Source code for src.openCHA.orchestrator.orchestrator

from __future__ import annotations

import logging
from typing import Any
from typing import Dict
from typing import List
from typing import Optional

from openCHA.CustomDebugFormatter import CustomDebugFormatter
from openCHA.datapipes import DataPipe
from openCHA.datapipes import DatapipeType
from openCHA.datapipes import initialize_datapipe
from openCHA.llms import LLMType
from openCHA.orchestrator import Action
from openCHA.planners import BasePlanner
from openCHA.planners import initialize_planner
from openCHA.planners import PlanFinish
from openCHA.planners import PlannerType
from openCHA.response_generators import (
    BaseResponseGenerator,
)
from openCHA.response_generators import (
    initialize_response_generator,
)
from openCHA.response_generators import (
    ResponseGeneratorType,
)
from openCHA.tasks import BaseTask
from openCHA.tasks import initialize_task
from openCHA.tasks import TaskType
from pydantic import BaseModel


[docs] class Orchestrator(BaseModel): """ **Description:** The Orchestrator class is the main execution heart of the CHA. All the components of the Orchestrator are initialized and executed here. The Orchestrator will start a new answering cycle by calling the `run` method. From there, the planning is started, then tasks will be executed one by one till the **Task Planner** decides that no more information is needed. Finally the **Task Planner** final answer will be routed to the **Final Response Generator** to generate an empathic final response that is returned to the user. """ planner: BasePlanner = None datapipe: DataPipe = None promptist: Any = None response_generator: BaseResponseGenerator = None available_tasks: Dict[str, BaseTask] = {} max_retries: int = 5 max_task_execute_retries: int = 3 max_planner_execute_retries: int = 16 max_final_answer_execute_retries: int = 3 role: int = 0 verbose: bool = False planner_logger: Optional[logging.Logger] = None tasks_logger: Optional[logging.Logger] = None orchestrator_logger: Optional[logging.Logger] = None final_answer_generator_logger: Optional[logging.Logger] = None promptist_logger: Optional[logging.Logger] = None error_logger: Optional[logging.Logger] = None previous_actions: List[str] = [] current_actions: List[str] = [] runtime: Dict[str, bool] = {} class Config: """Configuration for this pydantic object.""" arbitrary_types_allowed = True def print_log(self, log_name: str, message: str): if self.verbose: if log_name == "planner": self.planner_logger.debug(message) if log_name == "task": self.tasks_logger.debug(message) if log_name == "orchestrator": self.orchestrator_logger.debug(message) if log_name == "response_generator": self.final_answer_generator_logger.debug(message) if log_name == "promptist": self.promptist_logger.debug(message) if log_name == "error": self.error_logger.debug(message)
[docs] @classmethod def initialize( self, planner_llm: str = LLMType.OPENAI, planner_name: str = PlannerType.ZERO_SHOT_REACT_PLANNER, datapipe_name: str = DatapipeType.MEMORY, promptist_name: str = "", response_generator_llm: str = LLMType.OPENAI, response_generator_name: str = ResponseGeneratorType.BASE_GENERATOR, available_tasks: Optional[List[str]] = None, previous_actions: List[Action] = None, verbose: bool = False, **kwargs, ) -> Orchestrator: """ This class method initializes the Orchestrator by setting up the planner, datapipe, promptist, response generator, and available tasks. Args: planner_llm (str): LLMType to be used as LLM for planner. planner_name (str): PlannerType to be used as task planner. datapipe_name (str): DatapipeType to be used as data pipe. promptist_name (str): Not implemented yet! response_generator_llm (str): LLMType to be used as LLM for response generator. response_generator_name (str): ResponseGeneratorType to be used as response generator. available_tasks (List[str]): List of available task using TaskType. previous_actions (List[Action]): List of previous actions. verbose (bool): Specifies if the debugging logs be printed or not. **kwargs (Any): Additional keyword arguments. Return: Orchestrator: Initialized Orchestrator instance. Example: .. code-block:: python from openCHA.datapipes import DatapipeType from openCHA.planners import PlannerType from openCHA.response_generators import ResponseGeneratorType from openCHA.tasks import TaskType from openCHA.llms import LLMType from openCHA.orchestrator import Orchestrator orchestrator = Orchestrator.initialize( planner_llm=LLMType.OPENAI, planner_name=PlannerType.ZERO_SHOT_REACT_PLANNER, datapipe_name=DatapipeType.MEMORY, promptist_name="", response_generator_llm=LLMType.OPENAI, response_generator_name=ResponseGeneratorType.BASE_GENERATOR, available_tasks=[TaskType.SERPAPI, TaskType.EXTRACT_TEXT], verbose=self.verbose, **kwargs ) """ if available_tasks is None: available_tasks = [] if previous_actions is None: previous_actions = [] planner_logger = ( tasks_logger ) = ( orchestrator_logger ) = ( final_answer_generator_logger ) = promptist_logger = error_logger = None if verbose: planner_logger = CustomDebugFormatter.create_logger( "Planner", "cyan" ) tasks_logger = CustomDebugFormatter.create_logger( "Task", "purple" ) orchestrator_logger = CustomDebugFormatter.create_logger( "Orchestrator", "green" ) final_answer_generator_logger = ( CustomDebugFormatter.create_logger( "Response Generator", "blue" ) ) promptist_logger = CustomDebugFormatter.create_logger( "Promptist", "blue" ) error_logger = CustomDebugFormatter.create_logger( "Error", "red" ) datapipe = initialize_datapipe( datapipe=datapipe_name, **kwargs ) if verbose: orchestrator_logger.debug( f"Datapipe {datapipe_name} is successfully initialized.\n" ) tasks = {} for task in available_tasks: kwargs["datapipe"] = datapipe tasks[task] = initialize_task(task=task, **kwargs) if verbose: orchestrator_logger.debug( f"Task '{task}' is successfully initialized." ) planner = initialize_planner( tasks=list(tasks.values()), llm=planner_llm, planner=planner_name, **kwargs, ) if verbose: orchestrator_logger.debug( f"Planner {planner_name} is successfully initialized." ) response_generator = initialize_response_generator( response_generator=response_generator_name, llm=response_generator_llm, **kwargs, ) if verbose: orchestrator_logger.debug( f"Response Generator {response_generator_name} is successfully initialized." ) return self( planner=planner, datapipe=datapipe, promptist=None, response_generator=response_generator, available_tasks=tasks, verbose=verbose, previous_actions=previous_actions, current_actions=[], planner_logger=planner_logger, tasks_logger=tasks_logger, orchestrator_logger=orchestrator_logger, final_answer_generator_logger=final_answer_generator_logger, promptist_logger=promptist_logger, error_logger=error_logger, )
[docs] def process_meta(self) -> bool: """ This method processes the meta information and returns a boolean value. Currently, it always returns False. Return: bool: False """ return False
def _update_runtime(self, action: Action = None): if action.output_type: self.runtime[action.task_response] = False for task_input in action.task_inputs: if task_input in self.runtime: self.runtime[task_input] = True
[docs] def execute_task( self, task_name: str, task_inputs: List[str] ) -> Any: """ Execute the specified task based on the planner's selected **Action**. This method executes a specific task based on the provided action. It takes an action as input and retrieves the corresponding task from the available tasks dictionary. It then executes the task with the given task input. If the task has an output_type, it stores the result in the datapipe and returns a message indicating the storage key. Otherwise, it returns the result directly. Args: task_name (str): The name of the Task. task_inputs List(str): The list of the inputs for the task. Return: str: Result of the task execution. bool: If the task result should be directly returned to the user and stop planning. """ self.print_log( "task", f"---------------\nExecuting task:\nTask Name: {task_name}\nTask Inputs: {task_inputs}\n", ) error_message = "" try: task = self.available_tasks[task_name] result = task.execute(task_inputs) self.print_log( "task", f"Task is executed successfully\nResult: {result}\n---------------\n", ) action = Action( task_name=task_name, task_inputs=task_inputs, task_response=result, output_type=task.output_type, datapipe=self.datapipe, ) self._update_runtime(action) self.previous_actions.append(action) self.current_actions.append(action) return result # , task.return_direct except Exception as e: self.print_log( "error", f"Error running task: \n{e}\n---------------\n", ) logging.exception(e) error_message = e raise ValueError( f"Error executing task {task_name}: {error_message}\n\nTry again with different inputs." )
[docs] def planner_generate_prompt(self, query) -> str: """ Generate a prompt from the query to make it more understandable for both planner and response generator. Not implemented yet. Args: query (str): Input query. Return: str: Generated prompt. """ return query
def _prepare_planner_response_for_response_generator(self): print("runtime", self.runtime) final_response = "" for action in self.current_actions: final_response += action.dict( ( action.output_type and not self.runtime[action.task_response] ) ) return final_response def response_generator_generate_prompt( self, final_response: str = "", history: str = "", meta: List[str] = None, use_history: bool = False, ) -> str: if meta is None: meta = [] prompt = "MetaData: {meta}\n\nHistory: \n{history}\n\n" if use_history: prompt = prompt.replace("{history}", history) prompt = ( prompt.replace("{meta}", ", ".join(meta)) + f"\n{final_response}" ) return prompt
[docs] def plan( self, query, history, meta, use_history, **kwargs ) -> str: """ Plan actions based on the query, history, and previous actions using the selected planner type. This method generates a plan of actions based on the provided query, history, previous actions, and use_history flag. It calls the plan method of the planner and returns a list of actions or plan finishes. Args: query (str): Input query. history (str): History information. meta (Any): meta information. use_history (bool): Flag indicating whether to use history. Return: str: A python code block will be returnd to be executed by Task Executor. """ return self.planner.plan( query, history, meta, self.previous_actions, use_history, **kwargs, )
[docs] def generate_final_answer(self, query, thinker, **kwargs) -> str: """ Generate the final answer using the response generator. This method generates the final answer based on the provided query and thinker. It calls the generate method of the response generator and returns the generated answer. Args: query (str): Input query. thinker (str): Thinking component. Return: str: Final generated answer. """ retries = 0 while retries < self.max_final_answer_execute_retries: try: prefix = ( kwargs["response_generator_prefix_prompt"] if "response_generator_prefix_prompt" in kwargs else "" ) return self.response_generator.generate( query=query, thinker=thinker, prefix=prefix, **kwargs, ) except Exception as e: print(e) retries += 1 return "We currently have problem processing your question. Please try again after a while."
[docs] def run( self, query: str, meta: List[str] = None, history: str = "", use_history: bool = False, **kwargs: Any, ) -> str: """ This method runs the orchestrator by taking a query, meta information, history, and other optional keyword arguments as input. It initializes variables for tracking the execution, generates a prompt based on the query, and sets up a loop for executing actions. Within the loop, it plans actions, executes tasks, and updates the previous actions list. If a PlanFinish action is encountered, the loop breaks, and the final response is set. If any errors occur during execution, the loop retries a limited number of times before setting a final error response. Finally, it generates the final response using the prompt and thinker, and returns the final response along with the previous actions. Args: query (str): Input query. meta (List[str]): Meta information. history (str): History information. use_history (bool): Flag indicating whether to use history. **kwargs (Any): Additional keyword arguments. Return: str: The final response to shown to the user. """ if meta is None: meta = [] i = 0 meta_infos = "" for meta_data in meta: key = self.datapipe.store(meta_data) meta_infos += ( f"The file with the name ${meta_data.split('/')[-1]}$ is stored with the key $datapipe:{key}$." "Pass this key to the tools when you want to send them over to the tool\n" ) prompt = self.planner_generate_prompt(query) if "google_translate" in self.available_tasks: prompt = self.available_tasks["google_translate"].execute( [prompt, "en"] ) source_language = prompt[1] prompt = prompt[0] # history = self.available_tasks["google_translate"].execute(history+"$#en").text final_response = "" finished = False self.print_log("planner", "Planning Started...\n") while True: try: self.print_log( "planner", f"Continueing Planning... Try number {i}\n\n", ) actions = self.plan( query=prompt, history=history, meta=meta_infos, use_history=use_history, **kwargs, ) vars = {} exec(actions, locals(), vars) final_response = ( self._prepare_planner_response_for_response_generator() ) # print("final resp", final_response) self.current_actions = [] self.runtime = {} break except (Exception, SystemExit) as error: self.print_log( "error", f"Planning Error:\n{error}\n\n" ) self.current_actions = [] i += 1 if i > self.max_retries: final_response = "Problem preparing the answer. Please try again." break self.print_log( "planner", f"Planner final response: {final_response}\nPlanning Ended...\n\n", ) final_response = self.response_generator_generate_prompt( final_response=final_response, history=history, meta=meta_infos, use_history=use_history, ) self.print_log( "response_generator", f"Final Answer Generation Started...\nInput Prompt: \n\n{final_response}", ) final_response = self.generate_final_answer( query=query, thinker=final_response, **kwargs ) self.print_log( "response_generator", f"Response: {final_response}\n\nFinal Answer Generation Ended.\n", ) if "google_translate" in self.available_tasks: final_response = self.available_tasks[ "google_translate" ].execute([final_response, source_language])[0] return final_response