本文将介绍如何使用AgentScope框架构建一个简单的多智能体群聊系统,并解释其背后的实现逻辑。
首先写好设置文件。
agent_config.json
[
{
"class": "DialogAgent",
"args":{
"name": "Lingfeng",
"sys_prompt": "You are Lingfeng, a noble in the kingdom of Xi'an. You are a wise and powerful wizard.",
"model_config_name": "qwen_config",
"use_memory": true
}
},
{
"class": "DialogAgent",
"args":{
"name": "Boyu",
"sys_prompt": "You are Boyu, a friend of Lingfeng. Your speech is moderately humorous.",
"model_config_name": "qwen_config",
"use_memory": true
}
},
{
"class": "DialogAgent",
"args":{
"name": "Haotian",
"sys_prompt": "You are Haotian, Lingfeng's cousin, a young man who is studying at the university.",
"model_config_name": "qwen_config",
"use_memory": true
}
}
]
model_config.json
{
"config_name": "qwen_config",
"model_type": "dashscope_chat",
"model_name": "qwen-turbo",
"api_key": "xxxxx",
"generate_kwargs": {
"temperature": 0.5
}
}
实现代码
主逻辑 main.py
import agentscope
from agentscope.agents import DialogAgent, UserAgent
from agentscope.message import Msg
from agentscope.msghub import msghub
import logging
from groupchat_utils import (
select_next_one,
filter_agents,
)
# 定义用户发言超时时间
USER_TIME_TO_SPEAK = 30
# 默认聊天话题
DEFAULT_TOPIC = """
This is a chat room and you can speak freely and briefly.
"""
SYS_PROMPT = """
You can designate a member to reply to your message, you can use the @ symbol.
This means including the @ symbol in your message, followed by
that person's name, and leaving a space after the name.
All participants are: {agent_names}
"""
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
def main() -> None:
# 初始化NPC代理
npc_agents = agentscope.init(
model_configs="./model_config.json",
agent_configs="./agent_config.json",
project="Conversation with Mentions",
)
user = UserAgent()
agents = list(npc_agents) + [user]
# 创建初始提示消息
hint = Msg(
name="Host",
content=DEFAULT_TOPIC
+ SYS_PROMPT.format(
agent_names=[agent.name for agent in agents],
),
role="assistant",
)
rnd = 0
speak_list = []
# 使用消息中心管理代理之间的通信
with msghub(agents, announcement=hint):
while True:
try:
x = user(timeout=USER_TIME_TO_SPEAK)
if x.content == "exit":
break
except TimeoutError:
x = {"content": ""}
logger.info(
f"User has not typed text for "
f"{USER_TIME_TO_SPEAK} seconds, skip."
)
# 根据用户输入过滤出需要发言的代理
speak_list += filter_agents(x.content, npc_agents)
# print("speak_list_before:", speak_list)
# for agent in speak_list:
# print(f"Name: {agent.name}")
# 如果有需要发言的代理,则选择第一个
if len(speak_list) > 0:
next_agent = speak_list.pop(0)
x = next_agent()
else:
# 否则根据轮次选择下一个代理
next_agent = select_next_one(npc_agents, rnd)
x = next_agent()
speak_list += filter_agents(x.content, npc_agents)
# print("speak_list_after:", speak_list)
rnd += 1
if __name__ == "__main__":
main()
辅助函数 groupchat_utils.py
import re
from typing import Sequence
def filter_agents(string: str, agents: Sequence) -> Sequence:
"""
筛选出字符串中包含的代理名称,并返回找到的代理对象列表。
"""
if len(agents) == 0:
return []
pattern = (
r"@(" + "|".join(re.escape(agent.name) for agent in agents) + r")\b"
)
matches = re.findall(pattern, string)
agent_dict = {agent.name: agent for agent in agents}
ordered_agents = [
agent_dict[name] for name in matches if name in agent_dict
]
return ordered_agents
def select_next_one(agents: Sequence, rnd: int) -> Sequence:
"""
根据轮次选择下一个发言的代理。
"""
return agents[rnd % len(agents)]
代码解析
这段代码实现了多智能体之间的群聊功能。用户可以与多个智能体(NPC)进行互动,而这些智能体之间也能相互交流。关键在于:
- 初始化智能体:通过
agentscope.init()
函数加载配置文件,并创建智能体实例。 - 消息传递:使用
msghub
来管理智能体间的通讯。 - 指定回复者:通过在消息中包含
@
符号加上特定智能体的名字来指定回复的对象。 - 轮询机制:如果没有明确指定回复者,则按照轮询的方式选择下一个发言的智能体。
最终效果
运行上述代码后,用户可以参与到与多个智能体的对话中。用户输入的内容将被解析,如果含有@
符号,则直接指定了下一个应该回应的智能体。如果没有指定,则按顺序选择智能体回应。
代码中的print部分为我加入的输出,可以看到目前在运行中还存在一些问题。 1.如果名字不打的完全准确是无法识别的
2.如果一次@了两个人,中间会需要用户进行输入
3.让NPC互相聊天会出现一个NPC把所有的NPC对话全部生成的情况。
有一些推想的解决方案,还没有实行,但是暂且列出来:
1.引入模糊匹配机制解决名字识别问题。通过使用正则表达式或其他字符串匹配算法,系统可以识别近似的名称输入,并将其映射到正确的NPC上。
2.可以通过调整代码逻辑解决。
3.目前还没有解决方案。
标签:群聊,name,list,agent,AgentScope,构建,agents,config,speak From: https://blog.51cto.com/u_16958225/12031469