Kumo Engine Overview
Kumo Engine is an accelerator for context generation over structured data, designed for teams building and running predictive models in production at scale. Like DuckDB and Polars for analytics, it is purpose-built for preparing the data models consume, combining fast context retrieval with time-correct sampling for training and batch inference.
At two to four hops, Kumo Engine’s warm-cache runs generated context for one million entities 10 to 48× faster than prepared DuckDB and 9.5 to 22× faster than Polars on GPU. Its Rust-based engine and NVMe storage delivered this performance on a 2.6 TiB graph index using a machine with just 1 TiB of RAM. Kumo Engine gives developers and coding agents a powerful foundation for ingestion and large scale prediction.
What You Can Build
Different tasks can reuse the same connected data while defining their own targets, training examples, and evaluation criteria. Check out more use cases and examples in Predictive Query Examples.
Why Build on Kumo
Use Relationships and History as Model Inputs
A customer’s orders, the products in those orders, and the timing of each purchase provide context beyond a single customer row. A total-spend feature, for example, leaves out which products a customer bought and in what order. Kumo’s graph-learning workflow learns from the connected records you select, using the history available at each prediction time.
Reuse the Data Foundation as Experiments Change
Test whether more context, a different encoder, or a larger model helps your task by changing a model plan. Kumo handles context retrieval and sampling for each run, while you compare configurations using the same training examples and validation splits. For a new prediction target, generate historical examples from the same connected data.
Iterate with Coding Agents
Ask your coding agent to compare two sampling depths within a fixed training budget. It can update the model plan, run both configurations, and use training diagnostics to summarize validation scores, learning curves, and runtime. Kumo handles the data loading and temporal sampling, so the agent can move directly from a modeling question to an experiment. Use the results to choose the next change, such as testing a different encoder or a smaller model.
How It Works
DiskGraph, Kumo’s data engine built in Rust, indexes records and relationships on local SSD. It efficiently retrieves connected records and their features as model-ready batches, applying temporal constraints where relevant. The learning runtime consumes those batches for training and prediction; the Model Plan Guide covers configuration.
A predictive query specifies what to predict and for which entities, including a time window for temporal tasks. For example, “Which customers will buy in the next 90 days?” defines a purchase-propensity task with customers as the entities and purchases in that future window as the outcome. See Predictive Queries for the query and its Python workflow.
Your Environment, Your Models, Your Control
Run Kumo Engine on a supported Linux machine or in a Snowflake or Databricks notebook. Your credentials govern data access, you provision compute and storage, and model plans configure the modeling workflow. See the setup guidance and prerequisites for environment requirements.