottleneck has always been organisational governance: deciding who approves definitions and maintaining them over time. While NVIDIA claims a 90% reduction in build time, empirical accuracy metrics are still pending, and significant manual effort is required to verify the ontology.

Original post

Gul Saeed Khattak

@Ldataengineer · 3 November 2026

NVIDIA Open-Sources Auto Ontology, a Semantic Layer for AI Agents Your boss asks the AI agent: "What was revenue last quarter?" It answers in 3 seconds. Clean number. Nice chart. But it used the wrong "revenue" column. 𝗡𝗼𝗯𝗼𝗱𝘆 𝗻𝗼𝘁𝗶𝗰𝗲𝘀. That's the problem NVIDIA's

2.8k likes76 replies5.7k bookmarks235.7k views

View on X
More posts in the thread (10)
  1. @AlbertLath14811 @ShimazuSystems Could be. I only know Shimazu for TSUKUYOMI, their OSINT framework. Was it part of that? Send a link if you have one. To be fair, nobody here is first. Ontology-based access to SQL databases has been around for years (Ontop, R2RML mappings), and Snowflake and Databricks both
  2. @AlbertLath14811 @ShimazuSystems Also the goal was never the hard bit. Every data team wants agents to stop guessing columns. The hard bit is who signs off on what "revenue" means, and what happens when finance changes that definition next quarter. NVIDIA at least has a certified state and a Learn step for
  3. @seewonderai MCP is how the agent connects. It doesn't tell the agent which of your three revenue columns is the right one. Point an agent at raw Postgres over MCP and it will still guess from column names. Auto Ontology sits on the other end of that connection and holds the business terms
  4. @vioinabox The worse version is the boss never notices and the number ends up in a board pack. A bad join that doubles rows still gives you a perfect-looking chart. If you use an agent for reports, make it show the SQL every time and check which table it hit. It takes about 30 seconds.
  5. @KedemRam Yes. NVIDIA's own eval harness runs on BIRD, which uses public databases. That tells you nothing about your column names. I'd also count the hours your team spends certifying the ontology, since the 90% claim is about build time and no accuracy figures are out yet. Then run the
  6. @muayyadalsadi Agree, comments are the cheapest win there is, and any agent that reads the schema gets them for free. But look at your own example. "Datetime of record insert" tells me what dt is. It doesn't tell an agent whether "orders last quarter" should use dt or the order date. A
  7. @Sam4Jef Agree, that's the bit I would audit too. The README says it returns provenance with every answer, but doesn't say what's in it. You can get close yourself by logging the SQL from every ask_question call and match it to the term's SQL attributes with get_term_sql_attributes.
  8. @nickryg .added
  9. @MandzukRegan Yup it is but worth it 😊
  10. @finbobbb There is retrieval in it, to be fair. It does use embeddings and a reranker. The difference is what happens next. Basic RAG finds text that looks similar and hopes the model picks right, and to an embedding revenue_gross and revenue_net can look almost the same. Here the
Selected replies (5)
  • @Sam4Jef @Ldataengineer the wrong revenue column is the oldest bug in data lol, agents just ship it faster with a nicer chart. does auto ontology also log which definition it used per answer? thats the part id want to audit
  • @xarielx506 @Ldataengineer No classes, hierarchy, logical constraints or reasoning, so is this an actual ontology?
  • @AlbertLath14811 @Ldataengineer Huh. Pretty sure @ShimazuSystems had this running... Late last year? Same goal and everything.
  • @333nsab @Ldataengineer t.co/SHTZ5KWfLw
  • @vioinabox @Ldataengineer my boss is definitely going to notice eventually 💀

Mentioned links

Summary generated by AI (Gemini 3.8 Flash, 4 November 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Repository:NVIDIA/SkillSpector

    SkillSpector: Security Scanner for AI Agent Skills by NVIDIA

    SkillSpector is an open-source security scanner from NVIDIA designed to inspect extensions and skills for AI agents (such as Claude Code, Codex CLI, and MCP) for vulnerabilities and malicious patterns before installation.

    15.9kstarsPython

    AI & LLMs· Tool

  • Repository:NVIDIA/OpenShell

    NVIDIA OpenShell: Isolated Runtime for Autonomous AI Agents

    NVIDIA OpenShell is an open-source runtime written in Rust designed to run autonomous AI agents in secure sandboxes. It enforces declarative policies at the kernel level and guards access to network routes and credentials.

    16.6kstarsRust

    AI & LLMs· Tool

  • Article:Chris Alexiuk and Chintan Patel

    NVIDIA Nemotron 3.5 Lightning: Compact MoE Model for Agentic Execution

    NVIDIA has announced Nemotron 3.5 Lightning, an open 30B mixture-of-experts (MoE) model featuring 3B active parameters. It is tailored for the high-volume execution layer of long-running AI agents, targeting routine tasks such as tool calling and result validation with low latency.

    AI & LLMs· Announcement

  • Repository:GoogleCloudPlatform/knowledge-catalog

    Google Cloud Knowledge Catalog: Tools and Samples for Semantic AI Agents

    The GoogleCloudPlatform/knowledge-catalog repository gathers tools, agents, and sample code for Google Cloud's Knowledge Catalog (formerly Dataplex). The platform functions as an AI-powered data catalog and metadata management system, organising structured and unstructured data into a dynamic knowledge graph to deliver business context to AI agents.

    9.1kstarsTypeScript

    AI & LLMs· Tool

  • Link:Tawkit, Inc.

    OpenBot: Self-Hosted Enterprise Agent Platform Built on the AG-UI Protocol

    CopilotKit (Tawkit, Inc.) has introduced OpenBot, an open-source platform for enterprise AI agents designed to run entirely within an organisation's own infrastructure. It integrates internal company data, supports computer use, and leverages the AG-UI standard.

    AI & LLMs· Tool

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.