kumoai 2.7.0.dev202508201830__cp312-cp312-win_amd64.whl → 2.12.0.dev202511111731__cp312-cp312-win_amd64.whl

kumoai/experimental/rfm/rfm.py

@@ -1,17 +1,32 @@
 import json
+import time
 import warnings
-from typing import List, Literal, Optional, Union
+from collections import defaultdict
+from collections.abc import Generator
+from contextlib import contextmanager
+from dataclasses import dataclass, replace
+from typing import (
+    Any,
+    Dict,
+    Iterator,
+    List,
+    Literal,
+    Optional,
+    Tuple,
+    Union,
+    overload,
+)
 
 import numpy as np
 import pandas as pd
 from kumoapi.model_plan import RunMode
-from kumoapi.pquery import QueryType
+from kumoapi.pquery import QueryType, ValidatedPredictiveQuery
+from kumoapi.rfm import Context
+from kumoapi.rfm import Explanation as ExplanationConfig
 from kumoapi.rfm import (
-    Context,
-    PQueryDefinition,
     RFMEvaluateRequest,
+    RFMParseQueryRequest,
     RFMPredictRequest,
-    RFMValidateQueryRequest,
 )
 from kumoapi.task import TaskType
 
@@ -20,11 +35,18 @@ from kumoai.exceptions import HTTPException
 from kumoai.experimental.rfm import LocalGraph
 from kumoai.experimental.rfm.local_graph_sampler import LocalGraphSampler
 from kumoai.experimental.rfm.local_graph_store import LocalGraphStore
-from kumoai.experimental.rfm.local_pquery_driver import LocalPQueryDriver
-from kumoai.utils import ProgressLogger
+from kumoai.experimental.rfm.local_pquery_driver import (
+    LocalPQueryDriver,
+    date_offset_to_seconds,
+)
+from kumoai.mixin import CastMixin
+from kumoai.utils import InteractiveProgressLogger, ProgressLogger
 
 _RANDOM_SEED = 42
 
+_MAX_PRED_SIZE: dict[TaskType, int] = defaultdict(lambda: 1_000)
+_MAX_PRED_SIZE[TaskType.TEMPORAL_LINK_PREDICTION] = 200
+
 _MAX_CONTEXT_SIZE = {
     RunMode.DEBUG: 100,
     RunMode.FAST: 1_000,
@@ -39,7 +61,7 @@ _MAX_TEST_SIZE = {  # Share test set size across run modes for fair comparison:
 }
 
 _MAX_SIZE = 30 * 1024 * 1024
-_SIZE_LIMIT_MSG = ("Context size exceeds the 30MB limit. {stats_msg}\nPlease "
+_SIZE_LIMIT_MSG = ("Context size exceeds the 30MB limit. {stats}\nPlease "
                    "reduce either the number of tables in the graph, their "
                    "number of columns (e.g., large text columns), "
                    "neighborhood configuration, or the run mode. If none of "
@@ -48,6 +70,51 @@ _SIZE_LIMIT_MSG = ("Context size exceeds the 30MB limit. {stats_msg}\nPlease "
                    "beyond this for your use-case.")
 
 
+@dataclass(repr=False)
+class ExplainConfig(CastMixin):
+    """Configuration for explainability.
+
+    Args:
+        skip_summary: Whether to skip generating a human-readable summary of
+            the explanation.
+    """
+    skip_summary: bool = False
+
+
+@dataclass(repr=False)
+class Explanation:
+    prediction: pd.DataFrame
+    summary: str
+    details: ExplanationConfig
+
+    @overload
+    def __getitem__(self, index: Literal[0]) -> pd.DataFrame:
+        pass
+
+    @overload
+    def __getitem__(self, index: Literal[1]) -> str:
+        pass
+
+    def __getitem__(self, index: int) -> Union[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]]:
+        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
+
+        display(self.prediction)
+        display(Markdown(self.summary))
+
+
 class KumoRFM:
     r"""The Kumo Relational Foundation model (RFM) from the `KumoRFM: A
     Foundation Model for In-Context Learning on Relational Data
@@ -98,35 +165,127 @@ class KumoRFM:
         self,
         graph: LocalGraph,
         preprocess: bool = False,
-        verbose: bool = True,
+        verbose: Union[bool, ProgressLogger] = True,
     ) -> None:
         graph = graph.validate()
         self._graph_def = graph._to_api_graph_definition()
         self._graph_store = LocalGraphStore(graph, preprocess, verbose)
         self._graph_sampler = LocalGraphSampler(self._graph_store)
 
+        self._batch_size: Optional[int | Literal['max']] = None
+        self.num_retries: int = 0
+
     def __repr__(self) -> str:
         return f'{self.__class__.__name__}()'
 
+    @contextmanager
+    def batch_mode(
+        self,
+        batch_size: Union[int, Literal['max']] = 'max',
+        num_retries: int = 1,
+    ) -> Generator[None, None, None]:
+        """Context manager to predict in batches.
+
+        .. code-block:: python
+
+            with model.batch_mode(batch_size='max', num_retries=1):
+                df = model.predict(query, indices=...)
+
+        Args:
+            batch_size: The batch size. If set to ``"max"``, will use the
+                maximum applicable batch size for the given task.
+            num_retries: The maximum number of retries for failed queries due
+                to unexpected server issues.
+        """
+        if batch_size != 'max' and batch_size <= 0:
+            raise ValueError(f"'batch_size' must be greater than zero "
+                             f"(got {batch_size})")
+
+        if num_retries < 0:
+            raise ValueError(f"'num_retries' must be greater than or equal to "
+                             f"zero (got {num_retries})")
+
+        self._batch_size = batch_size
+        self.num_retries = num_retries
+        yield
+        self._batch_size = None
+        self.num_retries = 0
+
+    @overload
     def predict(
         self,
         query: str,
+        indices: Union[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,
         num_hops: int = 2,
         max_pq_iterations: int = 20,
         random_seed: Optional[int] = _RANDOM_SEED,
-        verbose: bool = True,
+        verbose: Union[bool, ProgressLogger] = True,
+        use_prediction_time: bool = False,
     ) -> pd.DataFrame:
+        pass
+
+    @overload
+    def predict(
+        self,
+        query: str,
+        indices: Union[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,
+        num_hops: int = 2,
+        max_pq_iterations: int = 20,
+        random_seed: Optional[int] = _RANDOM_SEED,
+        verbose: Union[bool, ProgressLogger] = True,
+        use_prediction_time: bool = False,
+    ) -> Explanation:
+        pass
+
+    def predict(
+        self,
+        query: str,
+        indices: Union[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,
+        num_hops: int = 2,
+        max_pq_iterations: int = 20,
+        random_seed: Optional[int] = _RANDOM_SEED,
+        verbose: Union[bool, ProgressLogger] = True,
+        use_prediction_time: bool = False,
+    ) -> Union[pd.DataFrame, Explanation]:
         """Returns predictions for a predictive query.
 
         Args:
             query: The predictive query.
-            anchor_time: The anchor timestamp for the query. If set to
-                :obj:`None`, will use the maximum timestamp in the data.
-                If set to :`"entity"`, will use the timestamp of the entity.
+            indices: The entity primary keys to predict for. Will override the
+                indices given as part of the predictive query. Predictions will
+                be generated for all indices, independent of whether they
+                fulfill entity filter constraints. To pre-filter entities, use
+                :meth:`~KumoRFM.is_valid_entity`.
+            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"``.
+            anchor_time: The anchor timestamp for the prediction. If set to
+                ``None``, will use the maximum timestamp in the data.
+                If set to ``"entity"``, will use the timestamp of the entity.
+            context_anchor_time: The maximum anchor timestamp for context
+                examples. If set to ``None``, ``anchor_time`` will
+                determine the anchor time for context examples.
             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.
@@ -138,79 +297,244 @@ class KumoRFM:
                 entities to find valid labels.
             random_seed: A manual seed for generating pseudo-random numbers.
             verbose: Whether to print verbose output.
+            use_prediction_time: Whether to use the anchor timestamp as an
+                additional feature during prediction. This is typically
+                beneficial for time series forecasting tasks.
 
         Returns:
-            The predictions as a :class:`pandas.DataFrame`
+            The predictions as a :class:`pandas.DataFrame`.
+            If ``explain`` is provided, returns an :class:`Explanation` object
+            containing the prediction, summary, and details.
         """
-        explain = False
+        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 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 explain and run_mode in {RunMode.NORMAL, RunMode.BEST}:
+        if explain_config is not None and run_mode in {
+                RunMode.NORMAL, RunMode.BEST
+        }:
             warnings.warn(f"Explainability is currently only supported for "
                           f"run mode 'FAST' (got '{run_mode}'). Provided run "
                           f"mode has been reset. Please lower the run mode to "
                           f"suppress this warning.")
 
-        if explain:
-            assert query_def.entity.ids is not None
-            if len(query_def.entity.ids.value) > 1:
-                raise ValueError(
-                    f"Cannot explain predictions for more than a single "
-                    f"entity (got {len(query_def.entity.ids.value)})")
+        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:
+        if explain_config is not None:
             msg = f'[bold]EXPLAIN[/bold] {query_repr}'
         else:
             msg = f'[bold]PREDICT[/bold] {query_repr}'
 
-        with ProgressLogger(msg, verbose=verbose) as logger:
-            context = self._get_context(
-                query_def,
-                anchor_time=anchor_time,
-                run_mode=RunMode(run_mode),
-                num_neighbors=num_neighbors,
-                num_hops=num_hops,
-                max_pq_iterations=max_pq_iterations,
-                evaluate=False,
-                random_seed=random_seed,
-                logger=logger,
-            )
-            request = RFMPredictRequest(
-                context=context,
-                run_mode=RunMode(run_mode),
+        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]
+            else:
+                batch_size = self._batch_size
+
+            if batch_size is not None:
+                offsets = range(0, len(indices), batch_size)
+                batches = [indices[step:step + batch_size] for step in offsets]
+            else:
+                batches = [indices]
+
+            if len(batches) > 1:
+                logger.log(f"Splitting {len(indices):,} entities into "
+                           f"{len(batches):,} batches of size {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.
+                context = self._get_context(
+                    query=query_def,
+                    indices=batch,
+                    anchor_time=anchor_time,
+                    context_anchor_time=context_anchor_time,
+                    run_mode=RunMode(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,
+                )
+                request = RFMPredictRequest(
+                    context=context,
+                    run_mode=RunMode(run_mode),
+                    query=query_str,
+                    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:
+                    logger.log(f"Generated context of size "
+                               f"{len(_bytes) / (1024*1024):.2f}MB")
+
+                if len(_bytes) > _MAX_SIZE:
+                    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',
+                    )
+
+                for attempt in range(self.num_retries + 1):
+                    try:
+                        if explain_config is not None:
+                            resp = global_state.client.rfm_api.explain(
+                                request=_bytes,
+                                skip_summary=explain_config.skip_summary,
+                            )
+                            summary = resp.summary
+                            details = resp.details
+                        else:
+                            resp = global_state.client.rfm_api.predict(_bytes)
+                        df = pd.DataFrame(**resp.prediction)
+
+                        # 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]))
+
+                        # Cast 'ANCHOR_TIMESTAMP' to correct data type:
+                        if 'ANCHOR_TIMESTAMP' in df:
+                            ser = df['ANCHOR_TIMESTAMP']
+                            if not pd.api.types.is_datetime64_any_dtype(ser):
+                                if isinstance(ser.iloc[0], str):
+                                    unit = None
+                                else:
+                                    unit = 'ms'
+                                df['ANCHOR_TIMESTAMP'] = pd.to_datetime(
+                                    ser, errors='coerce', unit=unit)
+
+                        predictions.append(df)
+
+                        if (isinstance(verbose, InteractiveProgressLogger)
+                                and len(batches) > 1):
+                            verbose.step()
+
+                        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 "
+                                f"create an issue at "
+                                f"'https://github.com/kumo-ai/kumo-rfm'. {msg}"
+                            ) from None
+
+                        time.sleep(2**attempt)  # 1s, 2s, 4s, 8s, ...
+
+        if len(predictions) == 1:
+            prediction = predictions[0]
+        else:
+            prediction = pd.concat(predictions, ignore_index=True)
+
+        if explain_config is not None:
+            assert len(predictions) == 1
+            assert summary is not None
+            assert details is not None
+            return Explanation(
+                prediction=prediction,
+                summary=summary,
+                details=details,
             )
-            with warnings.catch_warnings():
-                warnings.filterwarnings('ignore', message='Protobuf gencode')
-                request_msg = request.to_protobuf()
-                request_bytes = request_msg.SerializeToString()
-            logger.log(f"Generated context of size "
-                       f"{len(request_bytes) / (1024*1024):.2f}MB")
 
-            if len(request_bytes) > _MAX_SIZE:
-                stats_msg = Context.get_memory_stats(request_msg.context)
-                raise ValueError(_SIZE_LIMIT_MSG.format(stats_msg=stats_msg))
+        return prediction
 
-            try:
-                if explain:
-                    resp = global_state.client.rfm_api.explain(request_bytes)
-                else:
-                    resp = global_state.client.rfm_api.predict(request_bytes)
-            except HTTPException as e:
-                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
+    def is_valid_entity(
+        self,
+        query: str,
+        indices: Union[List[str], List[float], List[int], None] = None,
+        *,
+        anchor_time: Union[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)
+        entity filter constraints.
 
-        return pd.DataFrame(**resp.prediction)
+        Args:
+            query: The predictive query.
+            indices: The entity primary keys to predict for. Will override the
+                indices given as part of the predictive query.
+            anchor_time: The anchor timestamp for the prediction. If set to
+                ``None``, will use the maximum timestamp in the data.
+                If set to ``"entity"``, will use the timestamp of the entity.
+        """
+        query_def = self._parse_query(query)
+
+        if indices is None:
+            if query_def.rfm_entity_ids is None:
+                raise ValueError("Cannot find entities to predict for. Please "
+                                 "pass them via "
+                                 "`is_valid_entity(query, indices=...)`")
+            indices = query_def.get_rfm_entity_id_list()
+
+        if len(indices) == 0:
+            raise ValueError("At least one entity is required")
+
+        if anchor_time is None:
+            anchor_time = self._graph_store.max_time
+
+        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):
+                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)
 
     def evaluate(
         self,
@@ -218,21 +542,26 @@ class KumoRFM:
         *,
         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,
         num_hops: int = 2,
         max_pq_iterations: int = 20,
         random_seed: Optional[int] = _RANDOM_SEED,
-        verbose: bool = True,
+        verbose: Union[bool, ProgressLogger] = True,
+        use_prediction_time: bool = False,
     ) -> pd.DataFrame:
         """Evaluates a predictive query.
 
         Args:
             query: The predictive query.
             metrics: The metrics to use.
-            anchor_time: The anchor timestamp for the query. If set to
-                :obj:`None`, will use the maximum timestamp in the data.
-                If set to :`"entity"`, will use the timestamp of the entity.
+            anchor_time: The anchor timestamp for the prediction. If set to
+                ``None``, will use the maximum timestamp in the data.
+                If set to ``"entity"``, will use the timestamp of the entity.
+            context_anchor_time: The maximum anchor timestamp for context
+                examples. If set to ``None``, ``anchor_time`` will
+                determine the anchor time for context examples.
             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.
@@ -244,6 +573,9 @@ class KumoRFM:
                 entities to find valid labels.
             random_seed: A manual seed for generating pseudo-random numbers.
             verbose: Whether to print verbose output.
+            use_prediction_time: Whether to use the anchor timestamp as an
+                additional feature during prediction. This is typically
+                beneficial for time series forecasting tasks.
 
         Returns:
             The metrics as a :class:`pandas.DataFrame`
@@ -254,13 +586,24 @@ class KumoRFM:
             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}'
 
-        with ProgressLogger(msg, verbose=verbose) as logger:
+        if not isinstance(verbose, ProgressLogger):
+            verbose = InteractiveProgressLogger(msg, verbose=verbose)
+
+        with verbose as logger:
             context = self._get_context(
-                query_def,
+                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,
@@ -276,6 +619,7 @@ class KumoRFM:
                 context=context,
                 run_mode=RunMode(run_mode),
                 metrics=metrics,
+                use_prediction_time=use_prediction_time,
             )
             with warnings.catch_warnings():
                 warnings.filterwarnings('ignore', message='Protobuf gencode')
@@ -334,17 +678,19 @@ class KumoRFM:
 
         if anchor_time is None:
             anchor_time = self._graph_store.max_time
-            anchor_time = anchor_time - query_def.target.end_offset
+            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)
 
         assert anchor_time is not None
         if isinstance(anchor_time, pd.Timestamp):
-            self._validate_time(query_def, anchor_time, evaluate=True)
+            self._validate_time(query_def, anchor_time, None, evaluate=True)
         else:
             assert anchor_time == 'entity'
-            if (query_def.entity.pkey.table_name
-                    not in self._graph_store.time_dict):
+            if (query_def.entity_table not in self._graph_store.time_dict):
                 raise ValueError(f"Anchor time 'entity' requires the entity "
-                                 f"table '{query_def.entity.pkey.table_name}' "
+                                 f"table '{query_def.entity_table}' "
                                  f"to have a time column")
 
         query_driver = LocalPQueryDriver(self._graph_store, query_def,
@@ -355,18 +701,22 @@ class KumoRFM:
             anchor_time=anchor_time,
             batch_size=min(10_000, size),
             max_iterations=max_iterations,
+            guarantee_train_examples=False,
         )
 
+        entity = self._graph_store.pkey_map_dict[
+            query_def.entity_table].index[node]
+
         return pd.DataFrame({
-            'ENTITY': node,
+            'ENTITY': entity,
             'ANCHOR_TIMESTAMP': time,
             'TARGET': y,
         })
 
     # Helpers #################################################################
 
-    def _parse_query(self, query: str) -> PQueryDefinition:
-        if isinstance(query, PQueryDefinition):
+    def _parse_query(self, query: str) -> ValidatedPredictiveQuery:
+        if isinstance(query, ValidatedPredictiveQuery):
             return query
 
         if isinstance(query, str) and query.strip()[:9].lower() == 'evaluate ':
@@ -376,12 +726,12 @@ class KumoRFM:
                              "predictions or evaluations.")
 
         try:
-            request = RFMValidateQueryRequest(
+            request = RFMParseQueryRequest(
                 query=query,
                 graph_definition=self._graph_def,
             )
 
-            resp = global_state.client.rfm_api.validate_query(request)
+            resp = global_state.client.rfm_api.parse_query(request)
             # TODO Expose validation warnings.
 
             if len(resp.validation_response.warnings) > 0:
@@ -392,7 +742,7 @@ class KumoRFM:
                 warnings.warn(f"Encountered the following warnings during "
                               f"parsing:\n{msg}")
 
-            return resp.query_definition
+            return resp.query
         except HTTPException as e:
             try:
                 msg = json.loads(e.detail)['detail']
@@ -403,8 +753,9 @@ class KumoRFM:
 
     def _validate_time(
         self,
-        query: PQueryDefinition,
+        query: ValidatedPredictiveQuery,
         anchor_time: pd.Timestamp,
+        context_anchor_time: Union[pd.Timestamp, None],
         evaluate: bool,
     ) -> None:
 
@@ -416,22 +767,45 @@ class KumoRFM:
                              f"the earliest timestamp "
                              f"'{self._graph_store.min_time}' in the data.")
 
-        if anchor_time - query.target.end_offset < self._graph_store.min_time:
-            raise ValueError(f"Anchor timestamp is too early or aggregation "
-                             f"time range is too large. To make this "
-                             f"prediction, we would need data back to "
-                             f"'{anchor_time - query.target.end_offset}', "
-                             f"however, your data only contains data back to "
+        if (context_anchor_time is not None
+                and context_anchor_time < self._graph_store.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}'.")
 
-        if (anchor_time - 2 * query.target.end_offset
-                < self._graph_store.min_time):
-            warnings.warn(f"Anchor timestamp is too early or aggregation "
-                          f"time range is too large. To form proper input "
-                          f"data, we would need data back to "
-                          f"'{anchor_time - 2 * query.target.end_offset}', "
-                          f"however, your data only contains data back to "
-                          f"'{self._graph_store.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
+        if (context_anchor_time is not None
+                and context_anchor_time > anchor_time):
+            warnings.warn(f"Context anchor timestamp "
+                          f"(got '{context_anchor_time}') is set to a later "
+                          f"date than the prediction anchor timestamp "
+                          f"(got '{anchor_time}'). Please make sure this is "
+                          f"intended.")
+        elif (query.query_type == QueryType.TEMPORAL
+              and context_anchor_time is not None
+              and context_anchor_time + forecast_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 "
+                          f"'{anchor_time}'. Please make sure this is "
+                          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
+            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}'.")
 
         if (not evaluate and anchor_time
                 > self._graph_store.max_time + pd.DateOffset(days=1)):
@@ -439,17 +813,18 @@ class KumoRFM:
                           f"latest timestamp '{self._graph_store.max_time}' "
                           f"in the data. Please make sure this is intended.")
 
-        if (evaluate and anchor_time
-                > self._graph_store.max_time - query.target.end_offset):
+        max_eval_time = self._graph_store.max_time - forecast_end_offset
+        if evaluate and anchor_time > max_eval_time:
             raise ValueError(
                 f"Anchor timestamp for evaluation is after the latest "
-                f"supported timestamp "
-                f"'{self._graph_store.max_time - query.target.end_offset}'.")
+                f"supported timestamp '{max_eval_time}'.")
 
     def _get_context(
         self,
-        query: PQueryDefinition,
+        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,
@@ -474,8 +849,8 @@ class KumoRFM:
                              f"must go beyond this for your use-case.")
 
         query_driver = LocalPQueryDriver(self._graph_store, query, random_seed)
-        task_type = query.get_task_type(
-            stypes=self._graph_store.stype_dict,
+        task_type = LocalPQueryDriver.get_task_type(
+            query,
             edge_types=self._graph_store.edge_types,
         )
 
@@ -507,28 +882,42 @@ class KumoRFM:
             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)
+        else:
+            end_offset = query.target_ast.date_offset_range.end_date_offset
+        forecast_end_offset = end_offset * query.num_forecasts
         if anchor_time is None:
             anchor_time = self._graph_store.max_time
             if evaluate:
-                anchor_time = anchor_time - query.target.end_offset
+                anchor_time = anchor_time - forecast_end_offset
             if logger is not None:
                 assert isinstance(anchor_time, pd.Timestamp)
-                if (anchor_time.hour == 0 and anchor_time.minute == 0
-                        and anchor_time.second == 0
-                        and anchor_time.microsecond == 0):
+                if anchor_time == pd.Timestamp.min:
+                    pass  # Static graph
+                elif (anchor_time.hour == 0 and anchor_time.minute == 0
+                      and anchor_time.second == 0
+                      and anchor_time.microsecond == 0):
                     logger.log(f"Derived anchor time {anchor_time.date()}")
                 else:
                     logger.log(f"Derived anchor time {anchor_time}")
 
         assert anchor_time is not None
         if isinstance(anchor_time, pd.Timestamp):
-            self._validate_time(query, anchor_time, evaluate)
+            if context_anchor_time is None:
+                context_anchor_time = anchor_time - forecast_end_offset
+            self._validate_time(query, anchor_time, context_anchor_time,
+                                evaluate)
         else:
             assert anchor_time == 'entity'
-            if query.entity.pkey.table_name not in self._graph_store.time_dict:
+            if query.entity_table not in self._graph_store.time_dict:
                 raise ValueError(f"Anchor time 'entity' requires the entity "
-                                 f"table '{query.entity.pkey.table_name}' to "
+                                 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:
@@ -540,6 +929,7 @@ class KumoRFM:
                 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:
@@ -563,34 +953,31 @@ class KumoRFM:
                 logger.log(msg)
 
         else:
-            assert query.entity.ids is not None
+            assert indices is not None
 
-            max_num_test = 200 if task_type.is_link_pred else 1000
-            if len(query.entity.ids.value) > max_num_test:
+            if len(indices) > _MAX_PRED_SIZE[task_type]:
                 raise ValueError(f"Cannot predict for more than "
-                                 f"{max_num_test:,} entities at once "
-                                 f"(got {len(query.entity.ids.value):,})")
+                                 f"{_MAX_PRED_SIZE[task_type]:,} entities at "
+                                 f"once (got {len(indices):,}). Use "
+                                 f"`KumoRFM.batch_mode` to process entities "
+                                 f"in batches")
 
             test_node = self._graph_store.get_node_id(
-                table_name=query.entity.pkey.table_name,
-                pkey=pd.Series(
-                    query.entity.ids.value,
-                    dtype=query.entity.ids.dtype,
-                ),
+                table_name=query.entity_table,
+                pkey=pd.Series(indices),
             )
 
             if isinstance(anchor_time, pd.Timestamp):
                 test_time = pd.Series(anchor_time).repeat(
                     len(test_node)).reset_index(drop=True)
             else:
-                time = self._graph_store.time_dict[
-                    query.entity.pkey.table_name]
+                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=anchor_time,
+            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,
@@ -617,12 +1004,23 @@ class KumoRFM:
                 raise NotImplementedError
             logger.log(msg)
 
-        entity_table_names = query.get_entity_table_names(
-            self._graph_store.edge_types)
+        entity_table_names: Tuple[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:
+                if edge_fkey == f'{edge_type[0]}.{edge_type[1]}':
+                    entity_table_names = (
+                        query.entity_table,
+                        edge_type[2],
+                    )
+        else:
+            entity_table_names = (query.entity_table, )
 
         # 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.exclude_cols_dict
+        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]] = []
@@ -642,6 +1040,17 @@ class KumoRFM:
             exclude_cols_dict=exclude_cols_dict,
         )
 
+        if len(subgraph.table_dict) >= 15:
+            raise ValueError(f"Cannot query from a graph with more than 15 "
+                             f"tables (got {len(subgraph.table_dict)}). "
+                             f"Please create a feature request at "
+                             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)
+
         return Context(
             task_type=task_type,
             entity_table_names=entity_table_names,
@@ -649,6 +1058,7 @@ class KumoRFM:
             y_train=y_train,
             y_test=y_test,
             top_k=query.top_k,
+            step_size=step_size,
         )
 
     @staticmethod
@@ -664,7 +1074,7 @@ class KumoRFM:
         elif task_type == TaskType.MULTICLASS_CLASSIFICATION:
             supported_metrics = ['acc', 'precision', 'recall', 'f1', 'mrr']
         elif task_type == TaskType.REGRESSION:
-            supported_metrics = ['mae', 'mape', 'mse', 'rmse', 'smape']
+            supported_metrics = ['mae', 'mape', 'mse', 'rmse', 'smape', 'r2']
         elif task_type == TaskType.TEMPORAL_LINK_PREDICTION:
             supported_metrics = [
                 'map@', 'ndcg@', 'mrr@', 'precision@', 'recall@', 'f1@',