Files
crewAI/crewai/agent.py
2023-11-05 16:21:22 -03:00

102 lines
3.0 KiB
Python

"""Generic agent."""
from typing import List
from pydantic import BaseModel, Field
from langchain.tools import Tool
from langchain.agents import AgentExecutor
from langchain.chat_models import ChatOpenAI as OpenAI
from langchain.tools.render import render_text_description
from langchain.agents.format_scratchpad import format_log_to_str
from langchain.agents.output_parsers import ReActSingleInputOutputParser, PydanticOutputParser
from .prompts import Prompts
from .agent.agent_vote import AgentVote
class Agent(BaseModel):
"""Generic agent implementation."""
role: str = Field(description="Role of the agent")
goal: str = Field(description="Objective of the agent")
backstory: str = Field(description="Backstory of the agent")
tools: List[Tool] = Field(
description="Tools at agents disposal",
default=[]
)
prompts: Prompts = Field(
description="Prompts class for the agent.",
default=Prompts
)
llm: str = Field(
description="LLM of the agent",
default=OpenAI(
temperature=0.7,
model="gpt-4",
verbose=True
)
)
def vote_agent_for_task(self, task: str) -> AgentVote:
"""
Execute a task with the agent.
Parameters:
task (str): Task to execute
Returns:
output (AgentVote): The agent voted to execute the task
"""
parser = PydanticOutputParser(pydantic_object=AgentVote)
prompt = Prompts.AGENT_EXECUTION_PROMPT.partial(
tools=render_text_description(self.tools),
tool_names=self.__tools_names(),
backstory=self.backstory,
role=self.role,
goal=self.goal,
format_instructions=parser.get_format_instructions()
)
return self.__function_calling(task, prompt, parser)
def execute_task(self, task: str) -> str:
"""
Execute a task with the agent.
Parameters:
task (str): Task to execute
Returns:
output (str): Output of the agent
"""
prompt = Prompts.AGENT_EXECUTION_PROMPT.partial(
tools=render_text_description(self.tools),
tool_names=self.__tools_names(),
backstory=self.backstory,
role=self.role,
goal=self.goal,
)
return self.__execute_task(task, prompt)
def __function_calling(self, input: str, prompt: str, parser: str) -> str:
inner_agent = {
"input": lambda x: x["input"],
"agent_scratchpad": lambda x: format_log_to_str(x['intermediate_steps'])
} | prompt | parser
return self.__execute(inner_agent, input)
def __execute_task(self, input: str, prompt: str) -> str:
chat_with_bind = self.llm.bind(stop=["\nObservation"])
inner_agent = {
"input": lambda x: x["input"],
"agent_scratchpad": lambda x: format_log_to_str(x['intermediate_steps'])
} | prompt | chat_with_bind | ReActSingleInputOutputParser()
return self.__execute(inner_agent, input)
def __execute(self, inner_agent, input):
agent_executor = AgentExecutor(
agent=inner_agent,
tools=self.tools,
verbose=True,
handle_parsing_errors=True
)
return agent_executor.invoke({"input": input})['output']
def __tools_names(self) -> str:
return ", ".join([t.name for t in self.tools])