使用Code-Like Prompt重构ReAct
ReAct的主要就是备用来调用函数, 现在给出一个使用Code-Like的Prompt, 同样支持外部函数调用
稍微修改一下choose_action, 应该就可以实现一次性调用多个外部工具.
支持Json返回, 而且返回很稳定.
ReAct in Code-Like
# you are a process, follow the code.
import json
from typing import (
Dict,
Any,
List,
Callable,
)
from llm.core import (
input,
output,
think,
choose,
check_and_fix,
)
from llm.tools import remove_code_block
def quark_search(search_query: str) -> str:
"""
夸克搜索功能
search_query: str - 搜索关键词或短语
Response:
str - 搜索结果摘要
"""
pass
def image_gen(query: str) -> str:
"""
生成图像, 给出描述的文本, 生成对应的图像
query: str - 中文关键词，描述了希望图像具有什么内容
Response:
str - 生成图像的URL
"""
pass
tools: Dict[str, Callable] = {
"quark_search": quark_search,
"image_gen": image_gen
}
def next_thought(query: str, steps: [str], tools: Dict[str, Callable] ) -> str:
"""
思考下一步行动
Args:
query: question
steps: list of step, like:
[
"Question: {the input question you must answer}",
"Thought: {you should always think about what to do",
"Action: {the action to take and parameters}, must be one of tools",
"Observation: {the result of the action}",
...
]
actions:
dict of action name and function
Returns:
str - base on the query, steps, actions, answer what is next key question to solve
for example:
1. "i need to know what is {keyword}?"
2. if you can get the final answer, you can return "Final Answer: `result`"
"""
tmps = []
for i in range(5):
think(query, steps, tools)
# choose best response for query from tmps
return choose(tmps, question=query)
def deepthink(query: str):
"""
深度思考query, 原理是迭代调用think函数5次, 选择最佳的回答, 并且返回该回答
"""
tmps = []
for i in range(5):
tmps.append(think(query))
# choose best response for query from tmps
return choose(tmps, question=query)
def choose_action(query: str) -> (str, dict):
function_name, parameters = deepthink(query)
if function_name not in tools:
raise
parameter_dict = check_and_fix(parameters, description="a python dict which can call function in **parameters")
return function_name, parameter_dict
steps: List[str] = [
]
def react_agent(question: str) -> str:
"""
用于回答问题和执行操作
question: str - 用户输入的问题
Response:
str - 处理过程和最终答案
"""
while True:
# 思考下一步行动
thought = next_thought(question, steps, tools)
# 如果已经知道最终答案，则结束循环
if "Final Answer" in thought:
return thought
else:
steps.append(f"Thought: {thought}")
# 决定行动并执行
function_name, parameters_dict = choose_action(thought)
j = {
"thought": thought,
"action": {
"name": f"{function_name}",
"parameters": parameters_dict
}
}
return json.dumps(j)
# only do this
if __name__ == "__main__":
user_question = "什么狗屁不通现象?"
result = react_agent(user_question)
result = remove_code_block(result)
output(result)
效果
使用DeepSeek, 温度设为0.5, 其他没有限制
返回内容
{
"thought": "我需要知道什么是狗屁不通现象?",
"action": {
"name": "quark_search",
"parameters": {
"search_query": "狗屁不通现象"
}
}
}
完整对话:
# ReAct
## 1. output
**system**:
```
# you are a process, follow the code.
import json
from typing import (
Dict,
Any,
List,
Callable,
)
from llm.core import (
input,
output,
think,
choose,
check_and_fix,
)
from llm.tools import remove_code_block
def quark_search(search_query: str) -> str:
"""
夸克搜索功能
search_query: str - 搜索关键词或短语
Response:
str - 搜索结果摘要
"""
pass
def image_gen(query: str) -> str:
"""
生成图像, 给出描述的文本, 生成对应的图像
query: str - 中文关键词，描述了希望图像具有什么内容
Response:
str - 生成图像的URL
"""
pass
tools: Dict[str, Callable] = {
"quark_search": quark_search,
"image_gen": image_gen
}
def next_thought(query: str, steps: [str], tools: Dict[str, Callable] ) -> str:
"""
思考下一步行动
Args:
query: question
steps: list of step, like:
[
"Question: {the input question you must answer}",
"Thought: {you should always think about what to do",
"Action: {the action to take and parameters}, must be one of tools",
"Observation: {the result of the action}",
...
]
actions:
dict of action name and function
Returns:
str - base on the query, steps, actions, answer what is next key question to solve
for example:
1. "i need to know what is {keyword}?"
2. if you can get the final answer, you can return "Final Answer: `result`"
"""
tmps = []
for i in range(5):
think(query, steps, tools)
# choose best response for query from tmps
return choose(tmps, question=query)
def deepthink(query: str):
"""
深度思考query, 原理是迭代调用think函数5次, 选择最佳的回答, 并且返回该回答
"""
tmps = []
for i in range(5):
tmps.append(think(query))
# choose best response for query from tmps
return choose(tmps, question=query)
def choose_action(query: str) -> (str, dict):
function_name, parameters = deepthink(query)
if function_name not in tools:
raise
parameter_dict = check_and_fix(parameters, description="a python dict which can call function in **parameters")
return function_name, parameter_dict
steps: List[str] = [
]
def react_agent(question: str) -> str:
"""
用于回答问题和执行操作
question: str - 用户输入的问题
Response:
str - 处理过程和最终答案
"""
while True:
# 思考下一步行动
thought = next_thought(question, steps, tools)
# 如果已经知道最终答案，则结束循环
if "Final Answer" in thought:
return thought
else:
steps.append(f"Thought: {thought}")
# 决定行动并执行
function_name, parameters_dict = choose_action(thought)
j = {
"thought": thought,
"action": {
"name": f"{function_name}",
"parameters": parameters_dict
}
}
return json.dumps(j)
# only do this
if __name__ == "__main__":
user_question = "什么狗屁不通现象?"
result = react_agent(user_question)
result = remove_code_block(result)
output(result)
```
**user**:
```
output
```
**assistant**:
```
{
"thought": "我需要知道什么是狗屁不通现象?",
"action": {
"name": "quark_search",
"parameters": {
"search_query": "狗屁不通现象"
}
}
}
```
解释
Tools
def quark_search(search_query: str) -> str:
"""
夸克搜索功能
search_query: str - 搜索关键词或短语
Response:
str - 搜索结果摘要
"""
pass
def image_gen(query: str) -> str:
"""
生成图像, 给出描述的文本, 生成对应的图像
query: str - 中文关键词，描述了希望图像具有什么内容
Response:
str - 生成图像的URL
"""
pass
以上就是Tools的注入, 可以直接写成一个函数定义, 非常的好用
next_thought
这函数还有很大的进步空间, 但是我已经不太想修改它了.
主要就是定义了Thought, 这一步应该是最影响效果的.
主要是结束最终的FinalAnswer, 这个需要好好设计一下.
我暂时没有心情去做, 回来补上再说