NVIDIA/auto-ontology
Auto Ontology is a Python framework for building semantic, graph-based reasoning layers over data. It enables deterministic query generation, better context retrieval, and more explainable AI for LLM and agent workflows.
What it solves
Agents can retrieve documents and query databases, but reliability gates agentic analytics: a revenue metric answered faster is worth nothing if the number is wrong. Agents need a context layer that tells them where data lives, what it means, whether to use it, how to use it, where it came from, and what is similar. Auto Ontology builds that layer from enterprise metadata, reconciles inconsistent definitions across systems, and exposes it to any agent harness through MCP, CLI, and API. It complements the semantic models customers already own.
How it works
Auto Ontology works in two stages: it generates an ontology that people certify, and then the agent answers questions from it.
Generation. The ingestion service reads tables, columns, keys, views, and query history from source databases, embeds them in pgvector, and uses models to profile values, find foreign keys, name business terms, and suggest SQL attributes. It groups copies of schemas into term families, links foreign keys when the data confirms them, and produces a draft ontology. People refine and certify terms, descriptions, and SQL attributes in the UI or via the API. The result is a semantic graph with three kinds of definitions: terms (business concepts mapped to columns), SQL attributes (reviewed SQL expressions reused consistently), and relationships (how tables join).
Answering. For each question, the agent classifies intent, retrieves candidate terms and columns via embedding search, computes join paths from the ontology's semantic foreign keys (no model involved), writes SQL with a reasoning model, then checks and repairs it with static checks and a model judge. The answer comes back with its SQL, the steps taken, and provenance, so users can validate it. If no valid query can be built, the agent reports the question can't be answered from this data.
The system runs as a stack: a Next.js frontend (web UI, auth, API gateway), a FastAPI backend (text-to-SQL agent, catalog APIs), an ingestion worker, Postgres+pgvector for storage, an optional HashiCorp Vault for credentials, and a FastMCP-based MCP server that exposes read-only tools to any MCP client. Models are configured via environment-variable triplets (endpoint, API key, model) and default to NVIDIA NIM endpoints.
Who it's for
- Developers building agentic analytics or chat interfaces over enterprise databases.
- Data teams that want to standardize business definitions and query logic across systems.
- Teams that already have semantic models or want to adopt Apache Ossie for interchange.
- Anyone who needs reliable, provenance-backed natural-language question answering over structured data.
Highlights
- Natural-language querying of structured data, translated to SQL and executed against connected databases.
- Automatic ontology generation from metadata, with human-in-the-loop certification.
- Reconciliation of inconsistent definitions across systems.
- Validation of generated queries against certified definitions and lineage; answers include provenance.
- Exposed to any agent harness via MCP, CLI, and API.
- Web UI for chat, analysis, data catalog, analytics, and settings.
- Flexible deployment with Helm (Kubernetes) or Docker Compose.
- Compatible with Apache Ossie for importing/exporting semantic models as YAML.
- Model configuration via triplets with fallbacks, supporting NVIDIA NIM or OpenAI-compatible endpoints.
- Optional PII detection, predictive queries via KumoRFM, and connection credential storage in Vault.
Related
- Project
- Project
- Dispatch
- Project
- Project