Nvidia – open-source auto-ontology
Points and comments are a snapshot, not live.
NVIDIA's open-source Auto-Ontology framework adds semantic context to help agents generate fewer incorrect SQL queries.
Auto-Ontology is a Python service that compiles database metadata into an ontology of business terms, SQL attributes, and table relationships. Agents query through this ontology rather than raw schemas, enabling deterministic SQL generation, answerability checking, and explainable AI. The framework supports ten database engines and can be accessed via MCP, REST API, or agent skills. NVIDIA provides evaluation results comparing ontology-grounded agents against schema-in-prompt baselines across multiple benchmarks.
What commenters are saying
Commenters debated the reliability of using LLMs to build ontology layers for other LLMs, with some arguing that combining unreliable systems can compound errors rather than fix them. Others countered that reliable systems have been built atop unreliable components before (e.g., TCP over IP, biological chemistry). A middle view held that the approach improves reliability by separating ontology generation from query execution, allowing human review and deterministic SQL generation. Critics remained skeptical that LLMs can be constrained to disregard their training data in favor of provided context.