from typing import List, Union import traceback import copy from loguru import logger from coagent.connector.schema import ( Memory, Task, Env, Role, Message, ActionStatus, PromptField, LogVerboseEnum ) from coagent.connector.memory_manager import BaseMemoryManager from coagent.connector.configs.agent_config import REACT_PROMPT_INPUT from coagent.llm_models import LLMConfig, EmbedConfig from .base_agent import BaseAgent from coagent.connector.memory_manager import LocalMemoryManager from coagent.connector.prompt_manager import PromptManager class ReactAgent(BaseAgent): def __init__( self, role: Role, prompt_config: [PromptField], prompt_manager_type: str = "PromptManager", task: Task = None, memory: Memory = None, chat_turn: int = 1, focus_agents: List[str] = [], focus_message_keys: List[str] = [], # llm_config: LLMConfig = None, embed_config: EmbedConfig = None, sandbox_server: dict = {}, jupyter_work_path: str = "", kb_root_path: str = "", log_verbose: str = "0" ): super().__init__(role, prompt_config, prompt_manager_type, task, memory, chat_turn, focus_agents, focus_message_keys, llm_config, embed_config, sandbox_server, jupyter_work_path, kb_root_path, log_verbose ) def step(self, query: Message, history: Memory = None, background: Memory = None, memory_manager: BaseMemoryManager = None) -> Message: '''agent reponse from multi-message''' for message in self.astep(query, history, background, memory_manager): pass return message def astep(self, query: Message, history: Memory = None, background: Memory = None, memory_manager: BaseMemoryManager = None) -> Message: '''agent reponse from multi-message''' step_nums = copy.deepcopy(self.chat_turn) react_memory = Memory(messages=[]) # insert query output_message = Message( role_name=self.role.role_name, role_type="assistant", #self.role.role_type, role_content=query.input_query, step_content="", input_query=query.input_query, tools=query.tools, # parsed_output_list=[query.parsed_output], customed_kargs=query.customed_kargs ) query_c = copy.deepcopy(query) query_c = self.start_action_step(query_c) # if query.parsed_output: # query_c.parsed_output = {"Question": "\n".join([f"{v}" for k, v in query.parsed_output.items() if k not in ["Action Status"]])} # else: # query_c.parsed_output = {"Question": query.input_query} # react_memory.append(query_c) # self_memory = self.memory if self.do_use_self_memory else None idx = 0 # start to react while step_nums > 0: output_message.role_content = output_message.step_content # prompt = self.create_prompt(query, self.memory, history, background, react_memory, memory_manager.current_memory) if memory_manager is None: memory_manager = LocalMemoryManager( unique_name=self.role.role_name, do_init=True, kb_root_path = self.kb_root_path, embed_config=self.embed_config, llm_config=self.embed_config ) memory_manager.append(query) memory_pool = memory_manager.current_memory else: memory_pool = memory_manager.current_memory prompt = self.prompt_manager.generate_full_prompt( previous_agent_message=query_c, agent_long_term_memory=self.memory, ui_history=history, chain_summary_messages=background, react_memory=react_memory, memory_pool=memory_pool) try: content = self.llm.predict(prompt) except Exception as e: logger.error(f"error prompt: {prompt}") raise Exception(traceback.format_exc()) output_message.role_content = "\n"+content output_message.step_content += "\n"+output_message.role_content yield output_message if LogVerboseEnum.ge(LogVerboseEnum.Log2Level, self.log_verbose): logger.debug(f"{self.role.role_name}, {idx} iteration prompt: {prompt}") if LogVerboseEnum.ge(LogVerboseEnum.Log1Level, self.log_verbose): logger.info(f"{self.role.role_name}, {idx} iteration step_run: {output_message.role_content}") output_message = self.message_utils.parser(output_message) # when get finished signal can stop early if output_message.action_status == ActionStatus.FINISHED or output_message.action_status == ActionStatus.STOPPED: output_message.parsed_output_list.append(output_message.parsed_output) break # according the output to choose one action for code_content or tool_content output_message, observation_message = self.message_utils.step_router(output_message) output_message.parsed_output_list.append(output_message.parsed_output) react_message = copy.deepcopy(output_message) react_memory.append(react_message) if observation_message: react_memory.append(observation_message) output_message.parsed_output_list.append(observation_message.parsed_output) # logger.debug(f"{observation_message.role_name} content: {observation_message.role_content}") idx += 1 step_nums -= 1 yield output_message # react' self_memory saved at last self.append_history(output_message) output_message.input_query = query.input_query # end_action_step, BUG:it may cause slack some information output_message = self.end_action_step(output_message) # update memory pool memory_manager.append(output_message) yield output_message def pre_print(self, query: Message, history: Memory = None, background: Memory = None, memory_manager: BaseMemoryManager=None): react_memory = Memory(messages=[]) prompt = self.prompt_manager.pre_print( previous_agent_message=query, agent_long_term_memory=self.memory, ui_history=history, chain_summary_messages=background, react_memory=react_memory, memory_pool=memory_manager.current_memory) title = f"<<<<{self.role.role_name}'s prompt>>>>" print("#"*len(title) + f"\n{title}\n"+ "#"*len(title)+ f"\n\n{prompt}\n\n") # def create_prompt( # self, query: Message, memory: Memory =None, history: Memory = None, background: Memory = None, react_memory: Memory = None, memory_manager: BaseMemoryManager= None, # prompt_mamnger=None) -> str: # prompt_mamnger = PromptManager() # prompt_mamnger.register_standard_fields() # # input_keys = parse_section(self.role.role_prompt, 'Agent Profile') # data_dict = { # "agent_profile": extract_section(self.role.role_prompt, 'Agent Profile'), # "tool_information": query.tools, # "session_records": memory_manager, # "reference_documents": query, # "output_format": extract_section(self.role.role_prompt, 'Response Output Format'), # "response": "\n".join(["\n".join([f"**{k}:**\n{v}" for k,v in _dict.items()]) for _dict in react_memory.get_parserd_output()]), # } # # logger.debug(memory_pool) # return prompt_mamnger.generate_full_prompt(data_dict) def create_prompt( self, query: Message, memory: Memory =None, history: Memory = None, background: Memory = None, react_memory: Memory = None, memory_pool: Memory= None, prompt_mamnger=None) -> str: ''' role\task\tools\docs\memory ''' # doc_infos = self.create_doc_prompt(query) code_infos = self.create_codedoc_prompt(query) # formatted_tools, tool_names, _ = self.create_tools_prompt(query) task_prompt = self.create_task_prompt(query) background_prompt = self.create_background_prompt(background) history_prompt = self.create_history_prompt(history) selfmemory_prompt = self.create_selfmemory_prompt(memory, control_key="step_content") # # extra_system_prompt = self.role.role_prompt prompt = self.role.role_prompt.format(**{"formatted_tools": formatted_tools, "tool_names": tool_names}) # react 流程是自身迭代过程，另外二次触发的是需要作为历史对话信息 # input_query = react_memory.to_tuple_messages(content_key="step_content") # # input_query = query.input_query + "\n" + "\n".join([f"{v}" for k, v in input_query if v]) # input_query = "\n".join([f"{v}" for k, v in input_query if v]) input_query = "\n".join(["\n".join([f"**{k}:**\n{v}" for k,v in _dict.items()]) for _dict in react_memory.get_parserd_output()]) # logger.debug(f"input_query: {input_query}") prompt += "\n" + REACT_PROMPT_INPUT.format(**{"query": input_query}) task = query.task or self.task # if task_prompt is not None: # prompt += "\n" + task.task_prompt # if doc_infos is not None and doc_infos!="" and doc_infos!="不存在知识库辅助信息": # prompt += f"\n知识库信息: {doc_infos}" # if code_infos is not None and code_infos!="" and code_infos!="不存在代码库辅助信息": # prompt += f"\n代码库信息: {code_infos}" # if background_prompt: # prompt += "\n" + background_prompt # if history_prompt: # prompt += "\n" + history_prompt # if selfmemory_prompt: # prompt += "\n" + selfmemory_prompt # logger.debug(f"{self.role.role_name} extra_system_prompt: {self.role.role_prompt}") # logger.debug(f"{self.role.role_name} input_query: {input_query}") # logger.debug(f"{self.role.role_name} doc_infos: {doc_infos}") # logger.debug(f"{self.role.role_name} tool_names: {tool_names}") # prompt += "\n" + REACT_PROMPT_INPUT.format(**{"query": input_query}) # prompt = extra_system_prompt.format(**{"query": input_query, "doc_infos": doc_infos, "formatted_tools": formatted_tools, "tool_names": tool_names}) while "{{" in prompt or "}}" in prompt: prompt = prompt.replace("{{", "{") prompt = prompt.replace("}}", "}") return prompt