92 lines
3.5 KiB
Python
92 lines
3.5 KiB
Python
# Copyright (c) Microsoft. All rights reserved.
|
|
|
|
from typing import TYPE_CHECKING, ClassVar
|
|
|
|
from opentelemetry import trace
|
|
from pydantic import Field
|
|
|
|
from semantic_kernel.agents.strategies import TerminationStrategy
|
|
from semantic_kernel.connectors.ai.chat_completion_client_base import ChatCompletionClientBase
|
|
from semantic_kernel.connectors.ai.open_ai import (
|
|
AzureChatPromptExecutionSettings,
|
|
OpenAIChatCompletion,
|
|
)
|
|
from semantic_kernel.contents import ChatHistory
|
|
|
|
if TYPE_CHECKING:
|
|
from semantic_kernel.agents.agent import Agent
|
|
from semantic_kernel.contents.chat_message_content import ChatMessageContent
|
|
|
|
|
|
TERMINATE_TRUE_KEYWORD = "yes"
|
|
TERMINATE_FALSE_KEYWORD = "no"
|
|
|
|
NEWLINE = "\n"
|
|
|
|
|
|
class CustomTerminationStrategy(TerminationStrategy):
|
|
NUM_OF_RETRIES: ClassVar[int] = 3
|
|
|
|
maximum_iterations: int = 20
|
|
chat_completion_service: ChatCompletionClientBase = Field(default_factory=lambda: OpenAIChatCompletion())
|
|
|
|
async def should_agent_terminate(self, agent: "Agent", history: list["ChatMessageContent"]) -> bool:
|
|
"""Check if the agent should terminate.
|
|
|
|
Args:
|
|
agent: The agent to check.
|
|
history: The history of messages in the conversation.
|
|
"""
|
|
tracer = trace.get_tracer(__name__)
|
|
with tracer.start_as_current_span("terminate_strategy"):
|
|
chat_history = ChatHistory(system_message=self.get_system_message().strip())
|
|
|
|
for message in history:
|
|
content = message.content
|
|
# We don't want to add messages whose text content is empty.
|
|
# Those messages are likely messages from function calls and function results.
|
|
if content:
|
|
chat_history.add_message(message)
|
|
|
|
chat_history.add_user_message(
|
|
"Is the latest content approved by all agents? "
|
|
f"Answer with '{TERMINATE_TRUE_KEYWORD}' or '{TERMINATE_FALSE_KEYWORD}'."
|
|
)
|
|
|
|
for _ in range(self.NUM_OF_RETRIES):
|
|
completion = await self.chat_completion_service.get_chat_message_content(
|
|
chat_history,
|
|
AzureChatPromptExecutionSettings(),
|
|
)
|
|
|
|
if not completion:
|
|
continue
|
|
|
|
if TERMINATE_FALSE_KEYWORD in completion.content.lower():
|
|
return False
|
|
if TERMINATE_TRUE_KEYWORD in completion.content.lower():
|
|
return True
|
|
|
|
chat_history.add_message(completion)
|
|
chat_history.add_user_message(
|
|
f"You must only say either '{TERMINATE_TRUE_KEYWORD}' or '{TERMINATE_FALSE_KEYWORD}'."
|
|
)
|
|
|
|
raise ValueError(
|
|
"Failed to determine if the agent should terminate because the model did not return a valid response."
|
|
)
|
|
|
|
def get_system_message(self) -> str:
|
|
return f"""
|
|
You are in a chat with multiple agents collaborating to create a document.
|
|
Each message in the chat history contains the agent's name and the message content.
|
|
|
|
The chat history may start empty as no agents have spoken yet.
|
|
|
|
Here are the agents with their indices, names, and descriptions:
|
|
{NEWLINE.join(f"[{index}] {agent.name}:{NEWLINE}{agent.description}" for index, agent in enumerate(self.agents))}
|
|
|
|
Your task is NOT to continue the conversation. Determine if the latest content is approved by all agents.
|
|
If approved, say "{TERMINATE_TRUE_KEYWORD}". Otherwise, say "{TERMINATE_FALSE_KEYWORD}".
|
|
"""
|