ICLEstimator

ICLEstimator runs zero-shot in-context learning (ICL) inference. Unlike GNNEstimator, ICLEstimator has no training loop, no experiment tracking, and no model artifact — there is nothing to fit() or register. You construct it once and call predict() or score() directly; both run the full pipeline (feature generation + ICL inference) in a single job.

Every ICL job needs a context table: a table of already-labeled examples used directly as in-context examples. There is no analogue of a training loop that “learns” from this data ahead of time — the context table is read at inference time, every time you call predict() or score().

Two dataset kinds

ICLEstimator accepts either of the two dataset kinds already used across the SDK, and picks its pipeline based on which one you pass in:

Dataset kind	Pipeline	Supported task types
RelationalDataset (multi-table, with a NodeTask)	DFS (featuretools) generates cross-table features (e.g. a customer’s past orders), then ICL runs on the result. Depth is controlled by ICLConfig.dfs_max_depth.	binary_classification, multiclass_classification, regression
TabularDataset (single flat table, TabularTask)	Flat pipeline — no DFS, no cross-table joins. The table is fed to ICL as-is.	binary_classification, multiclass_classification, regression, forecasting

Only TabularDataset supports forecasting — there is no relational forecasting path (RelationalDataset has no analogous _forecast() handling), and neither dataset kind supports multilabel_classification or link prediction with ICLEstimator.

Note

ICLEstimator requires task.label_column to be set on every task, for both dataset kinds. Because there is no training loop, the model has nothing else to learn from — the context table’s labels are the only source of supervision. This is different from NodeTask, where label_column=None is valid for an inference-only task passed to GNNEstimator.

Parameters

Name	Type	Description	Optional
connector	SnowflakeConnector	The connector object used for sending requests to the engine.	No
config	ICLConfig	Configuration object containing inference knobs. Defaults to ICLConfig().	Yes

Returns

An instance of the ICLEstimator class.

Example — relational dataset

from relationalai_predictive import ICLEstimator, ICLConfig

model = ICLEstimator(connector=connector, config=ICLConfig())
job = model.predict(relational_dataset)

Example — flat / forecasting dataset

from relationalai_predictive import ICLEstimator, ICLConfig, TabularTask, TabularDataset, TaskType

task = TabularTask(
    name="sales_forecast",
    task_type=TaskType.FORECASTING,
    label_column="sales",
    context_table="DATABASE.SCHEMA.SALES_CONTEXT",
    test_table="DATABASE.SCHEMA.SALES_TEST",
    time_column="date",
    prediction_length=12,
    max_context_length=64,
)
dataset = TabularDataset(connector=connector, name="sales", task=task)

model = ICLEstimator(connector=connector)
job = model.predict(dataset)

Methods

• predict()
• score()