背景
在完整的AI Agent中,工具的调用这里采用ChatGPT的Function Calling的思想,通过定义“Tools Schema”来连接LLM和Tools的调用.
1,先定义Tools Schema
这里用到的工具: Search Database, Send Mail, Get Weather等等. 这些工具也就是自定义工具方法的名字.
tools = [
{
"type": "function",
"function": {
"name": "Search Database",
"description": "Retrive profile from database",
"parameters":{
"type": "object",
"properties": {
"name": {"type": "string"}
},
"required": ["name"]
}
}
},
{
"type": "function",
"function": {
"name": "Send Mail",
"description": "Send Mail to Gmail",
"parameters":{
"to": {"type": "string"},
"subject": {"type": "string"},
"body": {"type": "string"}
}
}
},
{
"type": "function",
"function": {
"name": "Get Weather",
"description": "Get Weather from location",
"parameters":{
"type": "object",
"properties": {
"location": {"type": "string"}
},
"required": ["location"]
}
}
},
......
]
2, 循环执行Execute (Plan-and-Execute框架)
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.
’
import json
# 1️⃣ 让 Planner 生成计划
planner_response = client.chat.completions.create(
model="gpt-4o",
messages=[
{"role": "system", "content": planner_prompt},
{"role": "user", "content": user_query}
]
)
plans = json.loads(planner_response.choices[0].message.content)["plans"]
# 2️⃣ 逐步执行
context_memory = {}
for step in plans:
tool_name = step["tool"]
args = step["arguments"]
result = execute_tool(tool_name, args, context_memory)
# 把结果加入上下文
context_memory[f"step_{step['step_id']}"] = result
def execute_tool(tool_name, args, memory):
if tool_name == "Search Database":
return search_Database(args["name"])
elif tool_name == "Send Mail":
return send_Mail(args["to"],args["subject"],args["body"])
elif tool_name == "Get Weather":
return get_Weather(args["location"])
else:
raise Exception("Unknown tool")
评价: (Plan-and-Execute框架 + Function Calling 推荐指数:🌟🌟)
优点:
1, 任务分解: 通过将大任务分解为小任务,可以有效管理和解决复杂任务.
2, 详细指导: 通过提供详细的指导来改善推理步骤的质量和准确性.
3, 适应性: 可以根据不同类型的任务进行调整,在各种复杂问题中表现出色.
缺点:
1, Planner Prompt一次性生成规划,步骤可能会丢失,乱序,中间状态无法感知
2, 循环执行Execute 过于机械,不具备条件跳转,分支,动态决策,失败重试
3, Planner 不知道Execute的结果,是否为空等等
适合场景:
1, 步骤固定,流程稳定
2, 不需要动态决策,不需要分支逻辑
不适合场景:
1, 决策型任务
2, 条件跳转,
3, 数据驱动推理
4, 长链条任务
