เฟรมเวิร์ก ReAct: คิด ลงมือทำ สังเกต

ReAct (Reasoning + Acting) is a prompting framework that interleaves the model's internal reasoning with external actions. Instead of producing a final answer immediately, the model generates a Thought, chooses an Action, receives an Observation from executing that action, and repeats the cycle until it can give a Final Answer.

Each iteration of the ReAct loop has three phases:

• Thought: The model reasons about what it knows and what it needs to find out next.
• Action: The model calls a tool — such as a web search, a calculator, or a database lookup.
• Observation: The result of the tool call is returned and appended to the model's context.

The loop continues until the model emits Final Answer: with its conclusion.

Here is a concrete trace of a ReAct agent answering 'What is the population of Tokyo?' using a search tool. Notice how each step builds on the previous observation.

Direct prompting asks the model to answer from memory, leading to hallucinations on factual questions. Chain-of-thought prompting improves reasoning but still cannot access external information. ReAct combines both: explicit reasoning traces like chain-of-thought plus the ability to gather fresh information through tool calls.

The ReAct system prompt teaches the model the exact format to use. It lists available tools with their descriptions and shows examples of the Thought/Action/Observation/Final Answer pattern. The model learns to follow this format precisely from the prompt.

Your application code acts as the runtime for the ReAct loop. After each model response, parse the output to extract the action name and arguments, execute the corresponding Python function, format the result as an observation, append it to the message history, and call the model again.

The execution loop calls the LLM, parses its output, dispatches the tool, appends the observation, and loops. A max_iterations guard prevents infinite loops when the model cannot resolve a task.

Tools are just Python functions. You register them in a dictionary mapping tool name to function. The functions receive a string argument and return a string result — this keeps the interface simple and consistent with what the model expects to see as observations.

ReAct shines on multi-hop questions that require chaining multiple pieces of information. For example: 'Who is the CEO of the company that makes GPT-4, and what year was that company founded?' The agent first searches for the CEO, then uses that result to look up the company's founding year — two separate tool calls chained by reasoning.

Chain-of-Thought (CoT) improves reasoning by having the model think step by step, but it cannot access external information. ReAct adds actions to CoT, letting the model fetch real data, run computations, and verify facts. The trade-off: ReAct is slower (multiple API calls) but far more accurate on tasks requiring current or specialized knowledge.

When a ReAct agent produces the wrong answer, inspect the full Thought/Action/Observation trace. Common failure patterns include: the model inventing an observation instead of calling the tool, parsing errors that skip tool execution, and incorrect reasoning in the Thought step that leads to a wrong action choice.

• Log every message appended to the context
• Check that tool output was actually appended before the next model call
• Verify the action regex matches the model's output format

Test your understanding of the ReAct framework.

In this lesson you learned: ReAct interleaves Thought, Action, and Observation in a loop, your application code acts as the runtime that executes tools and appends observations, and max_iterations guards prevent infinite agent loops. Next up we learn to define custom tools so your agent can do exactly what your application needs.