NVIDIA Open-Sources Its Semantic Layer: AI Agents Verify Metric Definitions Before Writing SQL

NVIDIA has released Auto Ontology, an open-source library under the Apache 2.0 license. It has a very targeted function: generating a semantic layer from existing enterprise metadata, so that AI agents can access certified, authoritative definitions before constructing SQL queries. As reported by Asif Razzaq, author of MarkTechPost, this semantic layer exposes context to external systems through MCP, CLI and APIs.

Common pain points in enterprise data systems are straightforward: table names, field names and metric calculation rules are scattered across BI tools, data catalogs, internal documents and historical SQL scripts. For the same revenue metric, there could be 3 to 4 different calculation logics. Human analysts can always turn to colleagues for clarification, but AI agents have no choice but to guess. Even if the SQL generated by the agent is syntactically correct, wrong metric definitions will lead to far more troublesome misleading results. The semantic layer unifies all these scattered definitions into a single machine-readable layer. Auto Ontology takes the approach of automatic generation from existing metadata, to reduce the upfront cost of building a semantic layer completely from scratch.

MCP, short for Model Context Protocol, is a standard interface that provides tools and context for agents. Auto Ontology delivers context via MCP, CLI and APIs, which effectively connects the semantic layer directly into the invocation chain of AI agents. Agents look up standardized definitions before generating SQL, reducing the randomness caused by forcing models to memorize full data schemas by rote.

It is suitable for enterprises that already have data catalogs, BI tools or metric platforms, and also for teams that are developing text-to-SQL systems, data question-answering tools and analytical AI agents. Its value lies in pushing metric governance to the agent runtime stage, reducing unnecessary post-hoc error correction work. The Apache 2.0 license lowers the barriers for internal evaluation and secondary development.

It is worth noting clearly that auto-generating a semantic layer does not automatically produce 100% correct semantic outputs. Certified definitions, access permissions, data lineage and update mechanisms still require human management. This open-source library only solves the problems of high startup cost and inconvenient integration, the governance responsibilities for data have not disappeared. Semantic layers are evolving from resources that humans look up in documents, to native runtime context for AI agents. This direction is far more maintainable than the solution that forces large models to memorize entire data schemas.

Source: MarkTechPost, by Asif Razzaq, Original Article Link.