Image: Nvidia
Why it mattersA team putting an AI agent on its own data can try an Apache-licensed grounding layer from a major vendor. Compare it against ad-hoc RAG prompts.
An AI agent that queries a company's database still gets the business meaning wrong. It joins the wrong table, picks the wrong revenue column or counts refunds as sales. Nvidia released auto-ontology this week, an Apache-licensed Python framework that gives the agent a trusted map of the data. The repository drew 2,393 stars by 3 November and reached rank 4 on Trendshift's daily list. A Hacker News thread about it rose to 242 points in 11 hours.
What the framework does
Nvidia says auto-ontology "enables deterministic query generation, better context retrieval, and more explainable AI for LLM and agent workflows". The framework splits the work into five steps. It discovers the shape of the data from database metadata and builds a knowledge graph. It grounds the agent in that graph. It validates the SQL the agent writes against the graph's definitions. It exposes the graph to the agent through an MCP endpoint, a CLI and an API. And it folds a human reviewer's corrections back into the graph.
The product does three concrete jobs. It lets a user ask a question in plain English and get a SQL query back. It serves a semantic catalog of the company's metrics and entities. It is a reference implementation of what Nvidia calls a semantic catalog layer over relational sources.
What runs with it
The repository supports four source databases: Databricks, Postgres, Snowflake and DuckDB. It carries a web UI for chat and analysis. Authentication is handled through Better Auth. Two deployment paths are offered: a Kubernetes Helm chart, and a Docker Compose file for a single machine.
The one hard dependency is an Nvidia API key. The README asks for a key from build.nvidia.com. Without it the framework will not call a model. A team that already runs its own model gateway has to add Nvidia's hosted inference on top. The repository is marked as not accepting external contributions.
Why it is a story
Three signals arrived at the same time. The repository was created on 27 April 2026. Its star count sat in the hundreds until this week, when it reached 2,393. It landed on Hacker News on 3 November and reached 242 points in 11 hours. It also ranks fourth on Trendshift's daily list. The 183 forks are a stronger signal than the stars: a fork is usually a team trying the tool on its own data.
The underlying problem is the one most teams building with retrieval-augmented generation hit second. The first version works on a demo. The second version joins the sales table to the wrong month. The reason is nobody told the model which column holds the booking date. A grounding layer is one answer. There are others, including hand-written tool descriptions and a hand-maintained metrics layer. Auto-ontology is one way to try the pattern in a day.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.