azure-ai-evaluation 1.0.1__py3-none-any.whl → 1.13.3__py3-none-any.whl

azure/ai/evaluation/red_team/_utils/_rai_service_true_false_scorer.py

@@ -0,0 +1,114 @@
+# ---------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# ---------------------------------------------------------
+
+import logging
+from typing import List, Optional
+
+from pyrit.models import Score, PromptRequestPiece, UnvalidatedScore
+from pyrit.score.scorer import Scorer
+
+from azure.ai.evaluation.simulator._model_tools._generated_rai_client import GeneratedRAIClient
+from ._rai_service_eval_chat_target import RAIServiceEvalChatTarget
+from .._attack_objective_generator import RiskCategory
+
+
+class AzureRAIServiceTrueFalseScorer(Scorer):
+    """Scorer for Azure RAI service.
+
+    This class extends the PyRIT Scorer class to support scoring prompts using
+    the Azure RAI service. It sends prompts to the RAI service and processes
+    the responses to generate scores.
+
+    This is a true/false scorer that determines whether a given prompt response
+    achieves the specified objective.
+    """
+
+    scorer_type = "true_false"
+
+    def __init__(
+        self,
+        *,
+        client: GeneratedRAIClient,
+        risk_category: RiskCategory,
+        credential,
+        azure_ai_project,
+        api_version: Optional[str] = None,
+        model: Optional[str] = None,
+        prompt_template_key: Optional[str] = None,
+        logger: Optional[logging.Logger] = None,
+        context: Optional[str] = None,
+    ) -> None:
+        """Initialize the scorer.
+
+        :param client: The RAI client to use for scoring
+        :param api_version: Optional API version to use
+        :param model: Optional model name to score with
+        :param prompt_template_key: Optional prompt template key to use
+        :param logger: Optional logger to use
+        """
+        super().__init__()
+        self._client = client
+        self._api_version = api_version
+        self._model = model
+        self._prompt_template_key = prompt_template_key
+        self.logger = logger
+        self.context = context
+
+        # Use AzureRAIServiceTarget for scoring as well
+        self._scoring_target_red_llm = RAIServiceEvalChatTarget(
+            logger=self.logger,
+            credential=credential,
+            azure_ai_project=azure_ai_project,
+            risk_category=risk_category,
+            context=context,
+        )
+
+    async def score_async(
+        self,
+        request_response: PromptRequestPiece,
+        *,
+        task: Optional[str] = None,
+    ) -> List[Score]:
+        """Score a prompt response asynchronously.
+
+        :param request_response: The prompt response to score (PromptRequestPiece)
+        :param task: Optional task description (objective)
+        :return: List of Score objects
+        """
+        self.logger.info(f"Starting to score prompt response")
+
+        # Validate request_response
+        self.validate(request_response, task=task)
+
+        if not self._scoring_target_red_llm.context:
+            self._scoring_target_red_llm.context = request_response.labels.get("context", "")
+
+        unvalidated_score: UnvalidatedScore = await self._score_value_with_llm(
+            prompt_target=self._scoring_target_red_llm,
+            system_prompt="Please evaluate the following response.",  # TODO: What should this be?
+            prompt_request_value=request_response.converted_value,
+            prompt_request_data_type=request_response.converted_value_data_type,
+            scored_prompt_id=request_response.id,
+            task=task,
+            orchestrator_identifier=request_response.orchestrator_identifier,
+        )
+
+        score = unvalidated_score.to_score(score_value=unvalidated_score.raw_score_value)
+
+        # self._memory.add_scores_to_memory(scores=[score])
+        return [score]
+
+    def validate(self, request_response, *, task: Optional[str] = None):
+        """Validates the request_response piece to score.
+
+        This method checks if the request_response is valid for scoring by this scorer.
+
+        :param request_response: The request response to be validated
+        :param task: The task based on which the text should be scored (the original attacker model's objective)
+        :raises: ValueError if the request_response is invalid
+        """
+
+        # Additional validation can be added here as needed
+        # For now we'll keep it simple since we handle conversion to PromptRequestResponse in score_async
+        pass

azure/ai/evaluation/red_team/_utils/constants.py

@@ -0,0 +1,72 @@
+"""
+Constants used in Red Team Agent.
+"""
+
+import os
+from .._attack_strategy import AttackStrategy
+from .._attack_objective_generator import RiskCategory
+
+# File extensions
+BASELINE_IDENTIFIER = "baseline"
+DATA_EXT = ".jsonl"
+RESULTS_EXT = ".json"
+
+# Mapping of attack strategies to complexity levels
+
+ATTACK_STRATEGY_COMPLEXITY_MAP = {
+    str(AttackStrategy.Baseline.value): "baseline",
+    str(AttackStrategy.AnsiAttack.value): "easy",
+    str(AttackStrategy.AsciiArt.value): "easy",
+    str(AttackStrategy.AsciiSmuggler.value): "easy",
+    str(AttackStrategy.Atbash.value): "easy",
+    str(AttackStrategy.Base64.value): "easy",
+    str(AttackStrategy.Binary.value): "easy",
+    str(AttackStrategy.Caesar.value): "easy",
+    str(AttackStrategy.CharacterSpace.value): "easy",
+    str(AttackStrategy.CharSwap.value): "easy",
+    str(AttackStrategy.Diacritic.value): "easy",
+    str(AttackStrategy.Flip.value): "easy",
+    str(AttackStrategy.Leetspeak.value): "easy",
+    str(AttackStrategy.Morse.value): "easy",
+    str(AttackStrategy.ROT13.value): "easy",
+    str(AttackStrategy.SuffixAppend.value): "easy",
+    str(AttackStrategy.StringJoin.value): "easy",
+    str(AttackStrategy.UnicodeConfusable.value): "easy",
+    str(AttackStrategy.UnicodeSubstitution.value): "easy",
+    str(AttackStrategy.Url.value): "easy",
+    str(AttackStrategy.EASY.value): "easy",
+    str(AttackStrategy.Tense.value): "moderate",
+    str(AttackStrategy.MODERATE.value): "moderate",
+    str(AttackStrategy.DIFFICULT.value): "difficult",
+    str(AttackStrategy.Jailbreak.value): "easy",
+    str(AttackStrategy.IndirectJailbreak.value): "easy",
+    str(AttackStrategy.MultiTurn.value): "difficult",
+    str(AttackStrategy.Crescendo.value): "difficult",
+}
+
+# Task timeouts and status codes
+INTERNAL_TASK_TIMEOUT = 120
+
+# Sampling constants
+# Multiplier for the maximum number of sampling iterations when round-robin sampling from risk subtypes.
+# This prevents infinite loops while allowing sufficient attempts to find unique objectives.
+# With N subtypes, this allows up to N * MAX_SAMPLING_ITERATIONS_MULTIPLIER total iterations.
+MAX_SAMPLING_ITERATIONS_MULTIPLIER = 100
+
+# Map of risk categories to their maximum number of subtypes
+# Used to calculate num_objectives_with_subtypes for adequate subtype coverage
+RISK_TO_NUM_SUBTYPE_MAP = {
+    RiskCategory.ProhibitedActions: 32,
+    RiskCategory.TaskAdherence: 9,
+    RiskCategory.SensitiveDataLeakage: 19,
+}
+
+# Task status definitions
+TASK_STATUS = {
+    "PENDING": "pending",
+    "RUNNING": "running",
+    "COMPLETED": "completed",
+    "FAILED": "failed",
+    "TIMEOUT": "timeout",
+    "INCOMPLETE": "incomplete",
+}

azure/ai/evaluation/red_team/_utils/exception_utils.py

@@ -0,0 +1,345 @@
+# ---------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# ---------------------------------------------------------
+"""
+Exception handling utilities for Red Team Agent.
+
+This module provides centralized exception handling, error categorization,
+and error reporting utilities for red team operations.
+"""
+
+import logging
+import traceback
+import asyncio
+from typing import Optional, Any, Dict, Union
+from enum import Enum
+
+
+class ErrorCategory(Enum):
+    """Categories of errors that can occur during red team operations."""
+
+    NETWORK = "network"
+    AUTHENTICATION = "authentication"
+    CONFIGURATION = "configuration"
+    DATA_PROCESSING = "data_processing"
+    ORCHESTRATOR = "orchestrator"
+    EVALUATION = "evaluation"
+    FILE_IO = "file_io"
+    TIMEOUT = "timeout"
+    UNKNOWN = "unknown"
+
+
+class ErrorSeverity(Enum):
+    """Severity levels for errors."""
+
+    LOW = "low"  # Warning level, operation can continue
+    MEDIUM = "medium"  # Error level, task failed but scan can continue
+    HIGH = "high"  # Critical error, scan should be aborted
+    FATAL = "fatal"  # Unrecoverable error
+
+
+class RedTeamError(Exception):
+    """Base exception for Red Team operations."""
+
+    def __init__(
+        self,
+        message: str,
+        category: ErrorCategory = ErrorCategory.UNKNOWN,
+        severity: ErrorSeverity = ErrorSeverity.MEDIUM,
+        context: Optional[Dict[str, Any]] = None,
+        original_exception: Optional[Exception] = None,
+    ):
+        super().__init__(message)
+        self.message = message
+        self.category = category
+        self.severity = severity
+        self.context = context or {}
+        self.original_exception = original_exception
+
+
+class ExceptionHandler:
+    """Centralized exception handling for Red Team operations."""
+
+    def __init__(self, logger: Optional[logging.Logger] = None):
+        """Initialize exception handler.
+
+        :param logger: Logger instance for error reporting
+        """
+        self.logger = logger or logging.getLogger(__name__)
+        self.error_counts: Dict[ErrorCategory, int] = {category: 0 for category in ErrorCategory}
+
+    def categorize_exception(self, exception: Exception) -> ErrorCategory:
+        """Categorize an exception based on its type and message.
+
+        :param exception: The exception to categorize
+        :return: The appropriate error category
+        """
+        import httpx
+        import httpcore
+
+        # Network-related errors
+        network_exceptions = (
+            httpx.ConnectTimeout,
+            httpx.ReadTimeout,
+            httpx.ConnectError,
+            httpx.HTTPError,
+            httpx.TimeoutException,
+            httpcore.ReadTimeout,
+            ConnectionError,
+            ConnectionRefusedError,
+            ConnectionResetError,
+        )
+
+        if isinstance(exception, network_exceptions):
+            return ErrorCategory.NETWORK
+
+        # Timeout errors (separate from network to handle asyncio.TimeoutError)
+        if isinstance(exception, (TimeoutError, asyncio.TimeoutError)):
+            return ErrorCategory.TIMEOUT
+
+        # File I/O errors
+        if isinstance(exception, (IOError, OSError, FileNotFoundError, PermissionError)):
+            return ErrorCategory.FILE_IO
+
+        # HTTP status code specific errors
+        if hasattr(exception, "response") and hasattr(exception.response, "status_code"):
+            status_code = exception.response.status_code
+            if 500 <= status_code < 600:
+                return ErrorCategory.NETWORK
+            elif status_code == 401:
+                return ErrorCategory.AUTHENTICATION
+            elif status_code == 403:
+                return ErrorCategory.CONFIGURATION
+
+        # String-based categorization
+        message = str(exception).lower()
+
+        # Define keyword mappings for cleaner logic
+        keyword_mappings = {
+            ErrorCategory.AUTHENTICATION: ["authentication", "unauthorized"],
+            ErrorCategory.CONFIGURATION: ["configuration", "config"],
+            ErrorCategory.ORCHESTRATOR: ["orchestrator"],
+            ErrorCategory.EVALUATION: ["evaluation", "evaluate", "model_error"],
+            ErrorCategory.DATA_PROCESSING: ["data", "json"],
+        }
+
+        for category, keywords in keyword_mappings.items():
+            if any(keyword in message for keyword in keywords):
+                return category
+
+        return ErrorCategory.UNKNOWN
+
+    def determine_severity(
+        self, exception: Exception, category: ErrorCategory, context: Optional[Dict[str, Any]] = None
+    ) -> ErrorSeverity:
+        """Determine the severity of an exception.
+
+        :param exception: The exception to evaluate
+        :param category: The error category
+        :param context: Additional context for severity determination
+        :return: The appropriate error severity
+        """
+        context = context or {}
+
+        # Critical system errors
+        if isinstance(exception, (MemoryError, SystemExit, KeyboardInterrupt)):
+            return ErrorSeverity.FATAL
+
+        # Authentication and configuration are typically high severity
+        if category in (ErrorCategory.AUTHENTICATION, ErrorCategory.CONFIGURATION):
+            return ErrorSeverity.HIGH
+
+        # File I/O errors can be high severity if they involve critical files
+        if category == ErrorCategory.FILE_IO:
+            if context.get("critical_file", False):
+                return ErrorSeverity.HIGH
+            return ErrorSeverity.MEDIUM
+
+        # Network and timeout errors are usually medium severity (retryable)
+        if category in (ErrorCategory.NETWORK, ErrorCategory.TIMEOUT):
+            return ErrorSeverity.MEDIUM
+
+        # Task-specific errors are medium severity
+        if category in (ErrorCategory.ORCHESTRATOR, ErrorCategory.EVALUATION, ErrorCategory.DATA_PROCESSING):
+            return ErrorSeverity.MEDIUM
+
+        return ErrorSeverity.LOW
+
+    def handle_exception(
+        self,
+        exception: Exception,
+        context: Optional[Dict[str, Any]] = None,
+        task_name: Optional[str] = None,
+        reraise: bool = False,
+    ) -> RedTeamError:
+        """Handle an exception with proper categorization and logging.
+
+        :param exception: The exception to handle
+        :param context: Additional context information
+        :param task_name: Name of the task where the exception occurred
+        :param reraise: Whether to reraise the exception after handling
+        :return: A RedTeamError with categorized information
+        """
+        context = context or {}
+
+        # If it's already a RedTeamError, just log and return/reraise
+        if isinstance(exception, RedTeamError):
+            self._log_error(exception, task_name)
+            if reraise:
+                raise exception
+            return exception
+
+        # Categorize the exception
+        category = self.categorize_exception(exception)
+        severity = self.determine_severity(exception, category, context)
+
+        # Update error counts
+        self.error_counts[category] += 1
+
+        # Create RedTeamError
+        message = f"{category.value.title()} error"
+        if task_name:
+            message += f" in {task_name}"
+        message += f": {str(exception)}"
+
+        red_team_error = RedTeamError(
+            message=message, category=category, severity=severity, context=context, original_exception=exception
+        )
+
+        # Log the error
+        self._log_error(red_team_error, task_name)
+
+        if reraise:
+            raise red_team_error
+
+        return red_team_error
+
+    def _log_error(self, error: RedTeamError, task_name: Optional[str] = None) -> None:
+        """Log an error with appropriate level based on severity.
+
+        :param error: The RedTeamError to log
+        :param task_name: Optional task name for context
+        """
+        # Determine log level based on severity
+        if error.severity == ErrorSeverity.FATAL:
+            log_level = logging.CRITICAL
+        elif error.severity == ErrorSeverity.HIGH:
+            log_level = logging.ERROR
+        elif error.severity == ErrorSeverity.MEDIUM:
+            log_level = logging.WARNING
+        else:
+            log_level = logging.INFO
+
+        # Create log message
+        message_parts = []
+        if task_name:
+            message_parts.append(f"[{task_name}]")
+        message_parts.append(f"[{error.category.value}]")
+        message_parts.append(f"[{error.severity.value}]")
+        message_parts.append(error.message)
+
+        log_message = " ".join(message_parts)
+
+        # Log with appropriate level
+        self.logger.log(log_level, log_message)
+
+        # Log additional context if available
+        if error.context:
+            self.logger.debug(f"Error context: {error.context}")
+
+        # Log original exception traceback for debugging
+        if error.original_exception and self.logger.isEnabledFor(logging.DEBUG):
+            self.logger.debug(f"Original exception traceback:\n{traceback.format_exc()}")
+
+    def should_abort_scan(self) -> bool:
+        """Determine if the scan should be aborted based on error patterns.
+
+        :return: True if the scan should be aborted
+        """
+        # Abort if we have too many high-severity errors
+        high_severity_categories = [ErrorCategory.AUTHENTICATION, ErrorCategory.CONFIGURATION]
+        high_severity_count = sum(self.error_counts[cat] for cat in high_severity_categories)
+
+        if high_severity_count > 2:
+            return True
+
+        # Abort if we have too many network errors (indicates systemic issue)
+        if self.error_counts[ErrorCategory.NETWORK] > 10:
+            return True
+
+        return False
+
+    def get_error_summary(self) -> Dict[str, Any]:
+        """Get a summary of all errors encountered.
+
+        :return: Dictionary containing error statistics
+        """
+        total_errors = sum(self.error_counts.values())
+
+        return {
+            "total_errors": total_errors,
+            "error_counts_by_category": dict(self.error_counts),
+            "most_common_category": max(self.error_counts, key=self.error_counts.get) if total_errors > 0 else None,
+            "should_abort": self.should_abort_scan(),
+        }
+
+    def log_error_summary(self) -> None:
+        """Log a summary of all errors encountered."""
+        summary = self.get_error_summary()
+
+        if summary["total_errors"] == 0:
+            self.logger.info("No errors encountered during operation")
+            return
+
+        self.logger.info(f"Error Summary: {summary['total_errors']} total errors")
+
+        for category, count in summary["error_counts_by_category"].items():
+            if count > 0:
+                self.logger.info(f"  {category}: {count}")
+
+        if summary["most_common_category"]:
+            self.logger.info(f"Most common error type: {summary['most_common_category']}")
+
+
+def create_exception_handler(logger: Optional[logging.Logger] = None) -> ExceptionHandler:
+    """Create an ExceptionHandler instance.
+
+    :param logger: Logger instance for error reporting
+    :return: Configured ExceptionHandler
+    """
+    return ExceptionHandler(logger=logger)
+
+
+# Convenience context manager for handling exceptions
+class exception_context:
+    """Context manager for handling exceptions in Red Team operations."""
+
+    def __init__(
+        self,
+        handler: ExceptionHandler,
+        task_name: str,
+        context: Optional[Dict[str, Any]] = None,
+        reraise_fatal: bool = True,
+    ):
+        self.handler = handler
+        self.task_name = task_name
+        self.context = context or {}
+        self.reraise_fatal = reraise_fatal
+        self.error: Optional[RedTeamError] = None
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if exc_val is not None:
+            self.error = self.handler.handle_exception(
+                exception=exc_val, context=self.context, task_name=self.task_name, reraise=False
+            )
+
+            # Reraise fatal errors unless specifically disabled
+            if self.reraise_fatal and self.error.severity == ErrorSeverity.FATAL:
+                raise self.error
+
+            # Suppress the original exception (we've handled it)
+            return True
+        return False