rasa-pro 3.14.0.dev7__py3-none-any.whl → 3.14.0.dev8__py3-none-any.whl

rasa/utils/io.py

@@ -26,6 +26,7 @@ from typing_extensions import Protocol
 
 import rasa.shared.constants
 import rasa.shared.utils.io
+from rasa.shared.exceptions import RasaException
 
 if TYPE_CHECKING:
     from prompt_toolkit.validation import Validator
@@ -124,9 +125,7 @@ def create_path(file_path: Text) -> None:
 def file_type_validator(
     valid_file_types: List[Text], error_message: Text
 ) -> Type["Validator"]:
-    """Creates a `Validator` class which can be used with `questionary` to validate
-    file paths.
-    """
+    """Creates a file type validator class for the questionary package."""
 
     def is_valid(path: Text) -> bool:
         return path is not None and any(
@@ -137,9 +136,7 @@ def file_type_validator(
 
 
 def not_empty_validator(error_message: Text) -> Type["Validator"]:
-    """Creates a `Validator` class which can be used with `questionary` to validate
-    that the user entered something other than whitespace.
-    """
+    """Creates a not empty validator class for the questionary package."""
 
     def is_valid(input: Text) -> bool:
         return input is not None and input.strip() != ""
@@ -150,9 +147,7 @@ def not_empty_validator(error_message: Text) -> Type["Validator"]:
 def create_validator(
     function: Callable[[Text], bool], error_message: Text
 ) -> Type["Validator"]:
-    """Helper method to create `Validator` classes from callable functions. Should be
-    removed when questionary supports `Validator` objects.
-    """
+    """Helper method to create a validator class from a callable function."""
     from prompt_toolkit.document import Document
     from prompt_toolkit.validation import ValidationError, Validator
 
@@ -250,3 +245,26 @@ def write_yaml(
 
     with Path(target).open("w", encoding="utf-8") as outfile:
         dumper.dump(data, outfile, transform=transform)
+
+
+class InvalidPathException(RasaException):
+    """Raised if a path is invalid - e.g. path traversal is detected."""
+
+
+def subpath(parent: str, child: str) -> str:
+    """Return the path to the child directory of the parent directory.
+
+    Ensures, that child doesn't navigate to parent directories. Prevents
+    path traversal. Raises an InvalidPathException if the path is invalid.
+
+    Based on Snyk's directory traversal mitigation:
+    https://learn.snyk.io/lesson/directory-traversal/
+    """
+    safe_path = os.path.abspath(os.path.join(parent, child))
+    parent = os.path.abspath(parent)
+
+    common_base = os.path.commonpath([parent, safe_path])
+    if common_base != parent:
+        raise InvalidPathException(f"Invalid path: {safe_path}")
+
+    return safe_path

rasa/utils/json_utils.py

@@ -1,6 +1,6 @@
 import json
 from decimal import Decimal
-from typing import Any, Text
+from typing import Any, Dict, List, Text
 
 
 class DecimalEncoder(json.JSONEncoder):
@@ -58,3 +58,8 @@ def replace_decimals_with_floats(obj: Any) -> Any:
         Input `obj` with all `Decimal` types replaced by `float`s.
     """
     return json.loads(json.dumps(obj, cls=DecimalEncoder))
+
+
+def extract_values(data: Dict, keys: List[Text]) -> Dict:
+    """Extracts values for given keys from a dictionary."""
+    return {key: data.get(key) for key in keys if data.get(key)}

rasa/utils/log_utils.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 import logging
 import os
 import sys
-from typing import Any, Optional
+from typing import Any, List, Optional
 
 import structlog
 from structlog.dev import ConsoleRenderer
@@ -37,6 +37,7 @@ class HumanConsoleRenderer(ConsoleRenderer):
 def configure_structlog(
     log_level: Optional[int] = None,
     include_time: bool = False,
+    additional_processors: Optional[List[structlog.typing.Processor]] = None,
 ) -> None:
     """Configure logging of the server."""
     if log_level is None:  # Log level NOTSET is 0 so we use `is None` here
@@ -75,6 +76,9 @@ def configure_structlog(
     if include_time:
         shared_processors.append(structlog.processors.TimeStamper(fmt="iso"))
 
+    if additional_processors:
+        shared_processors.extend(additional_processors)
+
     if not FORCE_JSON_LOGGING and sys.stderr.isatty():
         # Pretty printing when we run in a terminal session.
         # Automatically prints pretty tracebacks when "rich" is installed

rasa/utils/openapi.py

@@ -0,0 +1,144 @@
+from typing import Any, Dict, List, Type
+
+from pydantic.main import BaseModel
+from sanic_openapi import openapi
+from sanic_openapi.openapi3.types import Schema
+
+_SUPPORTED_ATTRIBUTES = frozenset(["format", "enum", "required", "example"])
+
+
+def _to_schema(
+    definition_stack: List[str], schema_def: Dict[str, Any], definitions: Dict[str, Any]
+) -> Schema:
+    type = schema_def.get("type")
+
+    if type == "object":
+        properties_spec = schema_def.get("properties", {})
+        properties = {}
+        for key in properties_spec:
+            properties[key] = _to_schema(
+                definition_stack=definition_stack,
+                schema_def=properties_spec[key],
+                definitions=definitions,
+            )
+        schema = openapi.Object(
+            title=schema_def.get("title"),
+            description=schema_def.get("description"),
+            required=schema_def.get("required"),
+            properties=properties,
+        )
+    elif type == "array":
+        schema = openapi.Array(
+            description=schema_def.get("description"),
+            required=schema_def.get("required"),
+            items=_to_schema(
+                definition_stack=definition_stack,
+                schema_def=schema_def.get("items"),
+                definitions=definitions,
+            ),
+        )
+    elif type is None:
+        if allof_spec := schema_def.get("allOf"):  # Model, Enum
+            definition = allof_spec[0]["$ref"].split("/")[-1]
+            definition_data = definitions.get(definition)
+            if definition_data is None:
+                schema = openapi.Object(
+                    title=definition, description=schema_def.get("description")
+                )
+            else:
+                schema = (
+                    _to_schema(
+                        definition_stack=definition_stack + [definition],
+                        schema_def={**definition_data},
+                        definitions=definitions,
+                    )
+                    if definition not in definition_stack
+                    else openapi.Object(
+                        title=definition, description=schema_def.get("description")
+                    )
+                )
+
+        elif anyof_spec := schema_def.get("anyOf"):  # Union
+            anyof = []
+            for any in anyof_spec:
+                if any.get("type"):
+                    schema_type_obj = Schema(
+                        **{
+                            "type": any.get("type"),
+                            "description": any.get("description"),
+                        }
+                    )
+                    anyof.append(schema_type_obj)
+                else:
+                    definition = any["$ref"].split("/")[-1]
+                    if definition not in definition_stack:
+                        definition_data = definitions.get(definition)
+                        if definition_data is not None:
+                            anyof.append(
+                                _to_schema(
+                                    definition_stack=definition_stack + [definition],
+                                    schema_def=definition_data,
+                                    definitions=definitions,
+                                )
+                            )
+                        else:
+                            anyof.append(
+                                openapi.Object(
+                                    title=definition,
+                                    description=schema_def.get(
+                                        "description", definition
+                                    ),
+                                    properties={},
+                                )
+                            )
+                    else:
+                        anyof.append(
+                            openapi.Object(
+                                title=definition,
+                                description=schema_def.get("description", definition),
+                                properties={},
+                            )
+                        )
+            schema = Schema(anyOf=anyof)
+        elif ref := schema_def.get("$ref"):  # $ref
+            definition = ref.split("/")[-1]
+            definition_data = definitions.get(definition)
+            if definition_data is not None:
+                schema = _to_schema(
+                    definition_stack=definition_stack,
+                    schema_def=definition_data,
+                    definitions=definitions,
+                )
+            else:
+                schema = openapi.Object(
+                    title=definition, description=schema_def.get("description")
+                )
+        else:  # Any type
+            schema = Schema(
+                **{"type": "object", "description": schema_def.get("description")}
+            )
+
+    else:
+        schema_spec = {
+            "type": schema_def.get("type"),
+            "description": schema_def.get("description"),
+        }
+        for spec in _SUPPORTED_ATTRIBUTES:
+            if schema_def.get(spec):
+                schema_spec[spec] = schema_def.get(spec)
+        schema = Schema(**schema_spec)
+
+    return schema
+
+
+def model_to_schema(model: Type[BaseModel]) -> Schema:
+    schema = model.model_json_schema()
+    # Handle both $defs (newer JSON Schema) and definitions (older JSON Schema)
+    definitions = schema.get("$defs") or schema.get("definitions") or {}
+    return _to_schema(
+        definition_stack=[],
+        schema_def=dict(
+            filter(lambda key: key[0] not in ("definitions", "$defs"), schema.items())
+        ),
+        definitions=definitions,
+    )

rasa/utils/tensorflow/__init__.py

@@ -0,0 +1,29 @@
+"""TensorFlow utilities for Rasa components.
+
+This module provides TensorFlow-related utilities and components.
+TensorFlow is an optional dependency and should be installed separately
+if you want to use components that require it. These are:
+- DIETClassifier
+- TEDPolicy
+- UnexpecTEDIntentPolicy
+- ResponseSelector
+- ConveRTFeaturizer
+- LanguageModelFeaturizer
+
+To install Rasa with TensorFlow support:
+`pip install "rasa[tensorflow]"`
+
+To install it with poetry:
+`poetry install --extras tensorflow`
+
+For macOS with Apple Silicon (M1/M2) (platform-specific TensorFlow installation):
+`pip install "rasa[tensorflow,metal]"`
+"""
+
+import importlib.util
+import logging
+
+logger = logging.getLogger(__name__)
+
+# check if TensorFlow is available
+TENSORFLOW_AVAILABLE = importlib.util.find_spec("tensorflow") is not None

rasa/utils/tensorflow/callback.py

@@ -55,7 +55,7 @@ class RasaModelCheckpoint(tf.keras.callbacks.Callback):
         """
         super().__init__()
 
-        self.checkpoint_file = checkpoint_dir / "checkpoint.tf_model"
+        self.checkpoint_file = checkpoint_dir / "checkpoint.weights.h5"
         self.best_metrics_so_far: Dict[Text, Any] = {}
 
     def on_epoch_end(self, epoch: int, logs: Optional[Dict[Text, Any]] = None) -> None:

rasa/utils/tensorflow/crf.py

@@ -9,7 +9,7 @@ from tensorflow.types.experimental import TensorLike
 # (modified to our neeeds)
 
 
-class CrfDecodeForwardRnnCell(tf.keras.layers.AbstractRNNCell):
+class CrfDecodeForwardRnnCell(tf.keras.layers.Layer):
     """Computes the forward decoding in a linear-chain CRF."""
 
     def __init__(self, transition_params: TensorLike, **kwargs: Any) -> None:

rasa/utils/tensorflow/data_generator.py

@@ -71,13 +71,22 @@ class RasaDataGenerator(Sequence):
         # balancing on the next epoch
         return data
 
+    @staticmethod
+    def _create_default_array() -> np.ndarray:
+        """Create a default empty array for missing features.
+
+        Returns:
+            A default empty array with shape (0, 1) and dtype float32.
+        """
+        return np.zeros((0, 1), dtype=np.float32)
+
     @staticmethod
     def prepare_batch(
         data: Data,
         start: Optional[int] = None,
         end: Optional[int] = None,
         tuple_sizes: Optional[Dict[Text, int]] = None,
-    ) -> Tuple[Optional[np.ndarray], ...]:
+    ) -> Tuple[np.ndarray, ...]:
         """Slices model data into batch using given start and end value.
 
         Args:
@@ -85,8 +94,8 @@ class RasaDataGenerator(Sequence):
             start: The start index of the batch
             end: The end index of the batch
             tuple_sizes: In case the feature is not present we propagate the batch with
-              None. Tuple sizes contains the number of how many None values to add for
-              what kind of feature.
+              default arrays. Tuple sizes contains the number of how many default values
+              to add for what kind of feature.
 
         Returns:
             The features of the batch.
@@ -95,12 +104,14 @@ class RasaDataGenerator(Sequence):
 
         for key, attribute_data in data.items():
             for sub_key, f_data in attribute_data.items():
-                # add None for not present values during processing
+                # add default arrays for not present values during processing
                 if not f_data:
                     if tuple_sizes:
-                        batch_data += [None] * tuple_sizes[key]
+                        batch_data += [
+                            RasaDataGenerator._create_default_array()
+                        ] * tuple_sizes[key]
                     else:
-                        batch_data.append(None)
+                        batch_data.append(RasaDataGenerator._create_default_array())
                     continue
 
                 for v in f_data:
@@ -409,8 +420,10 @@ class RasaBatchDataGenerator(RasaDataGenerator):
         end = start + self._current_batch_size
 
         # return input and target data, as our target data is inside the input
-        # data return None for the target data
-        return self.prepare_batch(self._data, start, end), None
+        # data return default array for the target data
+        return self.prepare_batch(
+            self._data, start, end
+        ), RasaDataGenerator._create_default_array()
 
     def on_epoch_end(self) -> None:
         """Update the data after every epoch."""

rasa/utils/tensorflow/layers.py

@@ -3,9 +3,7 @@ from typing import Any, Callable, List, Optional, Text, Tuple, Union
 
 import tensorflow as tf
 import tensorflow.keras.backend as K
-
-# TODO: The following is not (yet) available via tf.keras
-from keras.src.utils.control_flow_util import smart_cond
+from tensorflow.python.keras.utils.control_flow_util import smart_cond
 
 import rasa.utils.tensorflow.crf
 import rasa.utils.tensorflow.layers_utils as layers_utils
@@ -278,6 +276,7 @@ class RandomlyConnectedDense(tf.keras.layers.Dense):
             kernel_constraint: Constraint function applied to
                 the `kernel` weights matrix.
             bias_constraint: Constraint function applied to the bias vector.
+            **kwargs: Additional keyword arguments passed to the parent class.
         """
         super().__init__(**kwargs)
 
@@ -367,7 +366,12 @@ class RandomlyConnectedDense(tf.keras.layers.Dense):
         Returns:
             The processed inputs.
         """
-        if self.density < 1.0:
+        # Apply kernel masking if needed (Keras 3.x compatibility check)
+        if (
+            self.density < 1.0
+            and hasattr(self, "kernel_mask")
+            and self.kernel_mask is not None
+        ):
             # Set fraction of the `kernel` weights to zero according to precomputed mask
             self.kernel.assign(self.kernel * self.kernel_mask)
         return super().call(inputs)
@@ -724,6 +728,7 @@ class DotProductLoss(tf.keras.layers.Layer):
                 Currently, the only possible value is `SOFTMAX`.
             similarity_type: Similarity measure to use, either `cosine` or `inner`.
             name: Optional name of the layer.
+            **kwargs: Additional keyword arguments passed to the parent class.
 
         Raises:
             TFLayerConfigException: When `similarity_type` is not one of `COSINE` or
@@ -883,6 +888,7 @@ class SingleLabelDotProductLoss(DotProductLoss):
                 values are approximately bounded.
             model_confidence: Normalization of confidence values during inference.
                 Currently, the only possible value is `SOFTMAX`.
+            **kwargs: Additional keyword arguments passed to the parent class.
         """
         super().__init__(
             num_candidates,
@@ -1244,6 +1250,7 @@ class MultiLabelDotProductLoss(DotProductLoss):
                 Used inside _loss_cross_entropy() only.
             model_confidence: Normalization of confidence values during inference.
                 Currently, the only possible value is `SOFTMAX`.
+            **kwargs: Additional keyword arguments passed to the parent class.
         """
         super().__init__(
             num_candidates,

rasa/utils/tensorflow/metrics.py

@@ -1,7 +1,6 @@
 from typing import Any, Dict, Optional
 
 import tensorflow as tf
-from tensorflow.keras import backend as K
 from tensorflow.types.experimental import TensorLike
 
 # original code taken from
@@ -118,7 +117,7 @@ class FBetaScore(tf.keras.metrics.Metric):
 
         def _zero_wt_init(name: Any) -> Any:
             return self.add_weight(
-                name, shape=self.init_shape, initializer="zeros", dtype=self.dtype
+                name=name, shape=self.init_shape, initializer="zeros", dtype=self.dtype
             )
 
         self.true_positives = _zero_wt_init("true_positives")
@@ -197,7 +196,12 @@ class FBetaScore(tf.keras.metrics.Metric):
 
     def reset_state(self) -> None:
         reset_value = tf.zeros(self.init_shape, dtype=self.dtype)
-        K.batch_set_value([(v, reset_value) for v in self.variables])
+        # In Keras 3.x, self.variables contains string names, not variable objects,
+        # so each metric variable is reset using assign() instead of K.batch_set_value()
+        self.true_positives.assign(reset_value)
+        self.false_positives.assign(reset_value)
+        self.false_negatives.assign(reset_value)
+        self.weights_intermediate.assign(reset_value)
 
     def reset_states(self) -> None:
         # Backwards compatibility alias of `reset_state`. New classes should

rasa/utils/tensorflow/models.py

@@ -8,7 +8,6 @@ from typing import TYPE_CHECKING, Any, Dict, List, Optional, Text, Tuple, Union
 import numpy as np
 import tensorflow as tf
 from keras import Model
-from keras.src.utils import tf_utils
 
 import rasa.utils.train_utils
 from rasa.shared.constants import DIAGNOSTIC_DATA
@@ -78,6 +77,7 @@ class RasaModel(Model):
 
         Args:
             random_seed: set the random seed to get reproducible results
+            **kwargs: Additional keyword arguments passed to the parent class
         """
         # make sure that keras releases resources from previously trained model
         tf.keras.backend.clear_session()
@@ -273,7 +273,8 @@ class RasaModel(Model):
         if self._run_eagerly:
             # Once we take advantage of TF's distributed training, this is where
             # scheduled functions will be forced to execute and return actual values.
-            outputs = tf_utils.sync_to_numpy_or_python_type(self.predict_step(batch_in))
+            step_output = self.predict_step(batch_in)
+            outputs = self._convert_tensors_to_numpy(step_output)
             if DIAGNOSTIC_DATA in outputs:
                 outputs[DIAGNOSTIC_DATA] = self._empty_lists_to_none_in_dict(
                     outputs[DIAGNOSTIC_DATA]
@@ -287,9 +288,8 @@ class RasaModel(Model):
 
         # Once we take advantage of TF's distributed training, this is where
         # scheduled functions will be forced to execute and return actual values.
-        outputs = tf_utils.sync_to_numpy_or_python_type(
-            self._tf_predict_step(list(batch_in))
-        )
+        step_output = self._tf_predict_step(list(batch_in))
+        outputs = self._convert_tensors_to_numpy(step_output)
         if DIAGNOSTIC_DATA in outputs:
             outputs[DIAGNOSTIC_DATA] = self._empty_lists_to_none_in_dict(
                 outputs[DIAGNOSTIC_DATA]
@@ -388,6 +388,23 @@ class RasaModel(Model):
 
         return {k: _recurse(v) for k, v in input_dict.items()}
 
+    def _convert_tensors_to_numpy(
+        self, step_output: Dict[Text, Any]
+    ) -> Dict[Text, Any]:
+        """Convert TensorFlow tensors to numpy arrays for Keras 3.x compatibility.
+
+        Replaces the deprecated tf_utils.sync_to_numpy_or_python_type() function.
+        Converts tensors (objects with 'numpy' method) to numpy arrays,
+        leaves others unchanged.
+        """
+        outputs = {}
+        for key, value in step_output.items():
+            if hasattr(value, "numpy"):
+                outputs[key] = value.numpy()
+            else:
+                outputs[key] = value
+        return outputs
+
     def _get_metric_results(self, prefix: Optional[Text] = "") -> Dict[Text, float]:
         return {
             f"{prefix}{metric.name}": metric.result()
@@ -403,7 +420,18 @@ class RasaModel(Model):
             overwrite: If 'True' an already existing model with the same file name will
                        be overwritten.
         """
-        self.save_weights(model_file_name, overwrite=overwrite, save_format="tf")
+        # Ensure filename ends with .weights.h5 and model is built for Keras 3.x
+        # compatibility
+        model_file_name = str(model_file_name)
+        if not model_file_name.endswith(".weights.h5"):
+            model_file_name += ".weights.h5"
+
+        if not self.built:
+            import tensorflow as tf
+
+            _ = self(tf.zeros((1, 1)))
+
+        self.save_weights(model_file_name, overwrite=overwrite)
 
     @classmethod
     def load(
@@ -444,6 +472,13 @@ class RasaModel(Model):
         )
         data_generator = RasaBatchDataGenerator(model_data_example, batch_size=1)
         model.fit(data_generator, verbose=False)
+
+        # Ensure model is built before loading weights (Keras 3 compatibility)
+        if not model.built:
+            # Build the model by calling it on sample data
+            sample_batch = next(iter(data_generator))
+            _ = model(sample_batch)
+
         # load trained weights
         model.load_weights(model_file_name)
 

rasa/utils/tensorflow/rasa_layers.py

@@ -301,12 +301,12 @@ class ConcatenateSparseDenseFeatures(RasaCustomLayer):
     ) -> tf.Tensor:
         """Turns sparse tensor into dense, possibly adds dropout before and/or after."""
         if self.SPARSE_DROPOUT in self._tf_layers:
-            feature = self._tf_layers[self.SPARSE_DROPOUT](feature, training)
+            feature = self._tf_layers[self.SPARSE_DROPOUT](feature, training=training)
 
         feature = self._tf_layers[self.SPARSE_TO_DENSE](feature)
 
         if self.DENSE_DROPOUT in self._tf_layers:
-            feature = self._tf_layers[self.DENSE_DROPOUT](feature, training)
+            feature = self._tf_layers[self.DENSE_DROPOUT](feature, training=training)
 
         return feature
 
@@ -1002,7 +1002,9 @@ class RasaSequenceLayer(RasaCustomLayer):
         ]((sequence_features, sentence_features, sequence_feature_lengths))
 
         # Apply one or more dense layers.
-        seq_sent_features = self._tf_layers[self.FFNN](seq_sent_features, training)
+        seq_sent_features = self._tf_layers[self.FFNN](
+            seq_sent_features, training=training
+        )
 
         # If using masked language modeling, mask the transformer inputs and get labels
         # for the masked tokens and a boolean mask. Note that TED does not use MLM loss,
@@ -1031,7 +1033,7 @@ class RasaSequenceLayer(RasaCustomLayer):
         if self._has_transformer:
             mask_padding = 1 - mask_combined_sequence_sentence
             outputs, attention_weights = self._tf_layers[self.TRANSFORMER](
-                seq_sent_features_masked, mask_padding, training
+                seq_sent_features_masked, mask_padding, training=training
             )
             outputs = tf.nn.gelu(outputs)
         else:

rasa/utils/tensorflow/transformer.py

@@ -2,10 +2,8 @@ from typing import Optional, Text, Tuple, Union
 
 import numpy as np
 import tensorflow as tf
-
-# TODO: The following is not (yet) available via tf.keras
-from keras.src.utils.control_flow_util import smart_cond
 from tensorflow.keras import backend as K
+from tensorflow.python.keras.utils.control_flow_util import smart_cond
 
 from rasa.utils.tensorflow.exceptions import TFLayerConfigException
 from rasa.utils.tensorflow.layers import RandomlyConnectedDense
@@ -280,6 +278,7 @@ class MultiHeadAttention(tf.keras.layers.Layer):
             value: A tensor with shape (..., length, depth).
             pad_mask: Float tensor with shape broadcastable
                 to (..., length, length). Defaults to None.
+            training: A tensor
 
         Returns:
             output: A tensor with shape (..., length, depth).