kumoai 2.13.0.dev202512091732__cp311-cp311-macosx_11_0_arm64.whl → 2.14.0.dev202601051732__cp311-cp311-macosx_11_0_arm64.whl

kumoai/experimental/rfm/rfm.py

@@ -1,26 +1,24 @@
 import json
+import math
 import time
 import warnings
 from collections import defaultdict
-from collections.abc import Generator
+from collections.abc import Generator, Iterator
 from contextlib import contextmanager
 from dataclasses import dataclass, replace
-from typing import (
-    Any,
-    Dict,
-    Iterator,
-    List,
-    Literal,
-    Optional,
-    Tuple,
-    Union,
-    overload,
-)
+from typing import Any, Literal, overload
 
 import numpy as np
 import pandas as pd
 from kumoapi.model_plan import RunMode
 from kumoapi.pquery import QueryType, ValidatedPredictiveQuery
+from kumoapi.pquery.AST import (
+    Aggregation,
+    Column,
+    Condition,
+    Join,
+    LogicalOperation,
+)
 from kumoapi.rfm import Context
 from kumoapi.rfm import Explanation as ExplanationConfig
 from kumoapi.rfm import (
@@ -29,35 +27,35 @@ from kumoapi.rfm import (
     RFMPredictRequest,
 )
 from kumoapi.task import TaskType
+from kumoapi.typing import AggregationType, Stype
 
 from kumoai.client.rfm import RFMAPI
 from kumoai.exceptions import HTTPException
-from kumoai.experimental.rfm import Graph
-from kumoai.experimental.rfm.backend.local import LocalGraphStore
-from kumoai.experimental.rfm.local_graph_sampler import LocalGraphSampler
-from kumoai.experimental.rfm.local_pquery_driver import (
-    LocalPQueryDriver,
-    date_offset_to_seconds,
-)
+from kumoai.experimental.rfm import Graph, TaskTable
+from kumoai.experimental.rfm.base import DataBackend, Sampler
 from kumoai.mixin import CastMixin
-from kumoai.utils import InteractiveProgressLogger, ProgressLogger
+from kumoai.utils import ProgressLogger, display
 
 _RANDOM_SEED = 42
 
 _MAX_PRED_SIZE: dict[TaskType, int] = defaultdict(lambda: 1_000)
 _MAX_PRED_SIZE[TaskType.TEMPORAL_LINK_PREDICTION] = 200
 
+_MAX_TEST_SIZE: dict[TaskType, int] = defaultdict(lambda: 2_000)
+_MAX_TEST_SIZE[TaskType.TEMPORAL_LINK_PREDICTION] = 400
+
 _MAX_CONTEXT_SIZE = {
     RunMode.DEBUG: 100,
     RunMode.FAST: 1_000,
     RunMode.NORMAL: 5_000,
     RunMode.BEST: 10_000,
 }
-_MAX_TEST_SIZE = {  # Share test set size across run modes for fair comparison:
-    RunMode.DEBUG: 100,
-    RunMode.FAST: 2_000,
-    RunMode.NORMAL: 2_000,
-    RunMode.BEST: 2_000,
+
+_DEFAULT_NUM_NEIGHBORS = {
+    RunMode.DEBUG: [16, 16, 4, 4, 1, 1],
+    RunMode.FAST: [32, 32, 8, 8, 4, 4],
+    RunMode.NORMAL: [64, 64, 8, 8, 4, 4],
+    RunMode.BEST: [64, 64, 8, 8, 4, 4],
 }
 
 _MAX_SIZE = 30 * 1024 * 1024
@@ -95,24 +93,26 @@ class Explanation:
     def __getitem__(self, index: Literal[1]) -> str:
         pass
 
-    def __getitem__(self, index: int) -> Union[pd.DataFrame, str]:
+    def __getitem__(self, index: int) -> pd.DataFrame | str:
         if index == 0:
             return self.prediction
         if index == 1:
             return self.summary
         raise IndexError("Index out of range")
 
-    def __iter__(self) -> Iterator[Union[pd.DataFrame, str]]:
+    def __iter__(self) -> Iterator[pd.DataFrame | str]:
         return iter((self.prediction, self.summary))
 
     def __repr__(self) -> str:
         return str((self.prediction, self.summary))
 
-    def _ipython_display_(self) -> None:
-        from IPython.display import Markdown, display
+    def print(self) -> None:
+        r"""Prints the explanation."""
+        display.dataframe(self.prediction)
+        display.message(self.summary)
 
-        display(self.prediction)
-        display(Markdown(self.summary))
+    def _ipython_display_(self) -> None:
+        self.print()
 
 
 class KumoRFM:
@@ -151,20 +151,35 @@ class KumoRFM:
     Args:
         graph: The graph.
         verbose: Whether to print verbose output.
+        optimize: If set to ``True``, will optimize the underlying data backend
+            for optimal querying. For example, for transactional database
+            backends, will create any missing indices. Requires write-access to
+            the data backend.
     """
     def __init__(
         self,
         graph: Graph,
-        verbose: Union[bool, ProgressLogger] = True,
+        verbose: bool | ProgressLogger = True,
+        optimize: bool = False,
     ) -> None:
         graph = graph.validate()
         self._graph_def = graph._to_api_graph_definition()
-        self._graph_store = LocalGraphStore(graph, verbose)
-        self._graph_sampler = LocalGraphSampler(self._graph_store)
 
-        self._client: Optional[RFMAPI] = None
+        if graph.backend == DataBackend.LOCAL:
+            from kumoai.experimental.rfm.backend.local import LocalSampler
+            self._sampler: Sampler = LocalSampler(graph, verbose)
+        elif graph.backend == DataBackend.SQLITE:
+            from kumoai.experimental.rfm.backend.sqlite import SQLiteSampler
+            self._sampler = SQLiteSampler(graph, verbose, optimize)
+        elif graph.backend == DataBackend.SNOWFLAKE:
+            from kumoai.experimental.rfm.backend.snow import SnowSampler
+            self._sampler = SnowSampler(graph, verbose)
+        else:
+            raise NotImplementedError
 
-        self._batch_size: Optional[int | Literal['max']] = None
+        self._client: RFMAPI | None = None
+
+        self._batch_size: int | Literal['max'] | None = None
         self.num_retries: int = 0
 
     @property
@@ -182,7 +197,7 @@ class KumoRFM:
     @contextmanager
     def batch_mode(
         self,
-        batch_size: Union[int, Literal['max']] = 'max',
+        batch_size: int | Literal['max'] = 'max',
         num_retries: int = 1,
     ) -> Generator[None, None, None]:
         """Context manager to predict in batches.
@@ -216,17 +231,17 @@ class KumoRFM:
     def predict(
         self,
         query: str,
-        indices: Union[List[str], List[float], List[int], None] = None,
+        indices: list[str] | list[float] | list[int] | None = None,
         *,
         explain: Literal[False] = False,
-        anchor_time: Union[pd.Timestamp, Literal['entity'], None] = None,
-        context_anchor_time: Union[pd.Timestamp, None] = None,
-        run_mode: Union[RunMode, str] = RunMode.FAST,
-        num_neighbors: Optional[List[int]] = None,
+        anchor_time: pd.Timestamp | Literal['entity'] | None = None,
+        context_anchor_time: pd.Timestamp | None = None,
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
         num_hops: int = 2,
-        max_pq_iterations: int = 20,
-        random_seed: Optional[int] = _RANDOM_SEED,
-        verbose: Union[bool, ProgressLogger] = True,
+        max_pq_iterations: int = 10,
+        random_seed: int | None = _RANDOM_SEED,
+        verbose: bool | ProgressLogger = True,
         use_prediction_time: bool = False,
     ) -> pd.DataFrame:
         pass
@@ -235,17 +250,17 @@ class KumoRFM:
     def predict(
         self,
         query: str,
-        indices: Union[List[str], List[float], List[int], None] = None,
+        indices: list[str] | list[float] | list[int] | None = None,
         *,
-        explain: Union[Literal[True], ExplainConfig, Dict[str, Any]],
-        anchor_time: Union[pd.Timestamp, Literal['entity'], None] = None,
-        context_anchor_time: Union[pd.Timestamp, None] = None,
-        run_mode: Union[RunMode, str] = RunMode.FAST,
-        num_neighbors: Optional[List[int]] = None,
+        explain: Literal[True] | ExplainConfig | dict[str, Any],
+        anchor_time: pd.Timestamp | Literal['entity'] | None = None,
+        context_anchor_time: pd.Timestamp | None = None,
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
         num_hops: int = 2,
-        max_pq_iterations: int = 20,
-        random_seed: Optional[int] = _RANDOM_SEED,
-        verbose: Union[bool, ProgressLogger] = True,
+        max_pq_iterations: int = 10,
+        random_seed: int | None = _RANDOM_SEED,
+        verbose: bool | ProgressLogger = True,
         use_prediction_time: bool = False,
     ) -> Explanation:
         pass
@@ -253,19 +268,19 @@ class KumoRFM:
     def predict(
         self,
         query: str,
-        indices: Union[List[str], List[float], List[int], None] = None,
+        indices: list[str] | list[float] | list[int] | None = None,
         *,
-        explain: Union[bool, ExplainConfig, Dict[str, Any]] = False,
-        anchor_time: Union[pd.Timestamp, Literal['entity'], None] = None,
-        context_anchor_time: Union[pd.Timestamp, None] = None,
-        run_mode: Union[RunMode, str] = RunMode.FAST,
-        num_neighbors: Optional[List[int]] = None,
+        explain: bool | ExplainConfig | dict[str, Any] = False,
+        anchor_time: pd.Timestamp | Literal['entity'] | None = None,
+        context_anchor_time: pd.Timestamp | None = None,
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
         num_hops: int = 2,
-        max_pq_iterations: int = 20,
-        random_seed: Optional[int] = _RANDOM_SEED,
-        verbose: Union[bool, ProgressLogger] = True,
+        max_pq_iterations: int = 10,
+        random_seed: int | None = _RANDOM_SEED,
+        verbose: bool | ProgressLogger = True,
         use_prediction_time: bool = False,
-    ) -> Union[pd.DataFrame, Explanation]:
+    ) -> pd.DataFrame | Explanation:
         """Returns predictions for a predictive query.
 
         Args:
@@ -307,18 +322,133 @@ class KumoRFM:
             If ``explain`` is provided, returns an :class:`Explanation` object
             containing the prediction, summary, and details.
         """
-        explain_config: Optional[ExplainConfig] = None
-        if explain is True:
-            explain_config = ExplainConfig()
-        elif explain is not False:
-            explain_config = ExplainConfig._cast(explain)
-
         query_def = self._parse_query(query)
-        query_str = query_def.to_string()
 
+        if indices is None:
+            if query_def.rfm_entity_ids is None:
+                raise ValueError("Cannot find entities to predict for. Please "
+                                 "pass them via `predict(query, indices=...)`")
+            indices = query_def.get_rfm_entity_id_list()
+        else:
+            query_def = replace(query_def, rfm_entity_ids=None)
+
+        if not isinstance(verbose, ProgressLogger):
+            query_repr = query_def.to_string(rich=True, exclude_predict=True)
+            if explain is not False:
+                msg = f'[bold]EXPLAIN[/bold] {query_repr}'
+            else:
+                msg = f'[bold]PREDICT[/bold] {query_repr}'
+            verbose = ProgressLogger.default(msg=msg, verbose=verbose)
+
+        with verbose as logger:
+            task_table = self._get_task_table(
+                query=query_def,
+                indices=indices,
+                anchor_time=anchor_time,
+                context_anchor_time=context_anchor_time,
+                run_mode=run_mode,
+                max_pq_iterations=max_pq_iterations,
+                random_seed=random_seed,
+                logger=logger,
+            )
+            task_table._query = query_def.to_string()  # type: ignore
+
+            return self.predict_task(
+                task_table,
+                explain=explain,  # type: ignore
+                run_mode=run_mode,
+                num_neighbors=num_neighbors,
+                num_hops=num_hops,
+                verbose=verbose,
+                exclude_cols_dict=query_def.get_exclude_cols_dict(),
+                use_prediction_time=use_prediction_time,
+                top_k=query_def.top_k,
+            )
+
+    @overload
+    def predict_task(
+        self,
+        task: TaskTable,
+        *,
+        explain: Literal[False] = False,
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
+        num_hops: int = 2,
+        verbose: bool | ProgressLogger = True,
+        exclude_cols_dict: dict[str, list[str]] | None = None,
+        use_prediction_time: bool = False,
+        top_k: int | None = None,
+    ) -> pd.DataFrame:
+        pass
+
+    @overload
+    def predict_task(
+        self,
+        task: TaskTable,
+        *,
+        explain: Literal[True] | ExplainConfig | dict[str, Any],
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
+        num_hops: int = 2,
+        verbose: bool | ProgressLogger = True,
+        exclude_cols_dict: dict[str, list[str]] | None = None,
+        use_prediction_time: bool = False,
+        top_k: int | None = None,
+    ) -> Explanation:
+        pass
+
+    def predict_task(
+        self,
+        task: TaskTable,
+        *,
+        explain: bool | ExplainConfig | dict[str, Any] = False,
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
+        num_hops: int = 2,
+        verbose: bool | ProgressLogger = True,
+        exclude_cols_dict: dict[str, list[str]] | None = None,
+        use_prediction_time: bool = False,
+        top_k: int | None = None,
+    ) -> pd.DataFrame | Explanation:
+        """Returns predictions for a custom task specification.
+
+        Args:
+            task: The custom :class:`TaskTable`.
+            explain: Configuration for explainability.
+                If set to ``True``, will additionally explain the prediction.
+                Passing in an :class:`ExplainConfig` instance provides control
+                over which parts of explanation are generated.
+                Explainability is currently only supported for single entity
+                predictions with ``run_mode="FAST"``.
+            run_mode: The :class:`RunMode` for the query.
+            num_neighbors: The number of neighbors to sample for each hop.
+                If specified, the ``num_hops`` option will be ignored.
+            num_hops: The number of hops to sample when generating the context.
+            verbose: Whether to print verbose output.
+            exclude_cols_dict: Any column in any table to exclude from the
+                model input.
+            use_prediction_time: Whether to use the anchor timestamp as an
+                additional feature during prediction. This is typically
+                beneficial for time series forecasting tasks.
+            top_k: The number of predictions to return per entity.
+
+        Returns:
+            The predictions as a :class:`pandas.DataFrame`.
+            If ``explain`` is provided, returns an :class:`Explanation` object
+            containing the prediction, summary, and details.
+        """
         if num_hops != 2 and num_neighbors is not None:
             warnings.warn(f"Received custom 'num_neighbors' option; ignoring "
                           f"custom 'num_hops={num_hops}' option")
+        if num_neighbors is None:
+            key = RunMode.FAST if task.task_type.is_link_pred else run_mode
+            num_neighbors = _DEFAULT_NUM_NEIGHBORS[key][:num_hops]
+
+        explain_config: ExplainConfig | None = None
+        if explain is True:
+            explain_config = ExplainConfig()
+        elif explain is not False:
+            explain_config = ExplainConfig._cast(explain)
 
         if explain_config is not None and run_mode in {
                 RunMode.NORMAL, RunMode.BEST
@@ -327,83 +457,82 @@ class KumoRFM:
                           f"run mode 'FAST' (got '{run_mode}'). Provided run "
                           f"mode has been reset. Please lower the run mode to "
                           f"suppress this warning.")
+            run_mode = RunMode.FAST
 
-        if indices is None:
-            if query_def.rfm_entity_ids is None:
-                raise ValueError("Cannot find entities to predict for. Please "
-                                 "pass them via `predict(query, indices=...)`")
-            indices = query_def.get_rfm_entity_id_list()
-        else:
-            query_def = replace(query_def, rfm_entity_ids=None)
-
-        if len(indices) == 0:
-            raise ValueError("At least one entity is required")
-
-        if explain_config is not None and len(indices) > 1:
-            raise ValueError(
-                f"Cannot explain predictions for more than a single entity "
-                f"(got {len(indices)})")
-
-        query_repr = query_def.to_string(rich=True, exclude_predict=True)
-        if explain_config is not None:
-            msg = f'[bold]EXPLAIN[/bold] {query_repr}'
-        else:
-            msg = f'[bold]PREDICT[/bold] {query_repr}'
+        if explain_config is not None and task.num_prediction_examples > 1:
+            raise ValueError(f"Cannot explain predictions for more than a "
+                             f"single entity "
+                             f"(got {task.num_prediction_examples:,})")
 
         if not isinstance(verbose, ProgressLogger):
-            verbose = InteractiveProgressLogger(msg, verbose=verbose)
-
-        with verbose as logger:
-
-            batch_size: Optional[int] = None
-            if self._batch_size == 'max':
-                task_type = LocalPQueryDriver.get_task_type(
-                    query_def,
-                    edge_types=self._graph_store.edge_types,
-                )
-                batch_size = _MAX_PRED_SIZE[task_type]
+            if task.task_type == TaskType.BINARY_CLASSIFICATION:
+                task_type_repr = 'binary classification'
+            elif task.task_type == TaskType.MULTICLASS_CLASSIFICATION:
+                task_type_repr = 'multi-class classification'
+            elif task.task_type == TaskType.REGRESSION:
+                task_type_repr = 'regression'
+            elif task.task_type == TaskType.TEMPORAL_LINK_PREDICTION:
+                task_type_repr = 'link prediction'
             else:
-                batch_size = self._batch_size
+                task_type_repr = str(task.task_type)
 
-            if batch_size is not None:
-                offsets = range(0, len(indices), batch_size)
-                batches = [indices[step:step + batch_size] for step in offsets]
+            if explain_config is not None:
+                msg = f'Explain {task_type_repr} task'
             else:
-                batches = [indices]
+                msg = f'Predict {task_type_repr} task'
+            verbose = ProgressLogger.default(msg=msg, verbose=verbose)
 
-            if len(batches) > 1:
-                logger.log(f"Splitting {len(indices):,} entities into "
-                           f"{len(batches):,} batches of size {batch_size:,}")
+        with verbose as logger:
+            if task.num_context_examples > _MAX_CONTEXT_SIZE[run_mode]:
+                logger.log(f"Sub-sampled {_MAX_CONTEXT_SIZE[run_mode]:,} "
+                           f"out of {task.num_context_examples:,} in-context "
+                           f"examples")
+                task = task.narrow_context(0, _MAX_CONTEXT_SIZE[run_mode])
+
+            if self._batch_size is None:
+                batch_size = task.num_prediction_examples
+            elif self._batch_size == 'max':
+                batch_size = _MAX_PRED_SIZE[task.task_type]
+            else:
+                batch_size = self._batch_size
 
-            predictions: List[pd.DataFrame] = []
-            summary: Optional[str] = None
-            details: Optional[Explanation] = None
-            for i, batch in enumerate(batches):
-                # TODO Re-use the context for subsequent predictions.
+            if batch_size > _MAX_PRED_SIZE[task.task_type]:
+                raise ValueError(f"Cannot predict for more than "
+                                 f"{_MAX_PRED_SIZE[task.task_type]:,} "
+                                 f"entities at once (got {batch_size:,}). Use "
+                                 f"`KumoRFM.batch_mode` to process entities "
+                                 f"in batches with a sufficient batch size.")
+
+            if task.num_prediction_examples > batch_size:
+                num = math.ceil(task.num_prediction_examples / batch_size)
+                logger.log(f"Splitting {task.num_prediction_examples:,} "
+                           f"entities into {num:,} batches of size "
+                           f"{batch_size:,}")
+
+            predictions: list[pd.DataFrame] = []
+            summary: str | None = None
+            details: Explanation | None = None
+            for start in range(0, task.num_prediction_examples, batch_size):
                 context = self._get_context(
-                    query=query_def,
-                    indices=batch,
-                    anchor_time=anchor_time,
-                    context_anchor_time=context_anchor_time,
-                    run_mode=RunMode(run_mode),
+                    task=task.narrow_prediction(start, length=batch_size),
+                    run_mode=run_mode,
                     num_neighbors=num_neighbors,
-                    num_hops=num_hops,
-                    max_pq_iterations=max_pq_iterations,
-                    evaluate=False,
-                    random_seed=random_seed,
-                    logger=logger if i == 0 else None,
+                    exclude_cols_dict=exclude_cols_dict,
+                    top_k=top_k,
                 )
+                context.y_test = None
+
                 request = RFMPredictRequest(
                     context=context,
                     run_mode=RunMode(run_mode),
-                    query=query_str,
+                    query=getattr(task, '_query', ''),
                     use_prediction_time=use_prediction_time,
                 )
                 with warnings.catch_warnings():
                     warnings.filterwarnings('ignore', message='gencode')
                     request_msg = request.to_protobuf()
                     _bytes = request_msg.SerializeToString()
-                if i == 0:
+                if start == 0:
                     logger.log(f"Generated context of size "
                                f"{len(_bytes) / (1024*1024):.2f}MB")
 
@@ -411,12 +540,9 @@ class KumoRFM:
                     stats = Context.get_memory_stats(request_msg.context)
                     raise ValueError(_SIZE_LIMIT_MSG.format(stats=stats))
 
-                if (isinstance(verbose, InteractiveProgressLogger) and i == 0
-                        and len(batches) > 1):
-                    verbose.init_progress(
-                        total=len(batches),
-                        description='Predicting',
-                    )
+                if start == 0 and task.num_prediction_examples > batch_size:
+                    num = math.ceil(task.num_prediction_examples / batch_size)
+                    verbose.init_progress(total=num, description='Predicting')
 
                 for attempt in range(self.num_retries + 1):
                     try:
@@ -433,10 +559,10 @@ class KumoRFM:
 
                         # Cast 'ENTITY' to correct data type:
                         if 'ENTITY' in df:
-                            entity = query_def.entity_table
-                            pkey_map = self._graph_store.pkey_map_dict[entity]
-                            df['ENTITY'] = df['ENTITY'].astype(
-                                type(pkey_map.index[0]))
+                            table_dict = context.subgraph.table_dict
+                            table = table_dict[context.entity_table_names[0]]
+                            ser = table.df[table.primary_key]
+                            df['ENTITY'] = df['ENTITY'].astype(ser.dtype)
 
                         # Cast 'ANCHOR_TIMESTAMP' to correct data type:
                         if 'ANCHOR_TIMESTAMP' in df:
@@ -451,8 +577,7 @@ class KumoRFM:
 
                         predictions.append(df)
 
-                        if (isinstance(verbose, InteractiveProgressLogger)
-                                and len(batches) > 1):
+                        if task.num_prediction_examples > batch_size:
                             verbose.step()
 
                         break
@@ -490,9 +615,9 @@ class KumoRFM:
     def is_valid_entity(
         self,
         query: str,
-        indices: Union[List[str], List[float], List[int], None] = None,
+        indices: list[str] | list[float] | list[int] | None = None,
         *,
-        anchor_time: Union[pd.Timestamp, Literal['entity'], None] = None,
+        anchor_time: pd.Timestamp | Literal['entity'] | None = None,
     ) -> np.ndarray:
         r"""Returns a mask that denotes which entities are valid for the
         given predictive query, *i.e.*, which entities fulfill (temporal)
@@ -519,37 +644,32 @@ class KumoRFM:
             raise ValueError("At least one entity is required")
 
         if anchor_time is None:
-            anchor_time = self._graph_store.max_time
+            anchor_time = self._get_default_anchor_time(query_def)
 
         if isinstance(anchor_time, pd.Timestamp):
             self._validate_time(query_def, anchor_time, None, False)
         else:
             assert anchor_time == 'entity'
-            if (query_def.entity_table not in self._graph_store.time_dict):
+            if query_def.entity_table not in self._sampler.time_column_dict:
                 raise ValueError(f"Anchor time 'entity' requires the entity "
                                  f"table '{query_def.entity_table}' "
                                  f"to have a time column.")
 
-        node = self._graph_store.get_node_id(
-            table_name=query_def.entity_table,
-            pkey=pd.Series(indices),
-        )
-        query_driver = LocalPQueryDriver(self._graph_store, query_def)
-        return query_driver.is_valid(node, anchor_time)
+        raise NotImplementedError
 
     def evaluate(
         self,
         query: str,
         *,
-        metrics: Optional[List[str]] = None,
-        anchor_time: Union[pd.Timestamp, Literal['entity'], None] = None,
-        context_anchor_time: Union[pd.Timestamp, None] = None,
-        run_mode: Union[RunMode, str] = RunMode.FAST,
-        num_neighbors: Optional[List[int]] = None,
+        metrics: list[str] | None = None,
+        anchor_time: pd.Timestamp | Literal['entity'] | None = None,
+        context_anchor_time: pd.Timestamp | None = None,
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
         num_hops: int = 2,
-        max_pq_iterations: int = 20,
-        random_seed: Optional[int] = _RANDOM_SEED,
-        verbose: Union[bool, ProgressLogger] = True,
+        max_pq_iterations: int = 10,
+        random_seed: int | None = _RANDOM_SEED,
+        verbose: bool | ProgressLogger = True,
         use_prediction_time: bool = False,
     ) -> pd.DataFrame:
         """Evaluates a predictive query.
@@ -582,40 +702,51 @@ class KumoRFM:
             The metrics as a :class:`pandas.DataFrame`
         """
         query_def = self._parse_query(query)
-
-        if num_hops != 2 and num_neighbors is not None:
-            warnings.warn(f"Received custom 'num_neighbors' option; ignoring "
-                          f"custom 'num_hops={num_hops}' option")
-
         if query_def.rfm_entity_ids is not None:
             query_def = replace(
                 query_def,
                 rfm_entity_ids=None,
             )
 
-        query_repr = query_def.to_string(rich=True, exclude_predict=True)
-        msg = f'[bold]EVALUATE[/bold] {query_repr}'
+        task_type = self._get_task_type(
+            query=query_def,
+            edge_types=self._sampler.edge_types,
+        )
+
+        if num_hops != 2 and num_neighbors is not None:
+            warnings.warn(f"Received custom 'num_neighbors' option; ignoring "
+                          f"custom 'num_hops={num_hops}' option")
+        if num_neighbors is None:
+            key = RunMode.FAST if task_type.is_link_pred else run_mode
+            num_neighbors = _DEFAULT_NUM_NEIGHBORS[key][:num_hops]
+
+        if metrics is not None and len(metrics) > 0:
+            self._validate_metrics(metrics, task_type)
+            metrics = list(dict.fromkeys(metrics))
 
         if not isinstance(verbose, ProgressLogger):
-            verbose = InteractiveProgressLogger(msg, verbose=verbose)
+            query_repr = query_def.to_string(rich=True, exclude_predict=True)
+            msg = f'[bold]EVALUATE[/bold] {query_repr}'
+            verbose = ProgressLogger.default(msg=msg, verbose=verbose)
 
         with verbose as logger:
-            context = self._get_context(
+            task_table = self._get_task_table(
                 query=query_def,
                 indices=None,
                 anchor_time=anchor_time,
                 context_anchor_time=context_anchor_time,
-                run_mode=RunMode(run_mode),
-                num_neighbors=num_neighbors,
-                num_hops=num_hops,
+                run_mode=run_mode,
                 max_pq_iterations=max_pq_iterations,
-                evaluate=True,
                 random_seed=random_seed,
-                logger=logger if verbose else None,
+                logger=logger,
+            )
+            context = self._get_context(
+                task=task_table,
+                run_mode=run_mode,
+                num_neighbors=num_neighbors,
+                exclude_cols_dict=query_def.get_exclude_cols_dict(),
             )
-            if metrics is not None and len(metrics) > 0:
-                self._validate_metrics(metrics, context.task_type)
-                metrics = list(dict.fromkeys(metrics))
+
             request = RFMEvaluateRequest(
                 context=context,
                 run_mode=RunMode(run_mode),
@@ -633,17 +764,23 @@ class KumoRFM:
                 stats_msg = Context.get_memory_stats(request_msg.context)
                 raise ValueError(_SIZE_LIMIT_MSG.format(stats=stats_msg))
 
-            try:
-                resp = self._api_client.evaluate(request_bytes)
-            except HTTPException as e:
+            for attempt in range(self.num_retries + 1):
                 try:
-                    msg = json.loads(e.detail)['detail']
-                except Exception:
-                    msg = e.detail
-                raise RuntimeError(f"An unexpected exception occurred. "
-                                   f"Please create an issue at "
-                                   f"'https://github.com/kumo-ai/kumo-rfm'. "
-                                   f"{msg}") from None
+                    resp = self._api_client.evaluate(request_bytes)
+                    break
+                except HTTPException as e:
+                    if attempt == self.num_retries:
+                        try:
+                            msg = json.loads(e.detail)['detail']
+                        except Exception:
+                            msg = e.detail
+                        raise RuntimeError(
+                            f"An unexpected exception occurred. Please create "
+                            f"an issue at "
+                            f"'https://github.com/kumo-ai/kumo-rfm'. {msg}"
+                        ) from None
+
+                    time.sleep(2**attempt)  # 1s, 2s, 4s, 8s, ...
 
         return pd.DataFrame.from_dict(
             resp.metrics,
@@ -656,9 +793,9 @@ class KumoRFM:
         query: str,
         size: int,
         *,
-        anchor_time: Union[pd.Timestamp, Literal['entity'], None] = None,
-        random_seed: Optional[int] = _RANDOM_SEED,
-        max_iterations: int = 20,
+        anchor_time: pd.Timestamp | Literal['entity'] | None = None,
+        random_seed: int | None = _RANDOM_SEED,
+        max_iterations: int = 10,
     ) -> pd.DataFrame:
         """Returns the labels of a predictive query for a specified anchor
         time.
@@ -678,40 +815,37 @@ class KumoRFM:
         query_def = self._parse_query(query)
 
         if anchor_time is None:
-            anchor_time = self._graph_store.max_time
+            anchor_time = self._get_default_anchor_time(query_def)
             if query_def.target_ast.date_offset_range is not None:
-                anchor_time = anchor_time - (
-                    query_def.target_ast.date_offset_range.end_date_offset *
-                    query_def.num_forecasts)
+                offset = query_def.target_ast.date_offset_range.end_date_offset
+                offset *= query_def.num_forecasts
+                anchor_time -= offset
 
         assert anchor_time is not None
         if isinstance(anchor_time, pd.Timestamp):
             self._validate_time(query_def, anchor_time, None, evaluate=True)
         else:
             assert anchor_time == 'entity'
-            if (query_def.entity_table not in self._graph_store.time_dict):
+            if query_def.entity_table not in self._sampler.time_column_dict:
                 raise ValueError(f"Anchor time 'entity' requires the entity "
                                  f"table '{query_def.entity_table}' "
                                  f"to have a time column")
 
-        query_driver = LocalPQueryDriver(self._graph_store, query_def,
-                                         random_seed)
-
-        node, time, y = query_driver.collect_test(
-            size=size,
-            anchor_time=anchor_time,
-            batch_size=min(10_000, size),
-            max_iterations=max_iterations,
-            guarantee_train_examples=False,
+        train, test = self._sampler.sample_target(
+            query=query_def,
+            num_train_examples=0,
+            train_anchor_time=anchor_time,
+            num_train_trials=0,
+            num_test_examples=size,
+            test_anchor_time=anchor_time,
+            num_test_trials=max_iterations * size,
+            random_seed=random_seed,
         )
 
-        entity = self._graph_store.pkey_map_dict[
-            query_def.entity_table].index[node]
-
         return pd.DataFrame({
-            'ENTITY': entity,
-            'ANCHOR_TIMESTAMP': time,
-            'TARGET': y,
+            'ENTITY': test.entity_pkey,
+            'ANCHOR_TIMESTAMP': test.anchor_time,
+            'TARGET': test.target,
         })
 
     # Helpers #################################################################
@@ -726,63 +860,120 @@ class KumoRFM:
                              "`predict()` or `evaluate()` methods to perform "
                              "predictions or evaluations.")
 
-        try:
-            request = RFMParseQueryRequest(
-                query=query,
-                graph_definition=self._graph_def,
-            )
+        request = RFMParseQueryRequest(
+            query=query,
+            graph_definition=self._graph_def,
+        )
 
-            resp = self._api_client.parse_query(request)
+        for attempt in range(self.num_retries + 1):
+            try:
+                resp = self._api_client.parse_query(request)
+                break
+            except HTTPException as e:
+                if attempt == self.num_retries:
+                    try:
+                        msg = json.loads(e.detail)['detail']
+                    except Exception:
+                        msg = e.detail
+                    raise ValueError(f"Failed to parse query '{query}'. {msg}")
 
-            # TODO Expose validation warnings.
+                time.sleep(2**attempt)  # 1s, 2s, 4s, 8s, ...
 
-            if len(resp.validation_response.warnings) > 0:
-                msg = '\n'.join([
-                    f'{i+1}. {warning.title}: {warning.message}' for i, warning
-                    in enumerate(resp.validation_response.warnings)
-                ])
-                warnings.warn(f"Encountered the following warnings during "
-                              f"parsing:\n{msg}")
+        if len(resp.validation_response.warnings) > 0:
+            msg = '\n'.join([
+                f'{i+1}. {warning.title}: {warning.message}'
+                for i, warning in enumerate(resp.validation_response.warnings)
+            ])
+            warnings.warn(f"Encountered the following warnings during "
+                          f"parsing:\n{msg}")
 
-            return resp.query
-        except HTTPException as e:
-            try:
-                msg = json.loads(e.detail)['detail']
-            except Exception:
-                msg = e.detail
-            raise ValueError(f"Failed to parse query '{query}'. "
-                             f"{msg}") from None
+        return resp.query
+
+    @staticmethod
+    def _get_task_type(
+        query: ValidatedPredictiveQuery,
+        edge_types: list[tuple[str, str, str]],
+    ) -> TaskType:
+        if isinstance(query.target_ast, (Condition, LogicalOperation)):
+            return TaskType.BINARY_CLASSIFICATION
+
+        target = query.target_ast
+        if isinstance(target, Join):
+            target = target.rhs_target
+        if isinstance(target, Aggregation):
+            if target.aggr == AggregationType.LIST_DISTINCT:
+                table_name, col_name = target._get_target_column_name().split(
+                    '.')
+                target_edge_types = [
+                    edge_type for edge_type in edge_types
+                    if edge_type[0] == table_name and edge_type[1] == col_name
+                ]
+                if len(target_edge_types) != 1:
+                    raise NotImplementedError(
+                        f"Multilabel-classification queries based on "
+                        f"'LIST_DISTINCT' are not supported yet. If you "
+                        f"planned to write a link prediction query instead, "
+                        f"make sure to register '{col_name}' as a "
+                        f"foreign key.")
+                return TaskType.TEMPORAL_LINK_PREDICTION
+
+            return TaskType.REGRESSION
+
+        assert isinstance(target, Column)
+
+        if target.stype in {Stype.ID, Stype.categorical}:
+            return TaskType.MULTICLASS_CLASSIFICATION
+
+        if target.stype in {Stype.numerical}:
+            return TaskType.REGRESSION
+
+        raise NotImplementedError("Task type not yet supported")
+
+    def _get_default_anchor_time(
+        self,
+        query: ValidatedPredictiveQuery | None = None,
+    ) -> pd.Timestamp:
+        if query is not None and query.query_type == QueryType.TEMPORAL:
+            aggr_table_names = [
+                aggr._get_target_column_name().split('.')[0]
+                for aggr in query.get_all_target_aggregations()
+            ]
+            return self._sampler.get_max_time(aggr_table_names)
+
+        return self._sampler.get_max_time()
 
     def _validate_time(
         self,
         query: ValidatedPredictiveQuery,
         anchor_time: pd.Timestamp,
-        context_anchor_time: Union[pd.Timestamp, None],
+        context_anchor_time: pd.Timestamp | None,
         evaluate: bool,
     ) -> None:
 
-        if self._graph_store.min_time == pd.Timestamp.max:
+        if len(self._sampler.time_column_dict) == 0:
             return  # Graph without timestamps
 
-        if anchor_time < self._graph_store.min_time:
+        min_time = self._sampler.get_min_time()
+        max_time = self._sampler.get_max_time()
+
+        if anchor_time < min_time:
             raise ValueError(f"Anchor timestamp '{anchor_time}' is before "
-                             f"the earliest timestamp "
-                             f"'{self._graph_store.min_time}' in the data.")
+                             f"the earliest timestamp '{min_time}' in the "
+                             f"data.")
 
-        if (context_anchor_time is not None
-                and context_anchor_time < self._graph_store.min_time):
+        if context_anchor_time is not None and context_anchor_time < min_time:
             raise ValueError(f"Context anchor timestamp is too early or "
                              f"aggregation time range is too large. To make "
                              f"this prediction, we would need data back to "
                              f"'{context_anchor_time}', however, your data "
-                             f"only contains data back to "
-                             f"'{self._graph_store.min_time}'.")
+                             f"only contains data back to '{min_time}'.")
 
         if query.target_ast.date_offset_range is not None:
             end_offset = query.target_ast.date_offset_range.end_date_offset
         else:
             end_offset = pd.DateOffset(0)
-        forecast_end_offset = end_offset * query.num_forecasts
+        end_offset = end_offset * query.num_forecasts
+
         if (context_anchor_time is not None
                 and context_anchor_time > anchor_time):
             warnings.warn(f"Context anchor timestamp "
@@ -792,7 +983,7 @@ class KumoRFM:
                           f"intended.")
         elif (query.query_type == QueryType.TEMPORAL
               and context_anchor_time is not None
-              and context_anchor_time + forecast_end_offset > anchor_time):
+              and context_anchor_time + end_offset > anchor_time):
             warnings.warn(f"Aggregation for context examples at timestamp "
                           f"'{context_anchor_time}' will leak information "
                           f"from the prediction anchor timestamp "
@@ -800,62 +991,44 @@ class KumoRFM:
                           f"intended.")
 
         elif (context_anchor_time is not None
-              and context_anchor_time - forecast_end_offset
-              < self._graph_store.min_time):
-            _time = context_anchor_time - forecast_end_offset
+              and context_anchor_time - end_offset < min_time):
+            _time = context_anchor_time - end_offset
             warnings.warn(f"Context anchor timestamp is too early or "
                           f"aggregation time range is too large. To form "
                           f"proper input data, we would need data back to "
                           f"'{_time}', however, your data only contains "
-                          f"data back to '{self._graph_store.min_time}'.")
+                          f"data back to '{min_time}'.")
 
-        if (not evaluate and anchor_time
-                > self._graph_store.max_time + pd.DateOffset(days=1)):
+        if not evaluate and anchor_time > max_time + pd.DateOffset(days=1):
             warnings.warn(f"Anchor timestamp '{anchor_time}' is after the "
-                          f"latest timestamp '{self._graph_store.max_time}' "
-                          f"in the data. Please make sure this is intended.")
+                          f"latest timestamp '{max_time}' in the data. Please "
+                          f"make sure this is intended.")
 
-        max_eval_time = self._graph_store.max_time - forecast_end_offset
-        if evaluate and anchor_time > max_eval_time:
+        if evaluate and anchor_time > max_time - end_offset:
             raise ValueError(
                 f"Anchor timestamp for evaluation is after the latest "
-                f"supported timestamp '{max_eval_time}'.")
+                f"supported timestamp '{max_time - end_offset}'.")
 
-    def _get_context(
+    def _get_task_table(
         self,
         query: ValidatedPredictiveQuery,
-        indices: Union[List[str], List[float], List[int], None],
-        anchor_time: Union[pd.Timestamp, Literal['entity'], None],
-        context_anchor_time: Union[pd.Timestamp, None],
-        run_mode: RunMode,
-        num_neighbors: Optional[List[int]],
-        num_hops: int,
-        max_pq_iterations: int,
-        evaluate: bool,
-        random_seed: Optional[int] = _RANDOM_SEED,
-        logger: Optional[ProgressLogger] = None,
-    ) -> Context:
-
-        if num_neighbors is not None:
-            num_hops = len(num_neighbors)
-
-        if num_hops < 0:
-            raise ValueError(f"'num_hops' must be non-negative "
-                             f"(got {num_hops})")
-        if num_hops > 6:
-            raise ValueError(f"Cannot predict on subgraphs with more than 6 "
-                             f"hops (got {num_hops}). Please reduce the "
-                             f"number of hops and try again. Please create a "
-                             f"feature request at "
-                             f"'https://github.com/kumo-ai/kumo-rfm' if you "
-                             f"must go beyond this for your use-case.")
-
-        query_driver = LocalPQueryDriver(self._graph_store, query, random_seed)
-        task_type = LocalPQueryDriver.get_task_type(
-            query,
-            edge_types=self._graph_store.edge_types,
+        indices: list[str] | list[float] | list[int] | None,
+        anchor_time: pd.Timestamp | Literal['entity'] | None = None,
+        context_anchor_time: pd.Timestamp | None = None,
+        run_mode: RunMode = RunMode.FAST,
+        max_pq_iterations: int = 10,
+        random_seed: int | None = _RANDOM_SEED,
+        logger: ProgressLogger | None = None,
+    ) -> TaskTable:
+
+        task_type = self._get_task_type(
+            query=query,
+            edge_types=self._sampler.edge_types,
         )
 
+        num_train_examples = _MAX_CONTEXT_SIZE[run_mode]
+        num_test_examples = _MAX_TEST_SIZE[task_type] if indices is None else 0
+
         if logger is not None:
             if task_type == TaskType.BINARY_CLASSIFICATION:
                 task_type_repr = 'binary classification'
@@ -869,30 +1042,17 @@ class KumoRFM:
                 task_type_repr = str(task_type)
             logger.log(f"Identified {query.query_type} {task_type_repr} task")
 
-        if task_type.is_link_pred and num_hops < 2:
-            raise ValueError(f"Cannot perform link prediction on subgraphs "
-                             f"with less than 2 hops (got {num_hops}) since "
-                             f"historical target entities need to be part of "
-                             f"the context. Please increase the number of "
-                             f"hops and try again.")
-
-        if num_neighbors is None:
-            if run_mode == RunMode.DEBUG:
-                num_neighbors = [16, 16, 4, 4, 1, 1][:num_hops]
-            elif run_mode == RunMode.FAST or task_type.is_link_pred:
-                num_neighbors = [32, 32, 8, 8, 4, 4][:num_hops]
-            else:
-                num_neighbors = [64, 64, 8, 8, 4, 4][:num_hops]
-
         if query.target_ast.date_offset_range is None:
-            end_offset = pd.DateOffset(0)
+            step_offset = pd.DateOffset(0)
         else:
-            end_offset = query.target_ast.date_offset_range.end_date_offset
-        forecast_end_offset = end_offset * query.num_forecasts
+            step_offset = query.target_ast.date_offset_range.end_date_offset
+        end_offset = step_offset * query.num_forecasts
+
         if anchor_time is None:
-            anchor_time = self._graph_store.max_time
-            if evaluate:
-                anchor_time = anchor_time - forecast_end_offset
+            anchor_time = self._get_default_anchor_time(query)
+            if num_test_examples > 0:
+                anchor_time = anchor_time - end_offset
+
             if logger is not None:
                 assert isinstance(anchor_time, pd.Timestamp)
                 if anchor_time == pd.Timestamp.min:
@@ -904,114 +1064,98 @@ class KumoRFM:
                 else:
                     logger.log(f"Derived anchor time {anchor_time}")
 
-        assert anchor_time is not None
         if isinstance(anchor_time, pd.Timestamp):
+            if context_anchor_time == 'entity':
+                raise ValueError("Anchor time 'entity' needs to be shared "
+                                 "for context and prediction examples")
             if context_anchor_time is None:
-                context_anchor_time = anchor_time - forecast_end_offset
+                context_anchor_time = anchor_time - end_offset
             self._validate_time(query, anchor_time, context_anchor_time,
-                                evaluate)
+                                evaluate=num_test_examples > 0)
         else:
             assert anchor_time == 'entity'
-            if query.entity_table not in self._graph_store.time_dict:
+            if query.query_type != QueryType.STATIC:
+                raise ValueError("Anchor time 'entity' is only valid for "
+                                 "static predictive queries")
+            if query.entity_table not in self._sampler.time_column_dict:
                 raise ValueError(f"Anchor time 'entity' requires the entity "
                                  f"table '{query.entity_table}' to "
                                  f"have a time column")
-            if context_anchor_time is not None:
-                warnings.warn("Ignoring option 'context_anchor_time' for "
-                              "`anchor_time='entity'`")
-            context_anchor_time = None
-
-        y_test: Optional[pd.Series] = None
-        if evaluate:
-            max_test_size = _MAX_TEST_SIZE[run_mode]
-            if task_type.is_link_pred:
-                max_test_size = max_test_size // 5
-
-            test_node, test_time, y_test = query_driver.collect_test(
-                size=max_test_size,
-                anchor_time=anchor_time,
-                max_iterations=max_pq_iterations,
-                guarantee_train_examples=True,
-            )
-            if logger is not None:
-                if task_type == TaskType.BINARY_CLASSIFICATION:
-                    pos = 100 * int((y_test > 0).sum()) / len(y_test)
-                    msg = (f"Collected {len(y_test):,} test examples with "
-                           f"{pos:.2f}% positive cases")
-                elif task_type == TaskType.MULTICLASS_CLASSIFICATION:
-                    msg = (f"Collected {len(y_test):,} test examples "
-                           f"holding {y_test.nunique()} classes")
-                elif task_type == TaskType.REGRESSION:
-                    _min, _max = float(y_test.min()), float(y_test.max())
-                    msg = (f"Collected {len(y_test):,} test examples with "
-                           f"targets between {format_value(_min)} and "
-                           f"{format_value(_max)}")
-                elif task_type == TaskType.TEMPORAL_LINK_PREDICTION:
-                    num_rhs = y_test.explode().nunique()
-                    msg = (f"Collected {len(y_test):,} test examples with "
-                           f"{num_rhs:,} unique items")
-                else:
-                    raise NotImplementedError
-                logger.log(msg)
-
-        else:
-            assert indices is not None
-
-            if len(indices) > _MAX_PRED_SIZE[task_type]:
-                raise ValueError(f"Cannot predict for more than "
-                                 f"{_MAX_PRED_SIZE[task_type]:,} entities at "
-                                 f"once (got {len(indices):,}). Use "
-                                 f"`KumoRFM.batch_mode` to process entities "
-                                 f"in batches")
+            if isinstance(context_anchor_time, pd.Timestamp):
+                raise ValueError("Anchor time 'entity' needs to be shared "
+                                 "for context and prediction examples")
+            context_anchor_time = 'entity'
+
+        train, test = self._sampler.sample_target(
+            query=query,
+            num_train_examples=num_train_examples,
+            train_anchor_time=context_anchor_time,
+            num_train_trials=max_pq_iterations * num_train_examples,
+            num_test_examples=num_test_examples,
+            test_anchor_time=anchor_time,
+            num_test_trials=max_pq_iterations * num_test_examples,
+            random_seed=random_seed,
+        )
+        train_pkey, train_time, train_y = train
+        test_pkey, test_time, test_y = test
 
-            test_node = self._graph_store.get_node_id(
-                table_name=query.entity_table,
-                pkey=pd.Series(indices),
-            )
+        if num_test_examples > 0 and logger is not None:
+            if task_type == TaskType.BINARY_CLASSIFICATION:
+                pos = 100 * int((test_y > 0).sum()) / len(test_y)
+                msg = (f"Collected {len(test_y):,} test examples with "
+                       f"{pos:.2f}% positive cases")
+            elif task_type == TaskType.MULTICLASS_CLASSIFICATION:
+                msg = (f"Collected {len(test_y):,} test examples holding "
+                       f"{test_y.nunique()} classes")
+            elif task_type == TaskType.REGRESSION:
+                _min, _max = float(test_y.min()), float(test_y.max())
+                msg = (f"Collected {len(test_y):,} test examples with targets "
+                       f"between {format_value(_min)} and "
+                       f"{format_value(_max)}")
+            elif task_type == TaskType.TEMPORAL_LINK_PREDICTION:
+                num_rhs = test_y.explode().nunique()
+                msg = (f"Collected {len(test_y):,} test examples with "
+                       f"{num_rhs:,} unique items")
+            else:
+                raise NotImplementedError
+            logger.log(msg)
 
+        if num_test_examples == 0:
+            assert indices is not None
+            test_pkey = pd.Series(indices, dtype=train_pkey.dtype)
             if isinstance(anchor_time, pd.Timestamp):
-                test_time = pd.Series(anchor_time).repeat(
-                    len(test_node)).reset_index(drop=True)
+                test_time = pd.Series([anchor_time]).repeat(
+                    len(indices)).reset_index(drop=True)
             else:
-                time = self._graph_store.time_dict[query.entity_table]
-                time = time[test_node] * 1000**3
-                test_time = pd.Series(time, dtype='datetime64[ns]')
-
-        train_node, train_time, y_train = query_driver.collect_train(
-            size=_MAX_CONTEXT_SIZE[run_mode],
-            anchor_time=context_anchor_time or 'entity',
-            exclude_node=test_node if (query.query_type == QueryType.STATIC
-                                       or anchor_time == 'entity') else None,
-            max_iterations=max_pq_iterations,
-        )
+                train_time = test_time = 'entity'
 
         if logger is not None:
             if task_type == TaskType.BINARY_CLASSIFICATION:
-                pos = 100 * int((y_train > 0).sum()) / len(y_train)
-                msg = (f"Collected {len(y_train):,} in-context examples with "
+                pos = 100 * int((train_y > 0).sum()) / len(train_y)
+                msg = (f"Collected {len(train_y):,} in-context examples with "
                        f"{pos:.2f}% positive cases")
             elif task_type == TaskType.MULTICLASS_CLASSIFICATION:
-                msg = (f"Collected {len(y_train):,} in-context examples "
-                       f"holding {y_train.nunique()} classes")
+                msg = (f"Collected {len(train_y):,} in-context examples "
+                       f"holding {train_y.nunique()} classes")
             elif task_type == TaskType.REGRESSION:
-                _min, _max = float(y_train.min()), float(y_train.max())
-                msg = (f"Collected {len(y_train):,} in-context examples with "
+                _min, _max = float(train_y.min()), float(train_y.max())
+                msg = (f"Collected {len(train_y):,} in-context examples with "
                        f"targets between {format_value(_min)} and "
                        f"{format_value(_max)}")
             elif task_type == TaskType.TEMPORAL_LINK_PREDICTION:
-                num_rhs = y_train.explode().nunique()
-                msg = (f"Collected {len(y_train):,} in-context examples with "
+                num_rhs = train_y.explode().nunique()
+                msg = (f"Collected {len(train_y):,} in-context examples with "
                        f"{num_rhs:,} unique items")
             else:
                 raise NotImplementedError
             logger.log(msg)
 
-        entity_table_names: Tuple[str, ...]
+        entity_table_names: tuple[str] | tuple[str, str]
         if task_type.is_link_pred:
             final_aggr = query.get_final_target_aggregation()
             assert final_aggr is not None
             edge_fkey = final_aggr._get_target_column_name()
-            for edge_type in self._graph_store.edge_types:
+            for edge_type in self._sampler.edge_types:
                 if edge_fkey == f'{edge_type[0]}.{edge_type[1]}':
                     entity_table_names = (
                         query.entity_table,
@@ -1020,23 +1164,98 @@ class KumoRFM:
         else:
             entity_table_names = (query.entity_table, )
 
+        context_df = pd.DataFrame({'ENTITY': train_pkey, 'TARGET': train_y})
+        if isinstance(train_time, pd.Series):
+            context_df['ANCHOR_TIMESTAMP'] = train_time
+        pred_df = pd.DataFrame({'ENTITY': test_pkey})
+        if num_test_examples > 0:
+            pred_df['TARGET'] = test_y
+        if isinstance(test_time, pd.Series):
+            pred_df['ANCHOR_TIMESTAMP'] = test_time
+
+        return TaskTable(
+            task_type=task_type,
+            context_df=context_df,
+            pred_df=pred_df,
+            entity_table_name=entity_table_names,
+            entity_column='ENTITY',
+            target_column='TARGET',
+            time_column='ANCHOR_TIMESTAMP' if isinstance(
+                train_time, pd.Series) else TaskTable.ENTITY_TIME,
+        )
+
+    def _get_context(
+        self,
+        task: TaskTable,
+        run_mode: RunMode | str = RunMode.FAST,
+        num_neighbors: list[int] | None = None,
+        exclude_cols_dict: dict[str, list[str]] | None = None,
+        top_k: int | None = None,
+    ) -> Context:
+
+        # TODO Remove all
+        if task.num_context_examples > max(_MAX_CONTEXT_SIZE.values()):
+            raise ValueError(f"Cannot process a context with more than "
+                             f"{max(_MAX_CONTEXT_SIZE.values()):,} samples "
+                             f"(got {task.num_context_examples:,})")
+        if task.evaluate:
+            if task.num_prediction_examples > _MAX_TEST_SIZE[task.task_type]:
+                raise ValueError(f"Cannot process a test set with more than "
+                                 f"{_MAX_TEST_SIZE[task.task_type]:,} samples "
+                                 f"for evaluation "
+                                 f"(got {task.num_prediction_examples:,})")
+        else:
+            if task.num_prediction_examples > _MAX_PRED_SIZE[task.task_type]:
+                raise ValueError(f"Cannot predict for more than "
+                                 f"{_MAX_PRED_SIZE[task.task_type]:,} "
+                                 f"entities at once "
+                                 f"(got {task.num_prediction_examples:,})")
+
+        if num_neighbors is None:
+            key = RunMode.FAST if task.task_type.is_link_pred else run_mode
+            num_neighbors = _DEFAULT_NUM_NEIGHBORS[key][:2]
+
+        if len(num_neighbors) > 6:
+            raise ValueError(f"Cannot predict on subgraphs with more than 6 "
+                             f"hops (got {len(num_neighbors)}). Reduce the "
+                             f"number of hops and try again. Please create a "
+                             f"feature request at "
+                             f"'https://github.com/kumo-ai/kumo-rfm' if you "
+                             f"must go beyond this for your use-case.")
+
         # Exclude the entity anchor time from the feature set to prevent
         # running out-of-distribution between in-context and test examples:
-        exclude_cols_dict = query.get_exclude_cols_dict()
-        if anchor_time == 'entity':
-            if entity_table_names[0] not in exclude_cols_dict:
-                exclude_cols_dict[entity_table_names[0]] = []
-            time_column_dict = self._graph_store.time_column_dict
-            time_column = time_column_dict[entity_table_names[0]]
-            exclude_cols_dict[entity_table_names[0]].append(time_column)
-
-        subgraph = self._graph_sampler(
-            entity_table_names=entity_table_names,
-            node=np.concatenate([train_node, test_node]),
-            time=np.concatenate([
-                train_time.astype('datetime64[ns]').astype(int).to_numpy(),
-                test_time.astype('datetime64[ns]').astype(int).to_numpy(),
-            ]),
+        exclude_cols_dict = exclude_cols_dict or {}
+        if task.entity_table_name in self._sampler.time_column_dict:
+            if task.entity_table_name not in exclude_cols_dict:
+                exclude_cols_dict[task.entity_table_name] = []
+            time_col = self._sampler.time_column_dict[task.entity_table_name]
+            exclude_cols_dict[task.entity_table_name].append(time_col)
+
+        entity_pkey = pd.concat([
+            task._context_df[task._entity_column],
+            task._pred_df[task._entity_column],
+        ], axis=0, ignore_index=True)
+
+        if task.use_entity_time:
+            if task.entity_table_name not in self._sampler.time_column_dict:
+                raise ValueError(f"The given annchor time requires the entity "
+                                 f"table '{task.entity_table_name}' to have a "
+                                 f"time column")
+            anchor_time = 'entity'
+        elif task._time_column is not None:
+            anchor_time = pd.concat([
+                task._context_df[task._time_column],
+                task._pred_df[task._time_column],
+            ], axis=0, ignore_index=True)
+        else:
+            anchor_time = pd.Series(self._get_default_anchor_time()).repeat(
+                (len(entity_pkey))).reset_index(drop=True)
+
+        subgraph = self._sampler.sample_subgraph(
+            entity_table_names=task.entity_table_names,
+            entity_pkey=entity_pkey,
+            anchor_time=anchor_time,
             num_neighbors=num_neighbors,
             exclude_cols_dict=exclude_cols_dict,
         )
@@ -1048,23 +1267,26 @@ class KumoRFM:
                              f"'https://github.com/kumo-ai/kumo-rfm' if you "
                              f"must go beyond this for your use-case.")
 
-        step_size: Optional[int] = None
-        if query.query_type == QueryType.TEMPORAL:
-            step_size = date_offset_to_seconds(end_offset)
+        if (task.task_type.is_link_pred
+                and task.entity_table_names[-1] not in subgraph.table_dict):
+            raise ValueError("Cannot perform link prediction on subgraphs "
+                             "without any historical target entities. Please "
+                             "increase the number of hops and try again.")
 
         return Context(
-            task_type=task_type,
-            entity_table_names=entity_table_names,
+            task_type=task.task_type,
+            entity_table_names=task.entity_table_names,
             subgraph=subgraph,
-            y_train=y_train,
-            y_test=y_test,
-            top_k=query.top_k,
-            step_size=step_size,
+            y_train=task._context_df[task.target_column.name],
+            y_test=task._pred_df[task.target_column.name]
+            if task.evaluate else None,
+            top_k=top_k,
+            step_size=None,
         )
 
     @staticmethod
     def _validate_metrics(
-        metrics: List[str],
+        metrics: list[str],
         task_type: TaskType,
     ) -> None:
 
@@ -1121,7 +1343,7 @@ class KumoRFM:
                                  f"'https://github.com/kumo-ai/kumo-rfm'.")
 
 
-def format_value(value: Union[int, float]) -> str:
+def format_value(value: int | float) -> str:
     if value == int(value):
         return f'{int(value):,}'
     if abs(value) >= 1000: