Enterprise agents have a reliability problem. They can retrieve documents and query databases, but when a revenue metric comes back wrong, speed is worthless. NVIDIA's Auto Ontology attacks this by inserting a semantic layer — an ontology built from enterprise metadata — between the agent and the data. The idea: tell the agent what data means, where it lives, how to use it, and what lineage it carries before it generates a single SQL statement. The library discovers enterprise ontologies from existing metadata, reconciles inconsistent definitions across systems (the same metric named three different things in three databases), and builds an AI-ready knowledge graph stored in Postgres with pgvector. It then exposes this context to any agent harness through MCP, CLI, and API. Natural-language questions get translated to SQL via NVIDIA NIM, executed against connected databases, and returned with provenance. The stack is conventional but competently assembled: a Next.js 16 frontend with Better Auth and Prisma, a FastAPI backend running LangChain for NL-to-SQL, a separate ingestion worker writing embeddings, and Postgres as the central store for metadata, catalog, ontology, and vectors. Optional HashiCorp Vault integration handles credential storage. Deployment is Helm on Kubernetes or Docker Compose locally. The NIM dependency is the structural tell. Default models are Nemotron-3-Nano-30b for reasoning, Nemotron-3-Embed-1b for embeddings, and Llama-Nemotron-Rerank for retrieval — all served from NVIDIA's hosted API catalog or self-hosted NIMs. Every model role is configurable via environment variables, and the triplet system (endpoint, key, model) allows mixing providers. But the path of least resistance is NVIDIA's inference stack, and that's the point. The Apache 2.0 license is genuinely open, and the architecture documentation is thorough — connection management, RBAC, API tokens, SSE-streamed conversation threads with conversationid continuity, all documented to deployment-ready detail. But NVIDIA is explicitly not accepting external contributions. This is open-source as distribution strategy, not as community project. The code is a gift; the roadmap is NVIDIA's. Benchmarks are pointed at the BIRD text-to-SQL benchmark via a separate evaluation repository, which is the right target — BIRD is the community standard for complex enterprise SQL generation. The validation story is incomplete from the README alone (no published numbers in this source), but the benchmark choice signals seriousness about the problem. The twenty-year question is whether ontology layers become infrastructure or remain integration glue. If agents become the primary interface to enterprise data, whoever owns the semantic layer between agent and database owns the context — and context is where reliability lives. NVIDIA is betting that NIM becomes the default inference substrate, and Auto Ontology is the hook that makes enterprise data flows dependent on it.
