andre/crewAI
1
0
Fork 0
mirror of https://github.com/crewAIInc/crewAI.git synced 2026-01-10 00:28:31 +00:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Brandon/cre 509 hitl multiple rounds of followup (#1702)

Browse Source 
* v1 of HITL working

* Drop print statements

* HITL code more robust. Still needs to be refactored.

* refactor and more clear messages

* Fix type issue

* fix tests

* Fix test again

* Drop extra print
This commit is contained in:
Brandon Hancock (bhancock_ai)
2024-12-05 10:14:04 -05:00
committed by GitHub
parent 06d02c0f62
commit 03abf53ba9
6 changed files with 403 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/crewai/agents/agent_builder/base_agent_executor_mixin.py
View File

@@ -3,16 +3,15 @@ from typing import TYPE_CHECKING, Optional
from crewai.memory.entity.entity_memory_item import EntityMemoryItem
from crewai.memory.long_term.long_term_memory_item import LongTermMemoryItem
from crewai.utilities import I18N
from crewai.utilities.converter import ConverterError
from crewai.utilities.evaluators.task_evaluator import TaskEvaluator
from crewai.utilities import I18N
from crewai.utilities.printer import Printer
if TYPE_CHECKING:
from crewai.agents.agent_builder.base_agent import BaseAgent
from crewai.crew import Crew
from crewai.task import Task
from crewai.agents.agent_builder.base_agent import BaseAgent
class CrewAgentExecutorMixin:
@@ -100,14 +99,19 @@ class CrewAgentExecutorMixin:
print(f"Failed to add to long term memory: {e}")
pass
def _ask_human_input(self, final_answer: dict) -> str:
def _ask_human_input(self, final_answer: str) -> str:
"""Prompt human input for final decision making."""
self._printer.print(
content=f"\033[1m\033[95m ## Final Result:\033[00m \033[92m{final_answer}\033[00m"
)
self._printer.print(
content="\n\n=====\n## Please provide feedback on the Final Result and the Agent's actions:",
content=(
"\n\n=====\n"
"## Please provide feedback on the Final Result and the Agent's actions. "
"Respond with 'looks good' or a similar phrase when you're satisfied.\n"
"=====\n"
),
color="bold_yellow",
)
return input()

123
src/crewai/agents/crew_agent_executor.py
View File

@@ -16,7 +16,7 @@ from crewai.agents.tools_handler import ToolsHandler
from crewai.tools.base_tool import BaseTool
from crewai.tools.tool_usage import ToolUsage, ToolUsageErrorException
from crewai.utilities import I18N, Printer
from crewai.utilities.constants import TRAINING_DATA_FILE
from crewai.utilities.constants import MAX_LLM_RETRY, TRAINING_DATA_FILE
from crewai.utilities.exceptions.context_window_exceeding_exception import (
LLMContextLengthExceededException,
)
@@ -90,7 +90,6 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
if "system" in self.prompt:
system_prompt = self._format_prompt(self.prompt.get("system", ""), inputs)
user_prompt = self._format_prompt(self.prompt.get("user", ""), inputs)
self.messages.append(self._format_msg(system_prompt, role="system"))
self.messages.append(self._format_msg(user_prompt))
else:
@@ -103,17 +102,8 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
formatted_answer = self._invoke_loop()
if self.ask_for_human_input:
human_feedback = self._ask_human_input(formatted_answer.output)
if self.crew and self.crew._train:
self._handle_crew_training_output(formatted_answer, human_feedback)
formatted_answer = self._handle_human_feedback(formatted_answer)
# Making sure we only ask for it once, so disabling for the next thought loop
self.ask_for_human_input = False
self.messages.append(self._format_msg(f"Feedback: {human_feedback}"))
formatted_answer = self._invoke_loop()
if self.crew and self.crew._train:
self._handle_crew_training_output(formatted_answer)
self._create_short_term_memory(formatted_answer)
self._create_long_term_memory(formatted_answer)
return {"output": formatted_answer.output}
@@ -326,16 +316,14 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
def _handle_context_length(self) -> None:
if self.respect_context_window:
self._logger.log(
"debug",
"Context length exceeded. Summarizing content to fit the model context window.",
self._printer.print(
content="Context length exceeded. Summarizing content to fit the model context window.",
color="yellow",
)
self._summarize_messages()
else:
self._logger.log(
"debug",
"Context length exceeded. Consider using smaller text or RAG tools from crewai_tools.",
self._printer.print(
content="Context length exceeded. Consider using smaller text or RAG tools from crewai_tools.",
color="red",
)
raise SystemExit(
@@ -362,15 +350,13 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
] = result.output
training_handler.save(training_data)
else:
self._logger.log(
"error",
"Invalid train iteration type or agent_id not in training data.",
self._printer.print(
content="Invalid train iteration type or agent_id not in training data.",
color="red",
)
else:
self._logger.log(
"error",
"Crew is None or does not have _train_iteration attribute.",
self._printer.print(
content="Crew is None or does not have _train_iteration attribute.",
color="red",
)
@@ -388,15 +374,13 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
train_iteration, agent_id, training_data
)
else:
self._logger.log(
"error",
"Invalid train iteration type. Expected int.",
self._printer.print(
content="Invalid train iteration type. Expected int.",
color="red",
)
else:
self._logger.log(
"error",
"Crew is None or does not have _train_iteration attribute.",
self._printer.print(
content="Crew is None or does not have _train_iteration attribute.",
color="red",
)
@@ -412,3 +396,82 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
def _format_msg(self, prompt: str, role: str = "user") -> Dict[str, str]:
prompt = prompt.rstrip()
return {"role": role, "content": prompt}
def _handle_human_feedback(self, formatted_answer: AgentFinish) -> AgentFinish:
"""
Handles the human feedback loop, allowing the user to provide feedback
on the agent's output and determining if additional iterations are needed.
Parameters:
formatted_answer (AgentFinish): The initial output from the agent.
Returns:
AgentFinish: The final output after incorporating human feedback.
"""
while self.ask_for_human_input:
human_feedback = self._ask_human_input(formatted_answer.output)
print("Human feedback: ", human_feedback)
if self.crew and self.crew._train:
self._handle_crew_training_output(formatted_answer, human_feedback)
# Make an LLM call to verify if additional changes are requested based on human feedback
additional_changes_prompt = self._i18n.slice(
"human_feedback_classification"
).format(feedback=human_feedback)
retry_count = 0
llm_call_successful = False
additional_changes_response = None
while retry_count < MAX_LLM_RETRY and not llm_call_successful:
try:
additional_changes_response = (
self.llm.call(
[
self._format_msg(
additional_changes_prompt, role="system"
)
],
callbacks=self.callbacks,
)
.strip()
.lower()
)
llm_call_successful = True
except Exception as e:
retry_count += 1
self._printer.print(
content=f"Error during LLM call to classify human feedback: {e}. Retrying... ({retry_count}/{MAX_LLM_RETRY})",
color="red",
)
if not llm_call_successful:
self._printer.print(
content="Error processing feedback after multiple attempts.",
color="red",
)
self.ask_for_human_input = False
break
if additional_changes_response == "false":
self.ask_for_human_input = False
elif additional_changes_response == "true":
self.ask_for_human_input = True
# Add human feedback to messages
self.messages.append(self._format_msg(f"Feedback: {human_feedback}"))
# Invoke the loop again with updated messages
formatted_answer = self._invoke_loop()
if self.crew and self.crew._train:
self._handle_crew_training_output(formatted_answer)
else:
# Unexpected response
self._printer.print(
content=f"Unexpected response from LLM: '{additional_changes_response}'. Assuming no additional changes requested.",
color="red",
)
self.ask_for_human_input = False
return formatted_answer

3
src/crewai/translations/en.json
View File

@@ -22,7 +22,8 @@
"sumamrize_instruction": "Summarize the following text, make sure to include all the important information: {group}",
"summary": "This is a summary of our conversation so far:\n{merged_summary}",
"manager_request": "Your best answer to your coworker asking you this, accounting for the context shared.",
"formatted_task_instructions": "Ensure your final answer contains only the content in the following format: {output_format}\n\nEnsure the final output does not include any code block markers like ```json or ```python."
"formatted_task_instructions": "Ensure your final answer contains only the content in the following format: {output_format}\n\nEnsure the final output does not include any code block markers like ```json or ```python.",
"human_feedback_classification": "Determine if the following feedback indicates that the user is satisfied or if further changes are needed. Respond with 'True' if further changes are needed, or 'False' if the user is satisfied. **Important** Do not include any additional commentary outside of your 'True' or 'False' response.\n\nFeedback: \"{feedback}\""
},
"errors": {
"force_final_answer_error": "You can't keep going, this was the best you could do.\n {formatted_answer.text}",

1
src/crewai/utilities/constants.py
View File

@@ -2,3 +2,4 @@ TRAINING_DATA_FILE = "training_data.pkl"
TRAINED_AGENTS_DATA_FILE = "trained_agents_data.pkl"
DEFAULT_SCORE_THRESHOLD = 0.35
KNOWLEDGE_DIRECTORY = "knowledge"
MAX_LLM_RETRY = 3