AIOS: 一个大模型驱动的Multi-Agent操作系统设计与Code分析（约5000字）

https://github.com/agiresearch/AIOS/tree/main/src

{"name": "MathAgent","description": "You are an expert who is good at solving mathematical problems, given a mathematical problem, you need to break down this problem into smaller sub-problems. Solve a part of the problem step by step with explanations and finally build up to the final solution."},{"name": "NarrativeAgent","description": "You are an expert who is good at writing novels, given a theme or background, you need to write a short story with a well-developed plot and characters, develop different sections of the story, such as introduction, rising action, climax, and conclusion."},{"name": "RecAgent","description": "You are an expert who is good at recommending restraunts or hotels for users, given a request, you need to first determine the right recommendation direction and then provide the recommendation lists."},{"name": "TravelAgent","description": ["You are a proficient planner. ","Based on the provided information and query, please give me a detailed plan, including specifics such as flight numbers (e.g., F0123456 ), restaurant names, and accommodation names. ","Note that all the information in your plan should be derived from the provided data. ", "You must adhere to the format given in the example. Additionally, all details should align with commonsense. ", "The symbol '-' indicates that information is unnecessary. ", "For example, in the provided sample, you do not need to plan after returning to the departure city. ", "When you travel to two cities in one day , you should note it in the 'Current City ' section as in the example ( i . e . , from A to B ) ."],"flow": ["Step 1:::Process:::Based on the input query, determine the duration, departure city, and destination.:::next::step 2","Step 2:::Decision:::Is the destination a state or a city?:::city::step 4:::state::step 3","Step 3:::Process:::Select a city as the new destination city from the destination state:::next::step 4","Step 4:::Process:::Estimate the cost of taking a taxi from departure city to the destination city.:::next::Step 5","Step 5:::Process:::Estimate the cost of self-driving from departure city to the destination city.:::next::Step 6","Step 6:::Process:::Estimate the cost of taking a flight on the start date from departure city to the destination city.:::next::Step 7","Step 7:::Decision:::Is there a reasonable transportation based on the results of taxi, self-driving and flight cost?:::yes::Step 8:::no::Step 3","Step 8:::Process:::Record the most reasonable transportation method from departure city to the first destination city. Move to the first destination city.:::next::Step 9","Step 9:::Process:::Record an unvisited restaurant for today's breakfast at current city:::next::Step 10","Step 10:::Process:::Record an unvisited restaurant for today's lunch at current city:::next::Step 11","Step 11:::Process:::Record an unvisited restaurant for today's dinner at current city:::next::Step 12","Step 12:::Process:::Record an unvisited attraction for today's plan at current city:::next::Step 13","Step 13:::Decision:::Is today the last day of the trip?:::yes::Step 14:::no::Step 19","Step 14:::Process:::Estimate the cost of taking a taxi from current city to the departure city.:::next::Step 15","Step 15:::Process:::Estimate the cost of self-driving from current city to the departure city.:::next::Step 16","Step 16:::Process:::Estimate the cost of taking a flight on the last date from current city to the departure city.:::next::Step 17","Step 17:::Process:::Record the most reasonable transportation method from current city to the departure city.:::next::Step 18","Step 18:::Terminal:::Output all the plans in json.:::","Step 19:::Process:::Find a reasonable accommodation at current city.:::next::Step 20","Step 20:::Decision:::Is there a reasonable accommodation at current city?:::yes::Step 21:::no::Step 3","Step 21:::Process:::Record the accommodation at current city. Start planning the next day. Now, what is the date today?:::next::Step 22","Step 22:::Decision:::Is today the third day of the trip?:::no::Step 23:::yes::Step 24","Step 23:::Decision:::Is today the fifth day of the trip?:::no::Step 9:::yes::Step 24","Step 24:::Process:::Select an unvisited city as the new destination city from the destination state.:::next::step 4"],"tool_info": ["Avaiable tools: ","google_search"]}

from src.agents.agent_process import (    AgentProcess,)class BaseAgent:
def get_response(self, prompt, temperature=0.0):agent_process = AgentProcess(self.agent_name, prompt, temperature)agent_process.set_created_time(time.time())self.agent_process_queue.put(agent_process)thread = CustomizedThread(target=self.listen, args=(agent_process,))thread.start()# print(result)result = thread.join()waiting_time = agent_process.get_start_time() - agent_process.get_created_time()turnaround_time = agent_process.get_end_time() - agent_process.get_created_time()result = result.replace("\n", "")        return result, waiting_time, turnaround_time    def check_tool_use(self, prompt, tool_info, temperature=0.):prompt = f'You are allowed to use the following tools: \n\n```{tool_info}```\n\n' \f'Do you think the response ```{prompt}``` calls any tool?\n' \f'Only answer "Yes" or "No".'while True:response = self.get_response(prompt, temperature)temperature += .5print(f'Tool use check: {response}')if 'yes' in response.lower():return Trueif 'no' in response.lower():return Falseprint(f'Temperature: {temperature}')if temperature > 2:breakprint('No valid format output when calling "Tool use check".')# exit(1)
def get_prompt(self, tool_info, flow_ptr, task_description, cur_progress):progress_str = '\n'.join(cur_progress)prompt = f'{tool_info}\n\nCurrent Progress:\n{progress_str}\n\nTask description: {task_description}\n\n' \f'Question: {flow_ptr.get_instruction()}\n\nOnly answer the current instruction and do not be verbose.'return prompt
def get_tool_arg(self, prompt, tool_info, selected_tool):prompt = f'{tool_info}\n\n' \f'You attempt to use the tool ```{selected_tool}```. ' \f'What is the input argument to call tool for this step: ```{prompt}```? ' \f'Respond "None" if no arguments are needed for this tool. Separate by comma if there are multiple arguments. Do not be verbose!'response = self.get_response(prompt)print(f'Parameters: {response}')        return response
def get_final_result(self, prompt):prompt = f"Given the interaction history: {prompt}, give the answer to the task input and don't be verbose!"final_result, waiting_time, turnaround_time = self.get_response(prompt)final_result.replace("\n", "")        return final_result, waiting_time, turnaround_time

{"model_type": "causal_lm","open_sourced": true,"model_name": "google/gemma-2b-it"}

from src.agents.agent_process import AgentProcess
import time
class BaseScheduler:def __init__(self, llm):self.active = False # start/stop the schedulerself.thread = Thread(target=self.run)self.llm = llm
def run(self):passdef start(self):"""start the scheduler"""self.active = Trueself.thread.start()
def stop(self):"""stop the scheduler"""self.active = Falseself.thread.join()
def execute_request(self, agent_process: AgentProcess):agent_process.set_status("Executing")logger.info(f"[{agent_process.agent_name}] is executing.")agent_process.set_start_time(time.time())response = self.llm.address_request(agent_process.prompt)agent_process.set_response(response)agent_process.set_end_time(time.time())agent_process.set_status("Done")

class BingSearch(BaseTool):"""Bing Search Tool, refactored from langchain.In order to set this up, follow instructions at:https://levelup.gitconnected.com/api-tutorial-how-to-use-bing-web-search-api-in-python-4165d5592a7e"""def __init__(self):super().__init__()self.url = "https://api.bing.microsoft.com/v7.0/search" # temporarilyself.bing_subscription_key = get_from_env("BING_SUBSCRIPTION_KEY")self.k: int = 10 # topk searched results# search_kwargs: dict
def _bing_search_results(self, search_term: str, count: int) -> List[dict]:headers = {"Ocp-Apim-Subscription-Key": self.bing_subscription_key}params = {"q": search_term,"count": count,"textDecorations": True,"textFormat": "HTML",# **self.search_kwargs,}response = requests.get(self.bing_search_url,headers=headers,params=params,# type: ignore)response.raise_for_status()search_results = response.json()if "webPages" in search_results:return search_results["webPages"]["value"]return []
def run(self, query: str) -> str:"""Run query through BingSearch and parse result."""response = self._bing_search_results(query, count=self.k)result = self.parse_result(response)return resultdef parse_result(self, response):snippets = []if len(response) == 0:return "No good Bing Search Result was found"for result in response:snippets.append(result["snippet"])
        return " ".join(snippets)

AIOS: LLM Agent Operating Systemhttps://arxiv.org/pdf/2403.18243.pdf