NVIDIA presents Kumo Relational predictions from historical database examples

NVIDIA’s October 21 announcement describes Kumo Relational predictions for tasks such as churn, demand and fraud. The model uses up to 10,000 historical examples as context rather than requiring training for each task or flattening related tables into one. Meaningful relationships and time structure matter: the approach depends on useful signals in that connected data.

The current client prepares a request, sends it to a running NVIDIA NIM inference service and returns predictions convertible to a pandas table. The client itself needs no GPU and downloads no model weights; inference happens in the service. Users select KumoRelational as a model argument for connected tables. The relational installation includes a native engine requiring a matching platform wheel. There is no source-distribution fallback on unsupported systems, and installing the client does not supply the running inference service.

The quickstart illustrates predicting thirty-day order counts for selected users. Before prediction, a separate materialization call resolves the query, finds historical labelled examples, samples related records and encodes them. Entity selection is explicit; the query does not silently score every row. Features precede an anchor time, while labels cover the following window. A thirty-day query therefore needs history comfortably longer than thirty days; shorter data can leave no labelled examples. The window excludes the anchor itself but includes day thirty, a distinction that matters when checking the same aggregate.