Gemini in Practice: Real Google Examples · Lesson 1

ReAct Agent: Reasoning + Acting

Implement the ReAct (Reasoning + Acting) pattern with Gemini 2.0 Flash and the Wikipedia API. The model generates Thought→Action→Observation chains to answer complex questions.

30 min read3 questions in quizReady prompt includedIn progress

Practical exercise

What to do after this lesson

Implement a ReAct agent with two tools: Wikipedia and a safe calculator (restricted eval). Ask a question requiring both search and computation, e.g. 'How many years has the Eiffel Tower been standing as of 2025?'

Ready-to-use prompt

Template for this lesson

Copy and adapt to your context. Text in angle brackets should be replaced.

import google.generativeai as genai
import wikipedia, re

genai.configure(api_key=GOOGLE_API_KEY)
model = genai.GenerativeModel("gemini-2.0-flash")
chat = model.start_chat(history=[])

def search(query):
    try:
        return wikipedia.summary(query, sentences=3)
    except Exception as e:
        return f"Error: {e}"

# ReAct loop
response = chat.send_message(REACT_PROMPT + "\nQuestion\n" + "How tall is Eiffel Tower?")
while "<finish>" not in response.text:
    match = re.search(r"<search>(.*?)</search>", response.text)
    if match:
        obs = search(match.group(1))
        response = chat.send_message(f"Observation {obs}")

Common mistakes

What people get wrong

What is ReAct

ReAct is a prompting method that lets LLMs trace their thinking alongside concrete actions. The model generates Thought–Action–Observation triples on each iteration.

model_instructions = """Solve a question answering task with interleaving Thought, Action, Observation steps. Action can be of three types: (1) <search>entity</search> — search Wikipedia (2) <lookup>keyword</lookup> — lookup keyword in current context (3) <finish>answer</finish> — return answer and finish """

The ReAct Agent Class

import wikipedia import google.generativeai as genai class ReAct: def __init__(self, model: str, ReAct_prompt: str): self.model = genai.GenerativeModel(model) self.chat = self.model.start_chat(history=[]) self._search_history: list[str] = [] def search(self, query: str): """Performs Wikipedia search and returns summary.""" try: return wikipedia.summary(query, sentences=4) except Exception: return f"No page found for: {query}" def __call__(self, user_question: str): prompt = self._prompt + "\nQuestion\n" + user_question response = self.chat.send_message(prompt) return self._process(response.text)

Parsing Actions and the Execution Loop

import re def _process_response(text: str, agent: ReAct): action = re.search(r"<search>(.*?)</search>", text) if action: result = agent.search(action.group(1)) return agent.chat.send_message(f"Observation {result}") finish = re.search(r"<finish>(.*?)</finish>", text) if finish: return finish.group(1) return text react_agent = ReAct("gemini-2.0-flash", ReAct_prompt) answer = react_agent("How tall is the Eiffel Tower?")

Сообщить об ошибке

ReAct Agent: Reasoning + Acting

Квиз — 3 вопроса

What is ReAct

The ReAct Agent Class

Parsing Actions and the Execution Loop

Few-Shot Prompting for Format Control