ReActフレームワーク：Think、Act、Observe

ReActフレームワークとは

Think-Act-Observeループ

ReActトレースの例

# ReAct trace (pseudocode showing the reasoning loop)

# Iteration 1
Thought_1 = 'I need to find the current population of Tokyo. I will search for it.'
Action_1 = 'search("Tokyo population 2024")'
Observation_1 = 'According to the 2023 census, Tokyo has approximately 13.96 million people in the city proper.'

# Iteration 2
Thought_2 = 'I have the population. I can now give a final answer.'
Final_Answer = 'The population of Tokyo is approximately 13.96 million people.'

ReActが直接プロンプトより優れている理由

ReActプロンプトの構造

REACT_SYSTEM_PROMPT = '''You are an AI assistant with access to these tools:

- search(query: str): Search the web for up-to-date information.
- calculator(expression: str): Evaluate a mathematical expression.
- lookup(topic: str): Look up a topic in the knowledge base.

Always follow this exact format:
Thought: [your reasoning about what to do next]
Action: tool_name(arguments)
Observation: [result of the action, provided by the system]
... (repeat as needed)
Final Answer: [your final response to the user]

Begin!'''

Thought、Action、Observationの解析

import re

def parse_react_output(text: str):
    '''Extract action name and argument from a ReAct model output.'''
    action_match = re.search(r'Action:\s*(\w+)\((.*)\)', text)
    if action_match:
        tool_name = action_match.group(1)
        tool_input = action_match.group(2).strip('\"\' ')
        return tool_name, tool_input
    if 'Final Answer:' in text:
        answer = text.split('Final Answer:')[-1].strip()
        return 'final', answer
    return None, None

エージェントの実行ループ

from openai import OpenAI

client = OpenAI()

def run_react_agent(user_query: str, tools: dict, max_iterations: int = 10) -> str:
    messages = [
        {'role': 'system', 'content': REACT_SYSTEM_PROMPT},
        {'role': 'user', 'content': user_query}
    ]

    for i in range(max_iterations):
        resp = client.chat.completions.create(model='gpt-4o', messages=messages)
        output = resp.choices[0].message.content
        messages.append({'role': 'assistant', 'content': output})

        tool_name, tool_input = parse_react_output(output)

        if tool_name == 'final':
            return tool_input

        if tool_name and tool_name in tools:
            observation = tools[tool_name](tool_input)
            messages.append({'role': 'user', 'content': f'Observation: {observation}'})
        else:
            messages.append({'role': 'user', 'content': 'Observation: Tool not found.'})

    return 'Max iterations reached without a final answer.'

シンプルなツールを登録する

import math

def calculator_tool(expression: str) -> str:
    try:
        # Restrict to safe math expressions
        allowed_names = {k: v for k, v in math.__dict__.items() if not k.startswith('_')}
        result = eval(expression, {'__builtins__': {}}, allowed_names)
        return str(result)
    except Exception as e:
        return f'Error: {e}'

def search_tool(query: str) -> str:
    # Stub — replace with real web search API
    return f'Top result for "{query}": [placeholder result]'

# Register tools
tools = {
    'calculator': calculator_tool,
    'search': search_tool
}

ReActによるマルチホップ推論

ReActと純粋なChain-of-Thoughtの比較

トレースを使ってReActをデバッグする

理解度チェック

レッスンのまとめ