CoACT initialize (#292)
This commit is contained in:
Linxin Song
@@ -0,0 +1,7 @@
|
||||
# Copyright (c) 2023 - 2025, AG2ai, Inc., AG2ai open-source projects maintainers and core contributors
|
||||
#
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
from .deep_research import DeepResearchTool
|
||||
|
||||
__all__ = ["DeepResearchTool"]
|
||||
@@ -0,0 +1,328 @@
|
||||
# Copyright (c) 2023 - 2025, AG2ai, Inc., AG2ai open-source projects maintainers and core contributors
|
||||
#
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
import copy
|
||||
from typing import Annotated, Any, Callable, Union
|
||||
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
from ....agentchat import ConversableAgent
|
||||
from ....doc_utils import export_module
|
||||
from ....llm_config import LLMConfig
|
||||
from ... import Depends, Tool
|
||||
from ...dependency_injection import on
|
||||
|
||||
__all__ = ["DeepResearchTool"]
|
||||
|
||||
|
||||
class Subquestion(BaseModel):
|
||||
question: Annotated[str, Field(description="The original question.")]
|
||||
|
||||
def format(self) -> str:
|
||||
return f"Question: {self.question}\n"
|
||||
|
||||
|
||||
class SubquestionAnswer(Subquestion):
|
||||
answer: Annotated[str, Field(description="The answer to the question.")]
|
||||
|
||||
def format(self) -> str:
|
||||
return f"Question: {self.question}\n{self.answer}\n"
|
||||
|
||||
|
||||
class Task(BaseModel):
|
||||
question: Annotated[str, Field(description="The original question.")]
|
||||
subquestions: Annotated[list[Subquestion], Field(description="The subquestions that need to be answered.")]
|
||||
|
||||
def format(self) -> str:
|
||||
return f"Task: {self.question}\n\n" + "\n".join(
|
||||
"Subquestion " + str(i + 1) + ":\n" + subquestion.format()
|
||||
for i, subquestion in enumerate(self.subquestions)
|
||||
)
|
||||
|
||||
|
||||
class CompletedTask(BaseModel):
|
||||
question: Annotated[str, Field(description="The original question.")]
|
||||
subquestions: Annotated[list[SubquestionAnswer], Field(description="The subquestions and their answers")]
|
||||
|
||||
def format(self) -> str:
|
||||
return f"Task: {self.question}\n\n" + "\n".join(
|
||||
"Subquestion " + str(i + 1) + ":\n" + subquestion.format()
|
||||
for i, subquestion in enumerate(self.subquestions)
|
||||
)
|
||||
|
||||
|
||||
class InformationCrumb(BaseModel):
|
||||
source_url: str
|
||||
source_title: str
|
||||
source_summary: str
|
||||
relevant_info: str
|
||||
|
||||
|
||||
class GatheredInformation(BaseModel):
|
||||
information: list[InformationCrumb]
|
||||
|
||||
def format(self) -> str:
|
||||
return "Here is the gathered information: \n" + "\n".join(
|
||||
f"URL: {info.source_url}\nTitle: {info.source_title}\nSummary: {info.source_summary}\nRelevant Information: {info.relevant_info}\n\n"
|
||||
for info in self.information
|
||||
)
|
||||
|
||||
|
||||
@export_module("autogen.tools.experimental")
|
||||
class DeepResearchTool(Tool):
|
||||
"""A tool that delegates a web research task to the subteams of agents."""
|
||||
|
||||
ANSWER_CONFIRMED_PREFIX = "Answer confirmed:"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
llm_config: Union[LLMConfig, dict[str, Any]],
|
||||
max_web_steps: int = 30,
|
||||
):
|
||||
"""Initialize the DeepResearchTool.
|
||||
|
||||
Args:
|
||||
llm_config (LLMConfig, dict[str, Any]): The LLM configuration.
|
||||
max_web_steps (int, optional): The maximum number of web steps. Defaults to 30.
|
||||
"""
|
||||
self.llm_config = llm_config
|
||||
|
||||
self.summarizer_agent = ConversableAgent(
|
||||
name="SummarizerAgent",
|
||||
system_message=(
|
||||
"You are an agent with a task of answering the question provided by the user."
|
||||
"First you need to split the question into subquestions by calling the 'split_question_and_answer_subquestions' method."
|
||||
"Then you need to sintesize the answers the original question by combining the answers to the subquestions."
|
||||
),
|
||||
is_termination_msg=lambda x: x.get("content", "")
|
||||
and x.get("content", "").startswith(self.ANSWER_CONFIRMED_PREFIX),
|
||||
llm_config=llm_config,
|
||||
human_input_mode="NEVER",
|
||||
)
|
||||
|
||||
self.critic_agent = ConversableAgent(
|
||||
name="CriticAgent",
|
||||
system_message=(
|
||||
"You are a critic agent responsible for evaluating the answer provided by the summarizer agent.\n"
|
||||
"Your task is to assess the quality of the answer based on its coherence, relevance, and completeness.\n"
|
||||
"Provide constructive feedback on how the answer can be improved.\n"
|
||||
"If the answer is satisfactory, call the 'confirm_answer' method to end the task.\n"
|
||||
),
|
||||
is_termination_msg=lambda x: x.get("content", "")
|
||||
and x.get("content", "").startswith(self.ANSWER_CONFIRMED_PREFIX),
|
||||
llm_config=llm_config,
|
||||
human_input_mode="NEVER",
|
||||
)
|
||||
|
||||
def delegate_research_task(
|
||||
task: Annotated[str, "The task to perform a research on."],
|
||||
llm_config: Annotated[Union[LLMConfig, dict[str, Any]], Depends(on(llm_config))],
|
||||
max_web_steps: Annotated[int, Depends(on(max_web_steps))],
|
||||
) -> str:
|
||||
"""Delegate a research task to the agent.
|
||||
|
||||
Args:
|
||||
task (str): The task to perform a research on.
|
||||
llm_config (LLMConfig, dict[str, Any]): The LLM configuration.
|
||||
max_web_steps (int): The maximum number of web steps.
|
||||
|
||||
Returns:
|
||||
str: The answer to the research task.
|
||||
"""
|
||||
|
||||
@self.summarizer_agent.register_for_execution()
|
||||
@self.critic_agent.register_for_llm(description="Call this method to confirm the final answer.")
|
||||
def confirm_summary(answer: str, reasoning: str) -> str:
|
||||
return f"{self.ANSWER_CONFIRMED_PREFIX}" + answer + "\nReasoning: " + reasoning
|
||||
|
||||
split_question_and_answer_subquestions = DeepResearchTool._get_split_question_and_answer_subquestions(
|
||||
llm_config=llm_config,
|
||||
max_web_steps=max_web_steps,
|
||||
)
|
||||
|
||||
self.summarizer_agent.register_for_llm(description="Split the question into subquestions and get answers.")(
|
||||
split_question_and_answer_subquestions
|
||||
)
|
||||
self.critic_agent.register_for_execution()(split_question_and_answer_subquestions)
|
||||
|
||||
result = self.critic_agent.initiate_chat(
|
||||
self.summarizer_agent,
|
||||
message="Please answer the following question: " + task,
|
||||
# This outer chat should preserve the history of the conversation
|
||||
clear_history=False,
|
||||
)
|
||||
|
||||
return result.summary
|
||||
|
||||
super().__init__(
|
||||
name=delegate_research_task.__name__,
|
||||
description="Delegate a research task to the deep research agent.",
|
||||
func_or_tool=delegate_research_task,
|
||||
)
|
||||
|
||||
SUBQUESTIONS_ANSWER_PREFIX = "Subquestions answered:"
|
||||
|
||||
@staticmethod
|
||||
def _get_split_question_and_answer_subquestions(
|
||||
llm_config: Union[LLMConfig, dict[str, Any]], max_web_steps: int
|
||||
) -> Callable[..., Any]:
|
||||
def split_question_and_answer_subquestions(
|
||||
question: Annotated[str, "The question to split and answer."],
|
||||
llm_config: Annotated[Union[LLMConfig, dict[str, Any]], Depends(on(llm_config))],
|
||||
max_web_steps: Annotated[int, Depends(on(max_web_steps))],
|
||||
) -> str:
|
||||
decomposition_agent = ConversableAgent(
|
||||
name="DecompositionAgent",
|
||||
system_message=(
|
||||
"You are an expert at breaking down complex questions into smaller, focused subquestions.\n"
|
||||
"Your task is to take any question provided and divide it into clear, actionable subquestions that can be individually answered.\n"
|
||||
"Ensure the subquestions are logical, non-redundant, and cover all key aspects of the original question.\n"
|
||||
"Avoid providing answers or interpretations—focus solely on decomposition.\n"
|
||||
"Do not include banal, general knowledge questions\n"
|
||||
"Do not include questions that go into unnecessary detail that is not relevant to the original question\n"
|
||||
"Do not include question that require knowledge of the original or other subquestions to answer\n"
|
||||
"Some rule of thumb is to have only one subquestion for easy questions, 3 for medium questions, and 5 for hard questions.\n"
|
||||
),
|
||||
llm_config=llm_config,
|
||||
is_termination_msg=lambda x: x.get("content", "")
|
||||
and x.get("content", "").startswith(DeepResearchTool.SUBQUESTIONS_ANSWER_PREFIX),
|
||||
human_input_mode="NEVER",
|
||||
)
|
||||
|
||||
example_task = Task(
|
||||
question="What is the capital of France?",
|
||||
subquestions=[Subquestion(question="What is the capital of France?")],
|
||||
)
|
||||
decomposition_critic = ConversableAgent(
|
||||
name="DecompositionCritic",
|
||||
system_message=(
|
||||
"You are a critic agent responsible for evaluating the subquestions provided by the initial analysis agent.\n"
|
||||
"You need to confirm whether the subquestions are clear, actionable, and cover all key aspects of the original question.\n"
|
||||
"Do not accept redundant or unnecessary subquestions, focus solely on the minimal viable subset of subqestions necessary to answer the original question. \n"
|
||||
"Do not accept banal, general knowledge questions\n"
|
||||
"Do not accept questions that go into unnecessary detail that is not relevant to the original question\n"
|
||||
"Remove questions that can be answered with combining knowledge from other questions\n"
|
||||
"After you are satisfied with the subquestions, call the 'generate_subquestions' method to answer each subquestion.\n"
|
||||
"This is an example of an argument that can be passed to the 'generate_subquestions' method:\n"
|
||||
f"{{'task': {example_task.model_dump()}}}\n"
|
||||
"Some rule of thumb is to have only one subquestion for easy questions, 3 for medium questions, and 5 for hard questions.\n"
|
||||
),
|
||||
llm_config=llm_config,
|
||||
is_termination_msg=lambda x: x.get("content", "")
|
||||
and x.get("content", "").startswith(DeepResearchTool.SUBQUESTIONS_ANSWER_PREFIX),
|
||||
human_input_mode="NEVER",
|
||||
)
|
||||
|
||||
generate_subquestions = DeepResearchTool._get_generate_subquestions(
|
||||
llm_config=llm_config, max_web_steps=max_web_steps
|
||||
)
|
||||
decomposition_agent.register_for_execution()(generate_subquestions)
|
||||
decomposition_critic.register_for_llm(description="Generate subquestions for a task.")(
|
||||
generate_subquestions
|
||||
)
|
||||
|
||||
result = decomposition_critic.initiate_chat(
|
||||
decomposition_agent,
|
||||
message="Analyse and gather subqestions for the following question: " + question,
|
||||
)
|
||||
|
||||
return result.summary
|
||||
|
||||
return split_question_and_answer_subquestions
|
||||
|
||||
@staticmethod
|
||||
def _get_generate_subquestions(
|
||||
llm_config: Union[LLMConfig, dict[str, Any]],
|
||||
max_web_steps: int,
|
||||
) -> Callable[..., str]:
|
||||
"""Get the generate_subquestions method.
|
||||
|
||||
Args:
|
||||
llm_config (Union[LLMConfig, dict[str, Any]]): The LLM configuration.
|
||||
max_web_steps (int): The maximum number of web steps.
|
||||
|
||||
Returns:
|
||||
Callable[..., str]: The generate_subquestions method.
|
||||
"""
|
||||
|
||||
def generate_subquestions(
|
||||
task: Task,
|
||||
llm_config: Annotated[Union[LLMConfig, dict[str, Any]], Depends(on(llm_config))],
|
||||
max_web_steps: Annotated[int, Depends(on(max_web_steps))],
|
||||
) -> str:
|
||||
if not task.subquestions:
|
||||
task.subquestions = [Subquestion(question=task.question)]
|
||||
|
||||
subquestions_answers: list[SubquestionAnswer] = []
|
||||
for subquestion in task.subquestions:
|
||||
answer = DeepResearchTool._answer_question(
|
||||
subquestion.question, llm_config=llm_config, max_web_steps=max_web_steps
|
||||
)
|
||||
subquestions_answers.append(SubquestionAnswer(question=subquestion.question, answer=answer))
|
||||
|
||||
completed_task = CompletedTask(question=task.question, subquestions=subquestions_answers)
|
||||
|
||||
return f"{DeepResearchTool.SUBQUESTIONS_ANSWER_PREFIX} \n" + completed_task.format()
|
||||
|
||||
return generate_subquestions
|
||||
|
||||
@staticmethod
|
||||
def _answer_question(
|
||||
question: str,
|
||||
llm_config: Union[LLMConfig, dict[str, Any]],
|
||||
max_web_steps: int,
|
||||
) -> str:
|
||||
from ....agents.experimental.websurfer import WebSurferAgent
|
||||
|
||||
websurfer_config = copy.deepcopy(llm_config)
|
||||
|
||||
websurfer_config["config_list"][0]["response_format"] = GatheredInformation
|
||||
|
||||
def is_termination_msg(x: dict[str, Any]) -> bool:
|
||||
content = x.get("content", "")
|
||||
return (content is not None) and content.startswith(DeepResearchTool.ANSWER_CONFIRMED_PREFIX)
|
||||
|
||||
websurfer_agent = WebSurferAgent(
|
||||
llm_config=llm_config,
|
||||
web_tool_llm_config=websurfer_config,
|
||||
name="WebSurferAgent",
|
||||
system_message=(
|
||||
"You are a web surfer agent responsible for gathering information from the web to provide information for answering a question\n"
|
||||
"You will be asked to find information related to the question and provide a summary of the information gathered.\n"
|
||||
"The summary should include the URL, title, summary, and relevant information for each piece of information gathered.\n"
|
||||
),
|
||||
is_termination_msg=is_termination_msg,
|
||||
human_input_mode="NEVER",
|
||||
web_tool_kwargs={
|
||||
"agent_kwargs": {"max_steps": max_web_steps},
|
||||
},
|
||||
)
|
||||
|
||||
websurfer_critic = ConversableAgent(
|
||||
name="WebSurferCritic",
|
||||
system_message=(
|
||||
"You are a critic agent responsible for evaluating the answer provided by the web surfer agent.\n"
|
||||
"You need to confirm whether the information provided by the websurfer is correct and sufficient to answer the question.\n"
|
||||
"You can ask the web surfer to provide more information or provide and confirm the answer.\n"
|
||||
),
|
||||
llm_config=llm_config,
|
||||
is_termination_msg=is_termination_msg,
|
||||
human_input_mode="NEVER",
|
||||
)
|
||||
|
||||
@websurfer_agent.register_for_execution()
|
||||
@websurfer_critic.register_for_llm(
|
||||
description="Call this method when you agree that the original question can be answered with the gathered information and provide the answer."
|
||||
)
|
||||
def confirm_answer(answer: str) -> str:
|
||||
return f"{DeepResearchTool.ANSWER_CONFIRMED_PREFIX} " + answer
|
||||
|
||||
websurfer_critic.register_for_execution()(websurfer_agent.tool)
|
||||
|
||||
result = websurfer_critic.initiate_chat(
|
||||
websurfer_agent,
|
||||
message="Please find the answer to this question: " + question,
|
||||
)
|
||||
|
||||
return result.summary
|
||||
Reference in New Issue
Block a user