背景
LangChain的Plan-and-Execute框架受到关于Plan-and-Solve的论文的启发.Plan-and-Execute非常适合更复杂的长期规划,把复杂问题拆分成一个个子任务,逐个击破. 这会频繁的的调用大模型,但可以避免ReAct Agent循环过程中产生提示词过长的问题.
Plan-and-Solve 策略
上图是无样本COT和Plan-and-Solve的对比. Plan-and-Solve本质上将COT具体化为一个个小任务,然后根据这一个个小任务去一个个的解决任务. Plan-and-Execute框架和Plan-and-Solve的思路是一样的,但在执行上,Plan-and-Execute框架更偏向于AI实操.
Plan-and-Execute框架的应用举例
适合Question ReWriting设计和多Agent协调工作等等
场景: 在一个构建了RAG系统的完整AI Agent系统中
例如:
Question: 查找Louis和刘亦菲的个人信息, 并且根据他们的个人信息生成一个结婚请柬,将此结婚请柬的内容发送到worldXXTest@gmail.com邮箱中.
Plan-and-Execute将任务分解为:Plan,Execute
利用LLM将Question分解为大概5个Plan:
|-1- 在数据库中查找Louis的个人信息
|-2- 在网络查找刘亦菲的个人信息
|-3- 在网络上查找结婚贺卡请柬的模版
|-4- 生成Louis和刘亦菲的结婚贺卡请柬内容
|-5- 发送内容到worldXXTest@gmail.com邮箱
利用ReAct框架的LLM一一处理这5个Plan, Execute:
|-1- ”Louis的个人信息:iOS and AI Engineer,男, 性格好,温和,善良,幽默......“
|-2- 调用Tool工具Browser,得到:“刘亦菲的个人信息: 茜茜, 女,影视和电影明星,颜值与实力并存,爱好自由......”
|-3- 结婚请柬的模版:“我们结婚啦, 您好xxx先生/女士,欢迎您在2026年xx月xx日,参加我们在...”
|-4- Louis和刘亦菲的结婚请柬内容:“我们结婚啦, 您好xxx先生/女士,欢迎您在2026年2月31日,参加我们在月球的北极宫殿举行的太阳历大婚礼,届时我们会派遣星际飞船LL001号去接您,请在2月30日晚上12点上海的东方明珠塔顶楼乘坐飞船.”
|-5- 调用工sendMail,发送结婚请柬内容
总结:调用一次LLM生成5个plan, 调用5次 ReAct Agent处理这5个Plan.
Planner Prompt:
prompt = ‘
You are a task planner in a multi-step AI system.
User query: {Query}
Your job:
1, Break down the user query into executable steps.
2, Each step must be atomic and tool-executable.
3, Steps should be ordered logically.
4, Do NOT execute anything.
5, Only return JSON.
’
例如:
Query: 我想查找一个拥有英语6级证书,并且Python开发经验有3年的人.然后将这个人的个人信息发送到luochuanad@gmail.com
通过LLM(Planner Prompt)得出以下结果:
json
[
{
"step: 1,
"action": "Search Database"
"description": "Query the database to find individuals with an English CET-6 certificate and 3 years of Python development experience.",
"parameters": {
"certificate": "CET-6",
"experience": {
"Language": "Python",
"years": 3
}
}
},
{
"step": 2,
"action": "Retrieve Personal Information",
"description": "Extract the personal information of the individual(s) found in the previous step.",
"parameters": {
"fields": ["name", "email", "phone", "address"]
}
},
{
"step": 3,
"action": "Format Email",
"description": "Prepare an email containing the personal information of the individual(s) found.",
"parameters": {
"recipient": "luochuanad@gmail.com",
"subject": "Candidate Information",
"body": Include the personal information retrieved in step 2."
}
},
{
"step": 4,
"action": "Send Email"
"description": "Send the formatted email to the specified email address.",
"parameters": {
"recipient": "xxx@gmail.com",
"content": "Use the email content prepared in step 3."
}
}
]
参考
Lei Wang等人的论文“Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models” https://aclanthology.org/2023.acl-long.147.pdf
