# Copyright (c) Microsoft. All rights reserved. import asyncio import inspect import json import logging import uuid from abc import ABC, abstractmethod from collections.abc import Awaitable, Callable from typing import Generic, Union, get_args from pydantic import Field from typing_extensions import TypeVar from semantic_kernel.agents.agent import Agent from semantic_kernel.agents.runtime.core.cancellation_token import CancellationToken from semantic_kernel.agents.runtime.core.core_runtime import CoreRuntime from semantic_kernel.contents.chat_message_content import ChatMessageContent from semantic_kernel.contents.streaming_chat_message_content import StreamingChatMessageContent from semantic_kernel.contents.utils.author_role import AuthorRole from semantic_kernel.kernel_pydantic import KernelBaseModel from semantic_kernel.utils.feature_stage_decorator import experimental logger: logging.Logger = logging.getLogger(__name__) DefaultTypeAlias = Union[ChatMessageContent, list[ChatMessageContent]] TIn = TypeVar("TIn", default=DefaultTypeAlias) TOut = TypeVar("TOut", default=DefaultTypeAlias) @experimental class OrchestrationResult(KernelBaseModel, Generic[TOut]): """The result of an invocation of an orchestration.""" background_task: asyncio.Task | None = None value: TOut | None = None exception: BaseException | None = None event: asyncio.Event = Field(default_factory=asyncio.Event) cancellation_token: CancellationToken = Field(default_factory=CancellationToken) async def get(self, timeout: float | None = None) -> TOut: """Get the result of the invocation. If a timeout is specified, the method will wait for the result for the specified time. If the result is not available within the timeout, a TimeoutError will be raised but the invocation will not be aborted. Args: timeout (int | None): The timeout (seconds) for getting the result. If None, wait indefinitely. Returns: TOut: The result of the invocation. """ if timeout is not None: await asyncio.wait_for(self.event.wait(), timeout=timeout) else: await self.event.wait() if self.value is None: if self.cancellation_token.is_cancelled(): raise RuntimeError("The invocation was canceled before it could complete.") if self.exception is not None: raise self.exception raise RuntimeError("The invocation did not produce a result.") return self.value def cancel(self) -> None: """Cancel the invocation. This method will cancel the invocation. Actors that have received messages will continue to process them, but no new messages will be processed. """ if self.cancellation_token.is_cancelled(): raise RuntimeError("The invocation has already been canceled.") if self.event.is_set(): raise RuntimeError("The invocation has already been completed.") self.cancellation_token.cancel() self.event.set() @experimental class OrchestrationBase(ABC, Generic[TIn, TOut]): """Base class for multi-agent orchestration.""" t_in: type[TIn] | None = None t_out: type[TOut] | None = None def __init__( self, members: list[Agent], name: str | None = None, description: str | None = None, input_transform: Callable[[TIn], Awaitable[DefaultTypeAlias] | DefaultTypeAlias] | None = None, output_transform: Callable[[DefaultTypeAlias], Awaitable[TOut] | TOut] | None = None, agent_response_callback: Callable[[DefaultTypeAlias], Awaitable[None] | None] | None = None, streaming_agent_response_callback: Callable[[StreamingChatMessageContent, bool], Awaitable[None] | None] | None = None, ) -> None: """Initialize the orchestration base. Args: members (list[Agent]): The list of agents to be used. name (str | None): A unique name of the orchestration. If None, a unique name will be generated. description (str | None): The description of the orchestration. If None, use a default description. input_transform (Callable | None): A function that transforms the external input message. output_transform (Callable | None): A function that transforms the internal output message. agent_response_callback (Callable | None): A function that is called when a full response is produced by the agents. streaming_agent_response_callback (Callable | None): A function that is called when a streaming response is produced by the agents. """ if not members: raise ValueError("The members list cannot be empty.") self._members = members self.name = name or f"{self.__class__.__name__}_{uuid.uuid4().hex}" self.description = description or "A multi-agent orchestration." self._input_transform = input_transform or self._default_input_transform self._output_transform = output_transform or self._default_output_transform self._agent_response_callback = agent_response_callback self._streaming_agent_response_callback = streaming_agent_response_callback def _set_types(self) -> None: """Set the external input and output types from the class arguments. This method can only be run after the class has been initialized because it relies on the `__orig_class__` attributes to determine the type parameters. This method will first try to get the type parameters from the class itself. The `__orig_class__` attribute will contain the external input and output types if they are explicitly given, for example: ``` class MyOrchestration(OrchestrationBase[TIn, TOut]): pass my_orchestration = MyOrchestration[str, str](...) ``` If the type parameters are not explicitly given, for example when the TypeVars has defaults, for example: ``` TIn = TypeVar("TIn", default=str) TOut = TypeVar("TOut", default=str) class MyOrchestration(OrchestrationBase[TIn, TOut]): pass my_orchestration = MyOrchestration(...) ``` The type parameters can be inferred from the `__orig_bases__` attribute. """ if all([self.t_in is not None, self.t_out is not None]): return try: args = self.__orig_class__.__args__ # type: ignore[attr-defined] if len(args) == 1: self.t_in = args[0] self.t_out = DefaultTypeAlias # type: ignore[assignment] elif len(args) == 2: self.t_in = args[0] self.t_out = args[1] else: raise TypeError("Orchestration must have two type parameters.") except AttributeError: args = get_args(self.__orig_bases__[0]) # type: ignore[attr-defined] if len(args) != 2: raise TypeError("Orchestration must be subclassed with two type parameters.") self.t_in = args[0] if isinstance(args[0], type) else getattr(args[0], "__default__", None) # type: ignore[assignment] self.t_out = args[1] if isinstance(args[1], type) else getattr(args[1], "__default__", None) # type: ignore[assignment] if any([self.t_in is None, self.t_out is None]): raise TypeError("Orchestration must have concrete types for all type parameters.") async def invoke( self, task: str | DefaultTypeAlias | TIn, runtime: CoreRuntime, ) -> OrchestrationResult[TOut]: """Invoke the multi-agent orchestration. This method is non-blocking and will return immediately. To wait for the result, use the `get` method of the `OrchestrationResult` object. Args: task (str, DefaultTypeAlias, TIn): The task to be executed by the agents. runtime (CoreRuntime): The runtime environment for the agents. """ self._set_types() orchestration_result = OrchestrationResult[self.t_out]() # type: ignore[name-defined] async def result_callback(result: DefaultTypeAlias) -> None: nonlocal orchestration_result if inspect.iscoroutinefunction(self._output_transform): transformed_result = await self._output_transform(result) else: transformed_result = self._output_transform(result) orchestration_result.value = transformed_result orchestration_result.event.set() def inner_exception_callback(exception: BaseException) -> None: nonlocal orchestration_result orchestration_result.exception = exception orchestration_result.event.set() # This unique topic type is used to isolate the orchestration run from others. internal_topic_type = uuid.uuid4().hex await self._prepare( runtime, internal_topic_type=internal_topic_type, result_callback=result_callback, exception_callback=inner_exception_callback, ) if isinstance(task, str): prepared_task = ChatMessageContent(role=AuthorRole.USER, content=task) elif isinstance(task, ChatMessageContent) or ( isinstance(task, list) and all(isinstance(item, ChatMessageContent) for item in task) ): prepared_task = task # type: ignore[assignment] else: if inspect.iscoroutinefunction(self._input_transform): prepared_task = await self._input_transform(task) # type: ignore[arg-type] else: prepared_task = self._input_transform(task) # type: ignore[arg-type,assignment] background_task = asyncio.create_task( self._start( prepared_task, runtime, internal_topic_type, orchestration_result.cancellation_token, ) ) # Add a callback to surface any exceptions that occur during outside of the runtime. def outer_exception_callback(task: asyncio.Task) -> None: nonlocal orchestration_result try: task.result() except BaseException as e: orchestration_result.exception = e orchestration_result.event.set() background_task.add_done_callback(outer_exception_callback) orchestration_result.background_task = background_task return orchestration_result @abstractmethod async def _start( self, task: DefaultTypeAlias, runtime: CoreRuntime, internal_topic_type: str, cancellation_token: CancellationToken, ) -> None: """Start the multi-agent orchestration. Args: task (ChatMessageContent | list[ChatMessageContent]): The task to be executed by the agents. runtime (CoreRuntime): The runtime environment for the agents. internal_topic_type (str): The internal topic type for the orchestration that this actor is part of. cancellation_token (CancellationToken): The cancellation token for the orchestration. """ pass @abstractmethod async def _prepare( self, runtime: CoreRuntime, internal_topic_type: str, exception_callback: Callable[[BaseException], None], result_callback: Callable[[DefaultTypeAlias], Awaitable[None]], ) -> None: """Register the actors and orchestrations with the runtime and add the required subscriptions. Args: runtime (CoreRuntime): The runtime environment for the agents. internal_topic_type (str): The internal topic type for the orchestration that this actor is part of. exception_callback (Callable): A function that is called when an exception occurs. result_callback (Callable): A function that is called when the result is available. """ pass def _default_input_transform(self, input_message: TIn) -> DefaultTypeAlias: """Default input transform function. This function transforms the external input message to chat message content(s). If the input message is already in the correct format, it is returned as is. Args: input_message (TIn): The input message to be transformed. Returns: DefaultTypeAlias: The transformed input message. """ if isinstance(input_message, ChatMessageContent): return input_message if isinstance(input_message, list) and all(isinstance(item, ChatMessageContent) for item in input_message): return input_message if isinstance(input_message, self.t_in): # type: ignore[arg-type] return ChatMessageContent( role=AuthorRole.USER, content=json.dumps(input_message.__dict__), ) raise TypeError(f"Invalid input message type: {type(input_message)}. Expected {self.t_in}.") def _default_output_transform(self, output_message: DefaultTypeAlias) -> TOut: """Default output transform function. This function transforms the internal output message to the external output message. If the output message is already in the correct format, it is returned as is. Args: output_message (DefaultTypeAlias): The output message to be transformed. Returns: TOut: The transformed output message. """ if self.t_out == DefaultTypeAlias or self.t_out in get_args(DefaultTypeAlias): if isinstance(output_message, ChatMessageContent) or ( isinstance(output_message, list) and all(isinstance(item, ChatMessageContent) for item in output_message) ): return output_message # type: ignore[return-value] raise TypeError(f"Invalid output message type: {type(output_message)}. Expected {self.t_out}.") if isinstance(output_message, ChatMessageContent): return self.t_out(**json.loads(output_message.content)) # type: ignore[misc] raise TypeError(f"Unable to transform output message of type {type(output_message)} to {self.t_out}.")