NVIDIA Kumo Relational

Founded 2021

Make predictions directly from connected database tables using graph foundation models

You can use Kumo Relational to generate predictions directly from connected relational tables without building custom machine learning models for each task. The system uses relational foundation models that understand table relationships and historical data context. You define your prediction targets and entities using Predictive Query Language, a simple query syntax similar to SQL. The tool returns results directly into pandas DataFrames via a Python client. It handles common machine learning workflows including classification, numerical regression, time-series forecasting, and entity link predictions across complex enterprise databases.

What NVIDIA Kumo Relational does

Query relational schemas using Predictive Query LanguageRun tabular regression and classification tasksForecast numeric metrics across future time horizonsConnect tabular graphs to NVIDIA NIM inference endpointsExtract prediction outputs directly into pandas DataFramesGenerate link predictions between related database entitiesRelational foundation model trained on graph representations of tablesPredictive Query Language for writing prediction jobs in SQL-like syntaxDirect prediction export into pandas DataFrames through the Python clientSupport for classification, regression, forecasting, and entity link predictionsDirect connectivity to NVIDIA NIM inference endpointsNo requirement to train separate machine learning models per prediction task