> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/sdgm/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server.

# Kumo Relational Client

> Use the Kumo Relational Client Python library to prepare relational prediction requests and run inference through a Kumo Relational NIM.

The [**NVIDIA Kumo Relational Client**](https://github.com/kumo-ai/kumo-relational-client)
(`kumo-relational-client`) prepares relational prediction requests and submits
them to a **Kumo Relational NIM** endpoint. Its `RelationalClient` API binds a
relational graph to the endpoint and returns predictions as pandas DataFrames.

The client supports the `kumo-relational` model. Refer to
[Supported Models](/client/kumo-tabular) for supported execution paths.

## How the Client Works with NIM

Data preparation runs in the client process; model inference runs in the NIM:

| Component                                                         | Role                                                                                                                        |
| ----------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------- |
| Kumo Relational Client                                            | Manages the endpoint connection, prepares prediction requests, and returns results to your Python application.              |
| Relational engine, installed with the client’s `relational` extra | Builds the relational graph, processes Predictive Query Language (PQL) queries, and samples related records for prediction. |
| Kumo Relational NIM                                               | Hosts the model and runs inference on the supplied prediction context.                                                      |

Configure a reachable NIM URL before calling `predict()`. The NIM deployment
provides the GPU and model weights; the client process needs neither. Deploy
the NIM separately or connect to an existing compatible endpoint.

Explore available predictive AI NIMs, model details, and deployment options in
the [NVIDIA API Catalog](https://build.nvidia.com/explore/predictive-ai).

For endpoint setup and authentication requirements, follow
[Deploy, Install, and Connect](/kumo-relational/get-started/sdk-getting-started).

## Get Started

#### [Install and Connect](/kumo-relational/get-started/sdk-getting-started)

Set up the endpoint connection and install the client and optional connectors.

#### [Run the Quickstart](/kumo-relational/get-started/rfm)

Build a graph from sample tables and make your first prediction through NIM.

#### [Kumo Relational NIM](/kumo-relational/about-kumo-relational/overview)

Review supported prediction tasks and the model execution flow.

#### [Python API Reference](/kumo-relational/reference/python-api-reference)

Look up client connections, graph handles, predictions, and errors.

## Run Models Directly in Python

To load pretrained weights and run inference in the same Python process, use
the [Structured Data Models library](/structured-data-models/overview).