Skip to navigation
About Kumo Relational

Overview

View as Markdown

Kumo Relational is a pre-trained relational foundation model: a model designed to make predictions from connected tables and examples supplied with each request. It works with structured, multi-table data without requiring you to train a separate model for each prediction task.

For each request, Kumo Relational uses table relationships and relevant historical rows as context for the prediction. You describe the target and entities to predict with Predictive Query Language (PQL), a lightweight SQL-like language for prediction tasks, and receive the results as a pandas DataFrame through the Python SDK.

You access Kumo Relational through the NVIDIA Kumo Relational Client Python package (kumo-relational-client). The package connects to an NVIDIA NIM, the deployed inference service that hosts Kumo Relational and exposes its endpoint.

What you can predict

Choose the prediction type that matches the output you need:

Classification and Regression

Predict categories, Boolean outcomes, and numeric values.

Forecasting

Predict numeric values over a future time horizon.

Link Prediction

Rank or recommend related entities.

Get Started

If you are new to Kumo Relational, install and connect the client first, and then run the quickstart.

Install and Connect

Install kumo-relational-client[relational] and connect to a Kumo Relational NIM.

Run Your First Prediction

Build a graph from related tables, bind it to the client, and run a prediction.

Examples and Notebooks

Explore end-to-end predictive workflows and datasets.

API Reference

Look up the Python SDK, PQL, result schema, and NIM REST API.