FEAT add TargetConfiguration & pieces (microsoft#1573)

Author: hannahwestra25

pyrit/message_normalizer/__init__.py

@@ -8,6 +8,7 @@
 from pyrit.message_normalizer.chat_message_normalizer import ChatMessageNormalizer
 from pyrit.message_normalizer.conversation_context_normalizer import ConversationContextNormalizer
 from pyrit.message_normalizer.generic_system_squash import GenericSystemSquashNormalizer
+from pyrit.message_normalizer.history_squash_normalizer import HistorySquashNormalizer
 from pyrit.message_normalizer.message_normalizer import (
     MessageListNormalizer,
     MessageStringNormalizer,
@@ -18,6 +19,7 @@
     "MessageListNormalizer",
     "MessageStringNormalizer",
     "GenericSystemSquashNormalizer",
+    "HistorySquashNormalizer",
     "TokenizerTemplateNormalizer",
     "ConversationContextNormalizer",
     "ChatMessageNormalizer",

pyrit/message_normalizer/history_squash_normalizer.py

@@ -0,0 +1,63 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+from pyrit.message_normalizer.message_normalizer import MessageListNormalizer
+from pyrit.models import Message
+
+
+class HistorySquashNormalizer(MessageListNormalizer[Message]):
+    """
+    Squashes a multi-turn conversation into a single user message.
+
+    Previous turns are formatted as labeled context and prepended to the
+    latest message.  Used by the normalization pipeline to adapt prompts
+    for targets that do not support multi-turn conversations.
+    """
+
+    async def normalize_async(self, messages: list[Message]) -> list[Message]:
+        """
+        Combine all messages into a single user message.
+
+        When there is only one message it is returned unchanged.  Otherwise
+        all prior turns are formatted as ``Role: content`` lines under a
+        ``[Conversation History]`` header and the last message's content
+        appears under a ``[Current Message]`` header.
+
+        Args:
+            messages: The conversation messages to squash.
+
+        Returns:
+            list[Message]: A single-element list containing the squashed message.
+
+        Raises:
+            ValueError: If the messages list is empty.
+        """
+        if not messages:
+            raise ValueError("Messages list cannot be empty")
+
+        if len(messages) == 1:
+            return list(messages)
+
+        history_lines = self._format_history(messages=messages[:-1])
+        current_parts = [piece.converted_value for piece in messages[-1].message_pieces]
+
+        combined = (
+            "[Conversation History]\n" + "\n".join(history_lines) + "\n\n[Current Message]\n" + "\n".join(current_parts)
+        )
+
+        return [Message.from_prompt(prompt=combined, role="user")]
+
+    def _format_history(self, *, messages: list[Message]) -> list[str]:
+        """
+        Format prior messages as ``Role: content`` lines.
+
+        Args:
+            messages: The history messages to format.
+
+        Returns:
+            list[str]: One line per message piece.
+        """
+        lines: list[str] = []
+        for msg in messages:
+            lines.extend(f"{piece.api_role.capitalize()}: {piece.converted_value}" for piece in msg.message_pieces)
+        return lines

pyrit/prompt_target/__init__.py

@@ -10,9 +10,16 @@
 
 from pyrit.prompt_target.azure_blob_storage_target import AzureBlobStorageTarget
 from pyrit.prompt_target.azure_ml_chat_target import AzureMLChatTarget
+from pyrit.prompt_target.common.conversation_normalization_pipeline import ConversationNormalizationPipeline
 from pyrit.prompt_target.common.prompt_chat_target import PromptChatTarget
 from pyrit.prompt_target.common.prompt_target import PromptTarget
-from pyrit.prompt_target.common.target_capabilities import TargetCapabilities
+from pyrit.prompt_target.common.target_capabilities import (
+    CapabilityHandlingPolicy,
+    CapabilityName,
+    TargetCapabilities,
+    UnsupportedCapabilityBehavior,
+)
+from pyrit.prompt_target.common.target_configuration import TargetConfiguration
 from pyrit.prompt_target.common.utils import limit_requests_per_minute
 from pyrit.prompt_target.gandalf_target import GandalfLevel, GandalfTarget
 from pyrit.prompt_target.http_target.http_target import HTTPTarget
@@ -41,7 +48,10 @@
 __all__ = [
     "AzureBlobStorageTarget",
     "AzureMLChatTarget",
+    "CapabilityName",
+    "CapabilityHandlingPolicy",
     "CopilotType",
+    "ConversationNormalizationPipeline",
     "GandalfLevel",
     "GandalfTarget",
     "get_http_target_json_response_callback_function",
@@ -66,6 +76,8 @@
     "PromptTarget",
     "RealtimeTarget",
     "TargetCapabilities",
+    "TargetConfiguration",
+    "UnsupportedCapabilityBehavior",
     "TextTarget",
     "WebSocketCopilotTarget",
 ]

pyrit/prompt_target/common/conversation_normalization_pipeline.py

@@ -0,0 +1,134 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import logging
+
+from pyrit.message_normalizer import (
+    GenericSystemSquashNormalizer,
+    HistorySquashNormalizer,
+    MessageListNormalizer,
+)
+from pyrit.models import Message
+from pyrit.prompt_target.common.target_capabilities import (
+    CapabilityHandlingPolicy,
+    CapabilityName,
+    TargetCapabilities,
+    UnsupportedCapabilityBehavior,
+)
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Single registry: add new normalizable capabilities here and nowhere else.
+# Order in the list determines pipeline execution order.
+# ---------------------------------------------------------------------------
+_NORMALIZER_REGISTRY: list[tuple[CapabilityName, MessageListNormalizer[Message]]] = [
+    (CapabilityName.SYSTEM_PROMPT, GenericSystemSquashNormalizer()),
+    (CapabilityName.MULTI_TURN, HistorySquashNormalizer()),
+]
+
+# Derived constant — no manual maintenance required.
+NORMALIZABLE_CAPABILITIES: frozenset[CapabilityName] = frozenset(cap for cap, _ in _NORMALIZER_REGISTRY)
+
+
+class ConversationNormalizationPipeline:
+    """
+    Ordered sequence of message normalizers that adapt conversations when
+    the target lacks certain capabilities.
+
+    The pipeline is constructed via ``from_capabilities``, which resolves
+    capabilities and policy into a concrete, ordered tuple of normalizers.
+    ``normalize_async`` then simply executes that tuple in order.
+
+    To add a new normalizable capability, add a single entry to
+    ``_NORMALIZER_REGISTRY``.  ``NORMALIZABLE_CAPABILITIES``,
+    pipeline ordering, and default normalizers are all derived from it.
+    """
+
+    def __init__(self, normalizers: tuple[MessageListNormalizer[Message], ...] = ()) -> None:
+        """
+        Initialize the normalization pipeline with an ordered sequence of normalizers.
+
+        Args:
+            normalizers (tuple[MessageListNormalizer[Message], ...]):
+                Ordered normalizers to apply during ``normalize_async``.
+                Defaults to an empty tuple (pass-through).
+        """
+        self._normalizers = normalizers
+
+    @classmethod
+    def from_capabilities(
+        cls,
+        *,
+        capabilities: TargetCapabilities,
+        policy: CapabilityHandlingPolicy,
+        normalizer_overrides: dict[CapabilityName, MessageListNormalizer[Message]] | None = None,
+    ) -> "ConversationNormalizationPipeline":
+        """
+        Resolve capabilities and policy into a concrete pipeline of normalizers.
+
+        For each capability in ``_NORMALIZER_REGISTRY`` (in order):
+
+        * If the target already supports the capability, no normalizer is added.
+        * If the capability is missing and the policy is ``ADAPT``, the
+          corresponding normalizer (from overrides or defaults) is added.
+        * If the capability is missing and the policy is ``RAISE``, a
+          ``ValueError`` is raised immediately.
+
+        Args:
+            capabilities (TargetCapabilities): The target's declared capabilities.
+            policy (CapabilityHandlingPolicy): How to handle each missing capability.
+            normalizer_overrides (dict[CapabilityName, MessageListNormalizer[Message]] | None):
+                Optional overrides for specific capability normalizers.
+                Falls back to the defaults from ``_NORMALIZER_REGISTRY``.
+
+        Returns:
+            ConversationNormalizationPipeline: A pipeline with the resolved
+            ordered tuple of normalizers.
+
+        Raises:
+            ValueError: If a required capability is missing and the policy is RAISE.
+        """
+        overrides = normalizer_overrides or {}
+        normalizers: list[MessageListNormalizer[Message]] = []
+
+        for capability, default_normalizer in _NORMALIZER_REGISTRY:
+            if capabilities.includes(capability=capability):
+                continue
+
+            behavior = policy.get_behavior(capability=capability)
+
+            if behavior == UnsupportedCapabilityBehavior.RAISE:
+                raise ValueError(f"Target does not support '{capability.value}' and the handling policy is RAISE.")
+
+            normalizer = overrides.get(capability, default_normalizer)
+
+            normalizers.append(normalizer)
+
+        return cls(normalizers=tuple(normalizers))
+
+    async def normalize_async(self, *, messages: list[Message]) -> list[Message]:
+        """
+        Run the pre-resolved normalizer sequence over the messages.
+
+        Args:
+            messages (list[Message]): The full conversation to normalize.
+
+        Returns:
+            list[Message]: The (possibly adapted) message list.
+        """
+        result = list(messages)
+        for normalizer in self._normalizers:
+            result = await normalizer.normalize_async(result)
+        return result
+
+    @property
+    def normalizers(self) -> tuple[MessageListNormalizer[Message], ...]:
+        """
+        The ordered normalizers in this pipeline.
+
+        Returns:
+            tuple[MessageListNormalizer[Message], ...]: The normalizer sequence.
+        """
+        return self._normalizers

pyrit/prompt_target/common/target_capabilities.py

@@ -1,12 +1,109 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
-from dataclasses import dataclass
-from typing import Optional, cast
+from collections.abc import Mapping
+from dataclasses import dataclass, field
+from enum import Enum
+from types import MappingProxyType
+from typing import NoReturn, Optional, cast
 
 from pyrit.models import PromptDataType
 
 
+class CapabilityName(str, Enum):
+    """
+    Canonical identifiers for target capabilities.
+
+    This keeps capability identity in one place so policy, requirements, and
+    normalization code do not duplicate string field names.
+    """
+
+    MULTI_TURN = "supports_multi_turn"
+    MULTI_MESSAGE_PIECES = "supports_multi_message_pieces"
+    JSON_SCHEMA = "supports_json_schema"
+    JSON_OUTPUT = "supports_json_output"
+    EDITABLE_HISTORY = "supports_editable_history"
+    SYSTEM_PROMPT = "supports_system_prompt"
+
+
+class UnsupportedCapabilityBehavior(str, Enum):
+    """
+    Defines what happens when a caller requires a capability the target does not support.
+
+    ADAPT: apply a normalization step to work around the unsupported capability.
+    RAISE: fail immediately with an error.
+    """
+
+    ADAPT = "adapt"
+    RAISE = "raise"
+
+
+@dataclass(frozen=True)
+class CapabilityHandlingPolicy:
+    """
+    Per-capability policy consulted only when a capability is unsupported.
+
+    Design invariants
+    -----------------
+    * The policy is never consulted if the capability is already supported.
+    * Non-adaptable capabilities (e.g. ``supports_editable_history``) are not
+      represented here; requesting them on a target that lacks them always
+      raises immediately.
+    """
+
+    behaviors: Mapping[CapabilityName, UnsupportedCapabilityBehavior] = field(
+        default_factory=lambda: {
+            CapabilityName.MULTI_TURN: UnsupportedCapabilityBehavior.RAISE,
+            CapabilityName.SYSTEM_PROMPT: UnsupportedCapabilityBehavior.RAISE,
+        }
+    )
+
+    def get_behavior(self, *, capability: CapabilityName) -> UnsupportedCapabilityBehavior:
+        """
+        Return the configured handling behavior for a capability.
+
+        Args:
+            capability: The capability to look up.
+
+        Returns:
+            UnsupportedCapabilityBehavior: The configured behavior.
+
+        Raises:
+            KeyError: If no behavior exists for the capability. This occurs for
+            non-adaptable capabilities (e.g., supports_editable_history).
+        """
+        try:
+            return self.behaviors[capability]
+        except KeyError:
+            supported = ", ".join(sorted(cap.value for cap in self.behaviors))
+            raise KeyError(
+                f"No policy for capability '{capability.value}'. Supported capabilities: {supported}."
+            ) from None
+
+    def __getattr__(self, name: str) -> NoReturn:
+        """
+        Guard against accessing policies for non-adaptable or unknown capabilities.
+
+        Raises:
+            AttributeError: If the capability is not part of this policy.
+        """
+        for capability in CapabilityName:
+            if capability.value == name:
+                supported_names = ", ".join(sorted(cap.value for cap in self.behaviors))
+                raise AttributeError(
+                    f"'{type(self).__name__}' has no policy for '{name}'. "
+                    f"Only the following capabilities have handling policies: "
+                    f"{supported_names}."
+                )
+
+        raise AttributeError(name)
+
+    def __post_init__(self) -> None:
+        """Create a defensive read-only copy of the behaviors mapping."""
+        # object.__setattr__ is required because the dataclass is frozen.
+        object.__setattr__(self, "behaviors", MappingProxyType(dict(self.behaviors)))
+
+
 @dataclass(frozen=True)
 class TargetCapabilities:
     """
@@ -47,6 +144,18 @@ class attribute. Users can override individual capabilities per instance
     # The output modalities supported by the target (e.g., "text", "image").
     output_modalities: frozenset[frozenset[PromptDataType]] = frozenset({frozenset(["text"])})
 
+    def includes(self, *, capability: CapabilityName) -> bool:
+        """
+        Return whether this target supports the given capability.
+
+        Args:
+            capability: The capability to check.
+
+        Returns:
+            bool: True if supported, otherwise False.
+        """
+        return bool(getattr(self, capability.value))
+
     @staticmethod
     def get_known_capabilities(underlying_model: str) -> "Optional[TargetCapabilities]":
         """