> This page is for Structured Data Models Client.

> 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.

# Glossary

> Definitions of Kumo Relational data, schema, prediction, and configuration terms

Use this glossary for terms that have a specific meaning in Kumo Relational and sdm-client. Follow the links for procedures, examples, and complete reference information.

## A

**Anchor time.** The reference time associated with a prediction or context example, passed to `materialize()`. Kumo Relational uses the time boundary to determine which historical events are available to the prediction. See [Configuration](/sdgm/kumo-relational/user-guide/configure-and-scale/configuration).

## C

**Candidates.** The set of rows a ranking query is asked over, passed to `materialize(candidates=...)`. Required for `RANK TOP K` — no ranking task assumes a candidate set. See [Prediction Types](/sdgm/kumo-relational/user-guide/make-predictions/prediction-types#recommendations-ranking).

**Context.** Rows where the answer is already known, gathered by `materialize()` from the schema's historical data. See [How Kumo Relational Works](/sdgm/kumo-relational/about-kumo-relational/introduction).

## E

**Entity table.** A table representing objects for which Kumo Relational can make predictions. An entity table typically has a stable, unique primary key.

**Event table.** A table containing time-dependent activities or transactions related to one or more entity tables. Its time column identifies when each event became available to the prediction workflow.

**Explanation.** The structured attribution payload returned in `result.explanation` when `outputs` includes `'explanation'`, where the deployment produces one. Many shipping deployments refuse this output entirely. See [Prediction Explainability](/sdgm/kumo-relational/user-guide/explain-predictions/prediction-explainability).

## F

**Fan-out.** The maximum number of neighbors sampled at one graph hop. Each value in `num_neighbors` sets the fan-out for the corresponding hop.

**`fit`.** Pins a materialized request's context on the endpoint and returns a `FittedModel` for scoring several questions against it without re-uploading. Takes an optional `holdout=` naming every row the session will later score. See [Make Predictions](/sdgm/kumo-relational/user-guide/make-predictions/make-predictions).

**Foreign key.** A column whose values refer to the primary key of another schema table. A `Relationship` connects a foreign key to its target entity table.

## H

**Holdout.** The `fit(holdout=...)` argument naming every entity key a session will go on to score, so those rows can be excluded from the context pinned at `fit()` time. Without it, a later call can score a row that is sitting in the pinned context with its own answer attached.

**Hop.** One relationship traversal in the relational neighborhood around a prediction entity. The number of entries in `num_neighbors` determines the number of sampled hops.

## I

**`id`.** The entity's primary-key value on a result row, present when the model has an entity key. Join on this, or on `row_index` when there is none. See [Prediction Results](/sdgm/kumo-relational/user-guide/work-with-prediction-results/prediction-results).

**`indices`.** Entity primary-key values passed to `materialize()`. They select the entities to score, up to 100,000 per call. See [Make Predictions](/sdgm/kumo-relational/user-guide/make-predictions/make-predictions).

**Inference configuration.** Task-specific model settings passed through `inference_config` on `materialize()`. See [Configuration](/sdgm/kumo-relational/user-guide/configure-and-scale/configuration#inference-configuration).

## K

**`kumo-relational`.** The model id sent on the wire and reported by a NIM's `/v1/models`. Named through the `KumoRelational()` handle in sdm-client.

**`KumoRelational`.** The sdm-client handle for the relational model. Passed as the `model=` argument to `predict()`, `fit()`, and `materialize()`.

## N

**Neighborhood.** Relational context sampled around a prediction entity. Its depth and maximum fan-out are controlled by `num_neighbors`.

**NIM.** A deployed inference service that hosts Kumo Relational and exposes its endpoint. sdm-client connects to the NIM and sends prediction requests; it does not deploy or run the model itself. See [Deploy, Install, and Connect](/sdgm/kumo-relational/get-started/sdk-getting-started).

## P

**Predictive Query Language (PQL).** The language used to describe a predictive task over a schema. A runnable Kumo Relational query identifies a target and the entity or entities to score. See [Write Predictive Queries](/sdgm/kumo-relational/user-guide/make-predictions/writing-predictive-queries).

**Prediction entity.** The entity for which Kumo Relational returns a prediction. In PQL, the entity is identified by the table key in the `FOR` or `FOR EACH` clause.

**Prediction horizon.** The future interval covered by a temporal prediction target, expressed by the start, end, and time unit in a PQL aggregation.

**`prediction_statistic`.** A `materialize()` option choosing what a regression's `prediction` column reports of the model's predicted distribution: `"median"` (default) or `"mean"`. See [Prediction Results](/sdgm/kumo-relational/user-guide/work-with-prediction-results/prediction-results#what-a-regression-reports).

**`PredictionResult`.** The object every `predict()` and `fit().predict()` call returns. `to_pandas()` gives the DataFrame; `explanations` carries gradients and column cohorts when requested and available; `unscored_candidates` lists any ranking candidates left out. See [Prediction Results](/sdgm/kumo-relational/user-guide/work-with-prediction-results/prediction-results).

**Primary key.** A non-null, unique column (or ordered set of columns, for a composite key) that identifies each row in a schema table.

## R

**Relationship.** A schema connection from a foreign key in an event or junction table to the primary key of an entity table. See [Define Relationships](/sdgm/kumo-relational/user-guide/build-a-relational-graph/graph-definitions).

**`Request`.** What to ask, independent of which model answers it. For Kumo Relational, built only through `PredictiveQuery(...).materialize(...)`. See [Make Predictions](/sdgm/kumo-relational/user-guide/make-predictions/make-predictions).

**`row_index`.** The zero-based position of a scored row within the whole (possibly batched) call, carried by every result row.

**`rows_per_call`.** A `materialize()` option lowering how many entities one NIM call scores, below the endpoint's own cap. Only ever lowers it. See [Batch Prediction](/sdgm/kumo-relational/user-guide/configure-and-scale/batch-prediction).

## S

**`Schema`.** The tables, table metadata, and relationships sdm-client uses to sample predictive context. Built with `from_data()` (in memory) or `from_source()` (sampled in place). One table with no relationships is the same type Kumo Tabular uses. See [Build a Schema](/sdgm/kumo-relational/user-guide/build-a-relational-graph/graph-creation).

**`SDMClient`.** The public sdm-client class that owns the HTTP connection to one NIM endpoint. Serves both Kumo Relational and Kumo Tabular. See [Python SDK API](/sdgm/kumo-relational/reference/python-api-reference).

**Semantic model.** A Databricks metric view or Snowflake semantic view, converted into a `Schema` with `Schema.from_source(source.semantic_model(name))`.

**Semantic type (`stype`).** Metadata describing how sdm-client should interpret a column, such as `ID`, `numerical`, `categorical`, `text`, or `timestamp`. Assigned as a plain string. See [Data Types and Semantic Types](/sdgm/kumo-relational/user-guide/prepare-data/data-types).

**`Source`.** Describes where a schema's tables live and how to reach them, for `Schema.from_source()`: `DatabricksSource`, `SnowflakeSource`, `PostgresSource`, `DuckDBSource`, or `SQLiteSource`.

## T

**Target.** The value or event to predict. In PQL, the target can be a column, a future aggregation, or a condition applied to an aggregation.

**Task kind.** The prediction category: `classification`, `binary_classification`, `multiclass_classification`, `regression`, `forecasting`, or `temporal_link_prediction`. See [Prediction Types](/sdgm/kumo-relational/user-guide/make-predictions/prediction-types).

**Time column.** Table metadata identifying when an event became available to the prediction workflow. There is no supported `end_time_column`. See [Time Columns](/sdgm/kumo-relational/user-guide/prepare-data/time-columns).

## U

**`unscored_candidates`.** On a ranking's `PredictionResult`, the candidates no entity's sampled context could answer, left out of every row. See [Prediction Results](/sdgm/kumo-relational/user-guide/work-with-prediction-results/prediction-results#a-ranking-can-hold-fewer-than-k-candidates).