NVIDIA Auto Ontology: semantic layer review
/auto-ontology
NVIDIA Auto Ontology is a serious self-hosted semantic layer for teams tired of unreliable agent-generated SQL. It is worth operating when certified definitions, source lineage, and answerability checks justify the added infrastructure.
Why agent data queries need context
Enterprise agents can query databases quickly without understanding whether a metric means the same thing in every system. NVIDIA Auto Ontology addresses the missing business context: approved terms, reviewed SQL expressions, table relationships, lineage, and whether the connected data can answer a question at all.
What Auto Ontology does
NVIDIA Auto Ontology is a self-hosted semantic layer that compiles database metadata into a knowledge graph. It generates terms, relationships, and SQL attributes, lets people certify them, then retrieves them to answer questions with checked SQL. The resulting context can be exposed through a web UI, REST API, CLI, or MCP.
Core capabilities and interfaces
- โCompiles relational metadata into terms, join relationships, and reusable SQL attributes stored with pgvector embeddings.
- โRuns an answerability check before querying, then repairs SQL when validation against the source data fails.
- โReturns answers with the generated SQL and the steps used, giving reviewers a concrete audit trail.
- โOffers a web interface for chat, semantic catalog review, analytics, connection settings, and model import or export.
- โExposes read-only ontology operations through MCP, question answering through the API, and local operations through the CLI.
- โConnects to PostgreSQL, MySQL, Databricks, Snowflake, ClickHouse, Trino, Apache Kyuubi, HeavyDB, DuckDB, and SQLite.
- โImports and exports semantic definitions as Apache Ossie YAML.
Enterprise analytics and agent workflows
- โขTurn recurring analytics questions into queries that reuse certified revenue, margin, and other SQL definitions.
- โขLet an agent determine that a requested metric is missing before attempting an unsupported calculation.
- โขGive analysts an interface for reviewing terms, column mappings, relationships, and compiled SQL attributes.
- โขExpose governed enterprise context to MCP-compatible agents while preserving user permissions.
- โขTrace an answer back to its SQL, mapped source columns, ontology terms, and source data.
- โขUnify terminology across connected relational systems through a shared semantic catalog and knowledge graph.
Run the full stack locally
Run the full stack locally with Docker Compose, Compose v2, uv, and an NVIDIA API key. Clone the repository, copy `.env.example` to `.env`, set the model key, application secrets, admin password, Postgres password, and sample database connection string, then start Postgres and seed Wide World Importers from the separate `auto-ontology-eval` repository. Start Postgres first. Build and launch the remaining services with `docker compose up -d --build`, open `http://localhost:3000`, sign in with the configured admin credentials, and enable Semantic Compilation under Settings.
Strengths
- โPuts reviewed business definitions, table relationships, and SQL expressions into one inspectable semantic graph.
- โMakes uncertainty visible by checking whether connected sources cover a question before producing an answer.
- โShows SQL and execution steps, making results easier for analysts to verify than an unexplained natural-language response.
- โSupports both interactive certification and programmatic access through the API, CLI, and MCP.
- โPreserves semantic-layer intent across several relational database engines.
Operational, privacy, and security limits
- โณThe complete stack requires Postgres, Docker, model endpoints, configuration, and a separate evaluation repository for the documented sample-data setup.
- โณThe backend trusts the user identity forwarded in `x-auto-ontology-user-id`, while ingestion endpoints have no authentication of their own; ports 3001 and 3002 must remain on a private network.
- โณUsing KumoRFM can upload connected tables to a hosted service, so teams must review the transferred data and configure the table and row limits before enabling predictions.
- โณThe project uses the Apache-2.0 license but currently does not accept external contributions; integration requests and bug reports are handled through issues.
- โณThe deployment is operationally heavier than a plain Python package because it includes a frontend, backend, ingestion worker, MCP server, Postgres, and optional Vault.
Where Auto Ontology fits
Frequently asked questions
NVIDIA Auto Ontology is a self-hosted semantic layer that compiles database metadata into terms, relationships, and reviewed SQL attributes. It stores this context with pgvector embeddings and lets agents query the resulting knowledge graph through a UI, API, CLI, or MCP.
NVIDIA Auto Ontology gives agents business definitions, table joins, reusable SQL expressions, data provenance, and answerability checks. The agent retrieves this context, generates SQL, validates it against source data, and returns the answer with the SQL and steps taken.
NVIDIA Auto Ontology supports PostgreSQL, MySQL, Databricks, Snowflake, ClickHouse, Trino, Apache Kyuubi, HeavyDB, DuckDB, and SQLite. DuckDB and SQLite are configured only through connection strings.
Run the NVIDIA Auto Ontology MCP server against a deployment, connect an MCP client to its endpoint, and sign in. The agent can check readiness, search ontology terms, test answerability, ask questions, and receive the answer, SQL, and rows.
Yes. Under the Apache License 2.0, NVIDIA Auto Ontology is distributed. The README additionally indicates that bundled, linked, or otherwise used third-party open-source components have a separate listing in `THIRD_PARTY_NOTICES.md`.
No. NVIDIA Auto Ontology currently does not accept external pull requests. The maintained code is managed by the project team, which invites integration requests and bug reports through GitHub issues.
Who should try it โ and who should skip
Try NVIDIA Auto Ontology if a team operates important relational data and needs agents to distinguish certified business meaning from a plausible but incorrect query. Skip it if the use case only needs an informal chat interface over one small database, or if the organization cannot operate Postgres, secure the service boundary, and maintain certified semantic definitions.
Curious whether auto-ontology is right for you?
Let ChatGPT, Claude, or Perplexity look into it โ click below and see what AI actually says about auto-ontology.