rasa-pro 3.14.1__py3-none-any.whl → 3.15.0a3__py3-none-any.whl

rasa/builder/evaluator/copilot_executor.py

@@ -0,0 +1,89 @@
+"""Copilot execution utilities for evaluators.
+
+This module provides utilities for running copilot operations in evaluation contexts,
+independent of specific evaluation frameworks like Langfuse.
+"""
+
+from typing import List, Optional
+
+import structlog
+from pydantic import BaseModel
+
+from rasa.builder.config import COPILOT_HANDLER_ROLLING_BUFFER_SIZE
+from rasa.builder.copilot.models import (
+    CopilotContext,
+    CopilotGenerationContext,
+    GeneratedContent,
+    ReferenceSection,
+    ResponseCategory,
+)
+from rasa.builder.llm_service import llm_service
+
+structlogger = structlog.get_logger()
+
+
+class CopilotRunResult(BaseModel):
+    """Result from running the copilot with response handler."""
+
+    complete_response: Optional[str]
+    response_category: Optional[ResponseCategory]
+    reference_section: Optional[ReferenceSection]
+    generation_context: CopilotGenerationContext
+
+
+async def run_copilot_with_response_handler(
+    context: CopilotContext,
+) -> Optional[CopilotRunResult]:
+    """Run the copilot with response handler on the given context.
+
+    This function encapsulates the core copilot execution logic. It handles:
+    - Instantiating the copilot and response handler
+    - Generating a response and extracting the reference section from the given context
+    - Returning structured results
+
+    Args:
+        context: The copilot context to process.
+
+    Returns:
+        CopilotRunResult containing the complete response, category, and generation
+        context, or None if execution fails.
+
+    Raises:
+        Any exceptions from the copilot or response handler execution.
+    """
+    # Instantiate the copilot and response handler
+    copilot = llm_service.instantiate_copilot()
+    copilot_response_handler = llm_service.instantiate_handler(
+        COPILOT_HANDLER_ROLLING_BUFFER_SIZE
+    )
+
+    # Call the copilot to generate a response and handle it with the response
+    # handler
+    (original_stream, generation_context) = await copilot.generate_response(context)
+    intercepted_stream = copilot_response_handler.handle_response(original_stream)
+
+    # Exhaust the stream to get the complete response for evaluation
+    response_chunks: List[str] = []
+    response_category = None
+    async for chunk in intercepted_stream:
+        if not isinstance(chunk, GeneratedContent):
+            continue
+        response_chunks.append(chunk.content)
+        response_category = chunk.response_category
+
+    complete_response = "".join(response_chunks) if response_chunks else None
+
+    # Extract the reference section from the response handler
+    if generation_context.relevant_documents:
+        reference_section = copilot_response_handler.extract_references(
+            generation_context.relevant_documents
+        )
+    else:
+        reference_section = None
+
+    return CopilotRunResult(
+        complete_response=complete_response,
+        response_category=response_category,
+        reference_section=reference_section,
+        generation_context=generation_context,
+    )

rasa/builder/evaluator/dataset/models.py

@@ -0,0 +1,173 @@
+from typing import Any, Dict, List, Optional, Union
+
+import structlog
+from pydantic import BaseModel, Field, field_validator
+
+from rasa.builder.copilot.models import (
+    ChatMessage,
+    CopilotContext,
+    EventContent,
+    ReferenceEntry,
+    ResponseCategory,
+    create_chat_message_from_dict,
+)
+from rasa.builder.document_retrieval.models import Document
+from rasa.builder.shared.tracker_context import TrackerContext
+
+structlogger = structlog.get_logger()
+
+
+class DatasetInput(BaseModel):
+    """Model for the input field of a dataset entry."""
+
+    message: Optional[str] = None
+    tracker_event_attachments: List[EventContent] = Field(default_factory=list)
+
+
+class DatasetExpectedOutput(BaseModel):
+    """Model for the expected_output field of a dataset entry."""
+
+    answer: str
+    response_category: ResponseCategory
+    references: list[ReferenceEntry]
+
+
+class DatasetMetadataCopilotAdditionalContext(BaseModel):
+    """Model for the copilot_additional_context in metadata."""
+
+    relevant_documents: List[Document] = Field(default_factory=list)
+    relevant_assistant_files: Dict[str, str] = Field(default_factory=dict)
+    assistant_tracker_context: Optional[Dict[str, Any]] = None
+    assistant_logs: str = Field(default="")
+    copilot_chat_history: List[ChatMessage] = Field(default_factory=list)
+
+    @field_validator("copilot_chat_history", mode="before")
+    @classmethod
+    def parse_chat_history(
+        cls, v: Union[List[Dict[str, Any]], List[ChatMessage]]
+    ) -> List[ChatMessage]:
+        """Manually parse chat history messages based on role field."""
+        # If already parsed ChatMessage objects, return them as-is
+        if (
+            v
+            and isinstance(v, list)
+            and all(isinstance(item, ChatMessage) for item in v)
+        ):
+            return v  # type: ignore[return-value]
+
+        # Check for mixed types (some ChatMessage, some not)
+        if (
+            v
+            and isinstance(v, list)
+            and any(isinstance(item, ChatMessage) for item in v)
+        ):
+            message = (
+                "Mixed types in copilot_chat_history: cannot mix ChatMessage objects "
+                "with other types."
+            )
+            structlogger.error(
+                "dataset_entry.parse_chat_history.mixed_types",
+                event_info=message,
+                chat_history_types=[type(item) for item in v],
+            )
+            raise ValueError(message)
+
+        # Otherwise, parse from dictionaries
+        parsed_messages: List[ChatMessage] = []
+        for message_data in v:
+            chat_message = create_chat_message_from_dict(message_data)
+            parsed_messages.append(chat_message)
+        return parsed_messages
+
+
+class DatasetMetadata(BaseModel):
+    """Model for the metadata field of a dataset entry."""
+
+    ids: Dict[str, str] = Field(default_factory=dict)
+    copilot_additional_context: DatasetMetadataCopilotAdditionalContext = Field(
+        default_factory=DatasetMetadataCopilotAdditionalContext
+    )
+
+
+class DatasetEntry(BaseModel):
+    """Pydantic model for dataset entries from Langfuse ExperimentItem."""
+
+    # Basic fields from ExperimentItem
+    id: str
+    input: DatasetInput
+    expected_output: DatasetExpectedOutput
+    metadata: DatasetMetadata
+
+    def to_copilot_context(self) -> CopilotContext:
+        """Create a CopilotContext from the dataset entry.
+
+        Raises:
+            ValueError: If the metadata is None, as it's required for creating a valid
+                CopilotContext.
+
+        Returns:
+            CopilotContext with all the context information.
+        """
+        if self.metadata is None:
+            message = (
+                f"Cannot create CopilotContext from dataset item with id: {self.id}. "
+                f"Metadata is required but was None."
+            )
+            structlogger.error(
+                "dataset_entry.to_copilot_context.metadata_is_none",
+                event_info=message,
+                item_id=self.id,
+                item_metadata=self.metadata,
+            )
+            raise ValueError(message)
+
+        # Parse tracker context if available
+        tracker_context = None
+        if (
+            self.metadata.copilot_additional_context.assistant_tracker_context
+            is not None
+        ):
+            tracker_context = TrackerContext(
+                **self.metadata.copilot_additional_context.assistant_tracker_context
+            )
+
+        return CopilotContext(
+            tracker_context=tracker_context,
+            assistant_logs=self.metadata.copilot_additional_context.assistant_logs,
+            assistant_files=self.metadata.copilot_additional_context.relevant_assistant_files,
+            copilot_chat_history=self.metadata.copilot_additional_context.copilot_chat_history,
+        )
+
+    @classmethod
+    def from_raw_data(
+        cls,
+        id: str,
+        input_data: Dict[str, Any],
+        expected_output_data: Dict[str, Any],
+        metadata_data: Dict[str, Any],
+    ) -> "DatasetEntry":
+        """Create a DatasetEntry from raw dictionary data.
+
+        Args:
+            id: The dataset entry ID.
+            input_data: Raw input dictionary.
+            expected_output_data: Raw expected output dictionary.
+            metadata_data: Raw metadata dictionary with all the additional context
+                used to generate the Copilot response.
+
+        Returns:
+            DatasetEntry with parsed data.
+        """
+        # Use Pydantic's model_validate to parse nested structures
+        dataset_input = DatasetInput.model_validate(input_data)
+        dataset_expected_output = DatasetExpectedOutput.model_validate(
+            expected_output_data
+        )
+        dataset_metadata = DatasetMetadata.model_validate(metadata_data)
+
+        return cls(
+            id=id,
+            input=dataset_input,
+            expected_output=dataset_expected_output,
+            metadata=dataset_metadata,
+        )

rasa/builder/evaluator/exceptions.py

@@ -0,0 +1,4 @@
+class EvaluationError(Exception):
+    """Base exception for evaluation-related errors."""
+
+    pass

rasa/builder/evaluator/response_classification/constants.py

@@ -0,0 +1,66 @@
+"""Constants for the response classification evaluator."""
+
+from typing import List, Literal
+
+# Options for averaging methods for Response Classification Evaluator
+MICRO_AVERAGING_METHOD: Literal["micro"] = "micro"
+MACRO_AVERAGING_METHOD: Literal["macro"] = "macro"
+WEIGHTED_AVERAGING_METHOD: Literal["weighted"] = "weighted"
+
+AVERAGING_METHODS: List[Literal["micro", "macro", "weighted"]] = [
+    MICRO_AVERAGING_METHOD,
+    MACRO_AVERAGING_METHOD,
+    WEIGHTED_AVERAGING_METHOD,
+]
+
+# Overall evaluation metric names
+MICRO_PRECISION_METRIC = "micro_precision"
+MACRO_PRECISION_METRIC = "macro_precision"
+WEIGHTED_PRECISION_METRIC = "weighted_precision"
+
+MICRO_RECALL_METRIC = "micro_recall"
+MACRO_RECALL_METRIC = "macro_recall"
+WEIGHTED_RECALL_METRIC = "weighted_recall"
+
+MICRO_F1_METRIC = "micro_f1"
+MACRO_F1_METRIC = "macro_f1"
+WEIGHTED_F1_METRIC = "weighted_f1"
+
+# Skip count metric name due to invalid data
+SKIP_COUNT_METRIC = "skipped_items"
+
+# Per-class evaluation metric name templates
+PER_CLASS_PRECISION_METRIC_TEMPLATE = "{category}_precision"
+PER_CLASS_RECALL_METRIC_TEMPLATE = "{category}_recall"
+PER_CLASS_F1_METRIC_TEMPLATE = "{category}_f1"
+PER_CLASS_SUPPORT_METRIC_TEMPLATE = "{category}_support"
+
+# Description templates for evaluation metrics
+MICRO_PRECISION_DESCRIPTION = "Micro Precision: {value:.3f}"
+MACRO_PRECISION_DESCRIPTION = "Macro Precision: {value:.3f}"
+WEIGHTED_PRECISION_DESCRIPTION = "Weighted Precision: {value:.3f}"
+
+MICRO_RECALL_DESCRIPTION = "Micro Recall: {value:.3f}"
+MACRO_RECALL_DESCRIPTION = "Macro Recall: {value:.3f}"
+WEIGHTED_RECALL_DESCRIPTION = "Weighted Recall: {value:.3f}"
+
+MICRO_F1_DESCRIPTION = "Micro F1: {value:.3f}"
+MACRO_F1_DESCRIPTION = "Macro F1: {value:.3f}"
+WEIGHTED_F1_DESCRIPTION = "Weighted F1: {value:.3f}"
+
+# Skip count metric description
+SKIP_COUNT_DESCRIPTION = "Skipped {value} items due to invalid data"
+
+# Per-class description templates
+PER_CLASS_PRECISION_DESCRIPTION = "[{category}] Precision: {value:.3f}"
+PER_CLASS_RECALL_DESCRIPTION = "[{category}] Recall: {value:.3f}"
+PER_CLASS_F1_DESCRIPTION = "[{category}] F1: {value:.3f}"
+PER_CLASS_SUPPORT_DESCRIPTION = "[{category}] Support: {value}"
+
+# Experiment configuration
+EXPERIMENT_NAME = "Copilot Response Classification Evaluation"
+EXPERIMENT_DESCRIPTION = (
+    "Evaluating Copilot response classification performance with per-class metrics "
+    "and overall averages (micro, macro, weighted). The metric that are reported are: "
+    "precision, recall, F1, support."
+)

rasa/builder/evaluator/response_classification/evaluator.py

@@ -0,0 +1,346 @@
+from typing import Dict, List, Literal, Optional
+
+import structlog
+
+from rasa.builder.copilot.models import ResponseCategory
+from rasa.builder.evaluator.response_classification.constants import (
+    MACRO_AVERAGING_METHOD,
+    MICRO_AVERAGING_METHOD,
+    WEIGHTED_AVERAGING_METHOD,
+)
+from rasa.builder.evaluator.response_classification.models import (
+    ClassificationResult,
+    MetricsSummary,
+    OverallClassificationMetrics,
+    PerClassMetrics,
+)
+
+structlogger = structlog.get_logger()
+
+
+class ResponseClassificationEvaluator:
+    def __init__(self):  # type: ignore[no-untyped-def]
+        self._classes: List[ResponseCategory] = [
+            ResponseCategory.COPILOT,
+            ResponseCategory.OUT_OF_SCOPE_DETECTION,
+            ResponseCategory.ROLEPLAY_DETECTION,
+            ResponseCategory.KNOWLEDGE_BASE_ACCESS_REQUESTED,
+            ResponseCategory.ERROR_FALLBACK,
+            # TODO: Add the greetings and goodbyes as support once the orchestrator
+            #       aproach is implemented
+        ]
+        self._true_positives_per_class: Dict[ResponseCategory, int] = {
+            clazz: 0 for clazz in self._classes
+        }
+        self._false_positives_per_class: Dict[ResponseCategory, int] = {
+            clazz: 0 for clazz in self._classes
+        }
+        self._false_negatives_per_class: Dict[ResponseCategory, int] = {
+            clazz: 0 for clazz in self._classes
+        }
+        self._support_per_class: Dict[ResponseCategory, int] = {
+            clazz: 0 for clazz in self._classes
+        }
+
+        self._evaluated = False
+
+    @property
+    def metrics_summary(self) -> Optional[MetricsSummary]:
+        """Get the metrics summary.
+
+        Returns:
+            MetricsSummary with structured per-class and overall metrics if
+            the evaluator has been run on the data, otherwise None.
+        """
+        if not self._evaluated:
+            structlogger.warning(
+                "evaluator.response_classification_evaluator"
+                ".metrics_summary.not_evaluated",
+                event_info="Evaluator not evaluated. Returning empty metrics summary.",
+            )
+            return None
+
+        return self._get_metrics_summary()
+
+    def reset(self) -> None:
+        self._true_positives_per_class = {clazz: 0 for clazz in self._classes}
+        self._false_positives_per_class = {clazz: 0 for clazz in self._classes}
+        self._false_negatives_per_class = {clazz: 0 for clazz in self._classes}
+        self._support_per_class = {clazz: 0 for clazz in self._classes}
+        self._evaluated = False
+
+    def evaluate(self, item_results: List[ClassificationResult]) -> MetricsSummary:
+        """Evaluate the classifier on the given item results."""
+        if self._evaluated:
+            structlogger.warning(
+                "evaluator.response_classification_evaluator.evaluate.already_evaluated",
+                event_info="Evaluator already evaluated. Resetting evaluator.",
+            )
+            self.reset()
+
+        for result in item_results:
+            # Skip and raise a warning if the class is not in the list of classes
+            if result.expected not in self._classes:
+                structlogger.warning(
+                    "evaluator.response_classification_evaluator"
+                    ".evaluate.class_not_recognized",
+                    event_info=(
+                        f"Class '{result.expected}' is not recognized. "
+                        f"Skipping evaluation for this class."
+                    ),
+                    expected_class=result.expected,
+                    classes=self._classes,
+                )
+                continue
+
+            # Update support for the expected class
+            if result.expected in self._support_per_class:
+                self._support_per_class[result.expected] += 1
+
+            # Calculate TP, FP, FN per class
+            for clazz in self._classes:
+                if result.prediction == clazz and result.expected == clazz:
+                    self._true_positives_per_class[clazz] += 1
+
+                elif result.prediction == clazz and result.expected != clazz:
+                    self._false_positives_per_class[clazz] += 1
+
+                elif result.prediction != clazz and result.expected == clazz:
+                    self._false_negatives_per_class[clazz] += 1
+
+        self._evaluated = True
+        return self._get_metrics_summary()
+
+    def calculate_precision_per_class(self, clazz: ResponseCategory) -> float:
+        """Calculate precision for a specific response category."""
+        tp = self._true_positives_per_class.get(clazz, 0)
+        fp = self._false_positives_per_class.get(clazz, 0)
+
+        if tp + fp == 0:
+            return 0.0
+
+        return tp / (tp + fp)
+
+    def calculate_recall_per_class(self, clazz: ResponseCategory) -> float:
+        """Calculate recall for a specific response category."""
+        tp = self._true_positives_per_class.get(clazz, 0)
+        fn = self._false_negatives_per_class.get(clazz, 0)
+
+        if tp + fn == 0:
+            return 0.0
+
+        return tp / (tp + fn)
+
+    def calculate_f1_per_class(self, clazz: ResponseCategory) -> float:
+        """Calculate F1 score for a specific response category."""
+        precision = self.calculate_precision_per_class(clazz)
+        recall = self.calculate_recall_per_class(clazz)
+
+        if precision + recall == 0:
+            return 0.0
+
+        return 2 * (precision * recall) / (precision + recall)
+
+    def calculate_precision(
+        self, average: Literal["micro", "macro", "weighted"] = MICRO_AVERAGING_METHOD
+    ) -> float:
+        """Calculate precision with specified averaging method."""
+        if average == MICRO_AVERAGING_METHOD:
+            return self._calculate_micro_precision()
+        elif average == MACRO_AVERAGING_METHOD:
+            return self._calculate_macro_precision()
+        elif average == WEIGHTED_AVERAGING_METHOD:
+            return self._calculate_weighted_avg_precision()
+        else:
+            raise ValueError(f"Invalid averaging method: {average}")
+
+    def _calculate_micro_precision(self) -> float:
+        """Calculate overall precision with specified averaging method.
+
+        Calculates the metric globally by aggregating the total true positives, false
+        positives, across all classes. Each sample contributes equally to the final
+        score.
+        """
+        total_tp = sum(self._true_positives_per_class.values())
+        total_fp = sum(self._false_positives_per_class.values())
+
+        if total_tp + total_fp == 0:
+            return 0.0
+
+        return total_tp / (total_tp + total_fp)
+
+    def _calculate_macro_precision(self) -> float:
+        """Calculate macro-averaged precision.
+
+        Calculates the metric independently for each class and then takes the
+        unweighted average. Each class contributes equally.
+        """
+        precisions = [
+            self.calculate_precision_per_class(clazz) for clazz in self._classes
+        ]
+        return sum(precisions) / len(precisions) if precisions else 0.0
+
+    def _calculate_weighted_avg_precision(self) -> float:
+        """Calculate weighted-averaged precision.
+
+        Calculates the metric independently for each class and then takes the average
+        weighted by the class support (number of true samples per class).
+        """
+        total_support = sum(self._support_per_class.values())
+        if total_support == 0:
+            return 0.0
+
+        weighted_sum = 0.0
+        for clazz in self._classes:
+            precision = self.calculate_precision_per_class(clazz)
+            support = self._support_per_class.get(clazz, 0)
+            weighted_sum += precision * support
+
+        return weighted_sum / total_support
+
+    def calculate_recall(
+        self, average: Literal["micro", "macro", "weighted"] = MICRO_AVERAGING_METHOD
+    ) -> float:
+        """Calculate recall with specified averaging method."""
+        if average == MICRO_AVERAGING_METHOD:
+            return self._calculate_micro_recall()
+        elif average == MACRO_AVERAGING_METHOD:
+            return self._calculate_macro_recall()
+        elif average == WEIGHTED_AVERAGING_METHOD:
+            return self._calculate_weighted_avg_recall()
+        else:
+            raise ValueError(f"Invalid averaging method: {average}")
+
+    def _calculate_micro_recall(self) -> float:
+        """Calculate micro-averaged recall.
+
+        Calculates the metric globally by aggregating the total true positives, false
+        negatives, across all classes. Each sample contributes equally to the final
+        score.
+        """
+        total_tp = sum(self._true_positives_per_class.values())
+        total_fn = sum(self._false_negatives_per_class.values())
+
+        if total_tp + total_fn == 0:
+            return 0.0
+
+        return total_tp / (total_tp + total_fn)
+
+    def _calculate_macro_recall(self) -> float:
+        """Calculate macro-averaged recall.
+
+        Calculates the metric independently for each class and then takes the
+        unweighted average. Each class contributes equally.
+        """
+        recalls = [self.calculate_recall_per_class(clazz) for clazz in self._classes]
+        return sum(recalls) / len(recalls) if recalls else 0.0
+
+    def _calculate_weighted_avg_recall(self) -> float:
+        """Calculate weighted-averaged recall.
+
+        Calculates the metric independently for each class and then takes the average
+        weighted by the class support (number of true samples per class).
+        """
+        total_support = sum(self._support_per_class.values())
+        if total_support == 0:
+            return 0.0
+
+        weighted_sum = 0.0
+        for clazz in self._classes:
+            recall = self.calculate_recall_per_class(clazz)
+            support = self._support_per_class.get(clazz, 0)
+            weighted_sum += recall * support
+
+        return weighted_sum / total_support
+
+    def calculate_f1(
+        self, average: Literal["micro", "macro", "weighted"] = MICRO_AVERAGING_METHOD
+    ) -> float:
+        """Calculate F1 score with specified averaging method."""
+        if average == MICRO_AVERAGING_METHOD:
+            return self._calculate_micro_f1()
+        elif average == MACRO_AVERAGING_METHOD:
+            return self._calculate_macro_f1()
+        elif average == WEIGHTED_AVERAGING_METHOD:
+            return self._calculate_weighted_avg_f1()
+        else:
+            raise ValueError(f"Invalid averaging method: {average}")
+
+    def _calculate_micro_f1(self) -> float:
+        """Calculate micro-averaged F1 score.
+
+        Calculates the metric globally by aggregating the total true positives, false
+        positives, and false negatives across all classes. Each sample contributes
+        equally to the final score.
+        """
+        micro_precision = self._calculate_micro_precision()
+        micro_recall = self._calculate_micro_recall()
+
+        if micro_precision + micro_recall == 0:
+            return 0.0
+
+        return 2 * (micro_precision * micro_recall) / (micro_precision + micro_recall)
+
+    def _calculate_macro_f1(self) -> float:
+        """Calculate macro-averaged F1 score.
+
+        Calculates the metric independently for each class and then takes the
+        unweighted average. Each class contributes equally.
+        """
+        f1_scores = [self.calculate_f1_per_class(clazz) for clazz in self._classes]
+        return sum(f1_scores) / len(f1_scores) if f1_scores else 0.0
+
+    def _calculate_weighted_avg_f1(self) -> float:
+        """Calculate weighted F1 score.
+
+        Calculates the metric independently for each class and then takes the average
+        weighted by the class support (number of true samples per class).
+        """
+        total_support = sum(self._support_per_class.values())
+        if total_support == 0:
+            return 0.0
+
+        weighted_sum = 0.0
+        for clazz in self._classes:
+            f1 = self.calculate_f1_per_class(clazz)
+            support = self._support_per_class.get(clazz, 0)
+            weighted_sum += f1 * support
+
+        return weighted_sum / total_support
+
+    def _get_metrics_summary(self) -> MetricsSummary:
+        """Get the metrics summary without Optional wrapper.
+
+        This method assumes the evaluator has been evaluated and will always
+        return a MetricsSummary.
+        """
+        # Build per-class metrics
+        per_class_metrics: Dict[ResponseCategory, PerClassMetrics] = {}
+        for clazz in self._classes:
+            per_class_metrics[clazz] = PerClassMetrics(
+                precision=self.calculate_precision_per_class(clazz),
+                recall=self.calculate_recall_per_class(clazz),
+                f1=self.calculate_f1_per_class(clazz),
+                support=self._support_per_class.get(clazz, 0),
+                true_positives=self._true_positives_per_class.get(clazz, 0),
+                false_positives=self._false_positives_per_class.get(clazz, 0),
+                false_negatives=self._false_negatives_per_class.get(clazz, 0),
+            )
+
+        # Build overall metrics
+        overall_metrics = OverallClassificationMetrics(
+            micro_precision=self.calculate_precision(MICRO_AVERAGING_METHOD),
+            macro_precision=self.calculate_precision(MACRO_AVERAGING_METHOD),
+            weighted_avg_precision=self.calculate_precision(WEIGHTED_AVERAGING_METHOD),
+            micro_recall=self.calculate_recall(MICRO_AVERAGING_METHOD),
+            macro_recall=self.calculate_recall(MACRO_AVERAGING_METHOD),
+            weighted_avg_recall=self.calculate_recall(WEIGHTED_AVERAGING_METHOD),
+            micro_f1=self.calculate_f1(MICRO_AVERAGING_METHOD),
+            macro_f1=self.calculate_f1(MACRO_AVERAGING_METHOD),
+            weighted_avg_f1=self.calculate_f1(WEIGHTED_AVERAGING_METHOD),
+            support=sum(self._support_per_class.values()),
+            true_positives=sum(self._true_positives_per_class.values()),
+            false_positives=sum(self._false_positives_per_class.values()),
+            false_negatives=sum(self._false_negatives_per_class.values()),
+        )
+        return MetricsSummary(per_class=per_class_metrics, overall=overall_metrics)