Execução de código

A execução de código é uma ferramenta que permite que o modelo gere e execute código em Python. O modelo pode aprender de forma iterativa com os resultados da execução de código até chegar a uma saída final.

Você pode usar a execução de código para criar recursos que se beneficiam do raciocínio baseado em código e que geram saída de texto. Por exemplo, é possível usar a execução de código para resolver equações ou processar texto. Você também pode usar as bibliotecas incluídas no ambiente de execução de código para realizar tarefas mais especializadas.

Assim como acontece com todas as ferramentas fornecidas ao modelo, ele decide quando usar a execução de código.

Ir para a implementação do código

Comparação entre execução de código e chamada de função

A execução de código e a chamada de função são recursos semelhantes. Em geral, é melhor usar a execução de código se o modelo puder lidar com seu caso de uso. A execução de código também é mais simples de usar porque basta ativá-la.

Confira algumas outras diferenças entre a execução de código e a chamada de função:

Execução de código Chamada de função
Use a execução de código se quiser que o modelo escreva e execute código em Python para você e retorne o resultado. Use a chamada de função se você já tiver as suas próprias funções que você quer executar localmente.
A execução de código permite que o modelo execute código no back-end da API em um ambiente fixo, isolado. A chamada de função permite executar as funções solicitadas pelo modelo, no ambiente que você quiser.
A execução de código é resolvida em uma única solicitação. Embora você possa usar a execução de código com o recurso de chat, não há nenhum requisito. A chamada de função exige uma solicitação extra para enviar a saída de cada chamada de função. Portanto, é necessário usar o recurso de chat capacidade.

Modelos compatíveis

  • gemini-3.1-pro-preview
  • gemini-3-flash-preview
  • gemini-3.1-flash-lite-preview
  • gemini-2.5-pro
  • gemini-2.5-flash
  • gemini-2.5-flash-lite

Usar a execução de código

Você pode usar a execução de código com entrada somente de texto e multimodal, mas a resposta será sempre apenas texto ou código.

Antes de começar

Clique no seu provedor Gemini API para conferir o conteúdo específico do provedor e o código nesta página.

Se ainda não fez isso, conclua o guia de iniciação, que descreve como configurar seu projeto do Firebase, conectar seu app ao Firebase, adicionar o SDK, inicializar o serviço de back-end para o provedor Gemini API escolhido e criar uma instância de GenerativeModel.

Para testar e iterar nos comandos, recomendamos o uso do Google AI Studio.

Ativar a execução de código

Antes de testar este exemplo, conclua a seção Antes de começar deste guia para configurar seu projeto e app.
Nessa seção, você também vai clicar em um botão para o provedor escolhido para que o conteúdo específico do provedor seja exibido nesta página.
Gemini API

Ao criar a instância GenerativeModel, forneça CodeExecution como uma ferramenta que o modelo pode usar para gerar a resposta. Isso permite que o modelo gere e execute código em Python.

Swift


import FirebaseAILogic

// Initialize the Gemini Developer API backend service
let ai = FirebaseAI.firebaseAI(backend: .googleAI())

// Create a `GenerativeModel` instance with a model that supports your use case
let model = ai.generativeModel(
  modelName: "GEMINI_MODEL_NAME",
  // Provide code execution as a tool that the model can use to generate its response.
  tools: [.codeExecution()]
)

let prompt = """
What is the sum of the first 50 prime numbers?
Generate and run code for the calculation, and make sure you get all 50.
"""

let response = try await model.generateContent(prompt)

guard let candidate = response.candidates.first else {
  print("No candidates in response.")
  return
}
for part in candidate.content.parts {
  if let textPart = part as? TextPart {
    print("Text = \(textPart.text)")
  } else if let executableCode = part as? ExecutableCodePart {
    print("Code = \(executableCode.code), Language = \(executableCode.language)")
  } else if let executionResult = part as? CodeExecutionResultPart {
    print("Outcome = \(executionResult.outcome), Result = \(executionResult.output ?? "no output")")
  }
}

Kotlin


// Initialize the Gemini Developer API backend service
// Create a `GenerativeModel` instance with a model that supports your use case
val model = Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(
    modelName = "GEMINI_MODEL_NAME",
    // Provide code execution as a tool that the model can use to generate its response.
    tools = listOf(Tool.codeExecution())
)

val prompt =  "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50."

val response = model.generateContent(prompt)

response.candidates.first().content.parts.forEach {
    if(it is TextPart) {
        println("Text = ${it.text}")
    }
    if(it is ExecutableCodePart) {
        println("Code = ${it.code}, Language = ${it.language}")
    }
    if(it is CodeExecutionResultPart) {
       println("Outcome = ${it.outcome}, Result = ${it.output}")
    }
}

Java


// Initialize the Gemini Developer API backend service
// Create a `GenerativeModel` instance with a model that supports your use case
GenerativeModel ai = FirebaseAI.getInstance(GenerativeBackend.googleAI())
                .generativeModel("GEMINI_MODEL_NAME",
                        null,
                        null,
                        // Provide code execution as a tool that the model can use to generate its response.
                        List.of(Tool.codeExecution()));

// Use the GenerativeModelFutures Java compatibility layer which offers
// support for ListenableFuture and Publisher APIs
GenerativeModelFutures model = GenerativeModelFutures.from(ai);

String text = "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50.";

Content prompt = new Content.Builder()
    .addText(text)
    .build();

ListenableFuture response = model.generateContent(prompt);

Futures.addCallback(response, new FutureCallback() {
   @Override
public void onSuccess(GenerateContentResponse response)   {
    // Access the first candidate's content parts
    List parts = response.getCandidates().get(0).getContent().getParts();
    for (Part part : parts) {
        if (part instanceof TextPart) {
            TextPart textPart = (TextPart) part;
            System.out.println("Text = " + textPart.getText());
        } else if (part instanceof ExecutableCodePart) {
            ExecutableCodePart codePart = (ExecutableCodePart) part;
            System.out.println("Code = " + codePart.getCode() + ", Language = " + codePart.getLanguage());
        } else if (part instanceof CodeExecutionResultPart) {
            CodeExecutionResultPart resultPart = (CodeExecutionResultPart) part;
            System.out.println("Outcome = " + resultPart.getOutcome() + ", Result = " + resultPart.getOutput());
        }
    }
}

    @Override
    public void onFailure(Throwable t) {
        t.printStackTrace();
    }
}, executor);

Web


import { initializeApp } from "firebase/app";
import { getAI, getGenerativeModel, GoogleAIBackend } from "firebase/ai";

// TODO(developer) Replace the following with your app's Firebase configuration
// See: https://firebase.google.com/docs/web/learn-more#config-object
const firebaseConfig = {
  // ...
};

// Initialize FirebaseApp
const firebaseApp = initializeApp(firebaseConfig);

// Initialize the Gemini Developer API backend service
const ai = getAI(firebaseApp, { backend: new GoogleAIBackend() });

// Create a `GenerativeModel` instance with a model that supports your use case
const model = getGenerativeModel(
  ai,
  {
    model: "GEMINI_MODEL_NAME",
    // Provide code execution as a tool that the model can use to generate its response.
    tools: [{ codeExecution: {} }]
  }
);

const prompt =  "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50."

const result = await model.generateContent(prompt);
const response = await result.response;

const parts = response.candidates?.[0].content.parts;

if (parts) {
  parts.forEach((part) => {
    if (part.text) {
        console.log(`Text: ${part.text}`);
    } else if (part.executableCode) {
      console.log(
        `Code: ${part.executableCode.code}, Language: ${part.executableCode.language}`
      );
    } else if (part.codeExecutionResult) {
      console.log(
        `Outcome: ${part.codeExecutionResult.outcome}, Result: ${part.codeExecutionResult.output}`
      );
    }
  });
}

Dart


import 'package:firebase_core/firebase_core.dart';
import 'package:firebase_ai/firebase_ai.dart';
import 'firebase_options.dart';

// Initialize FirebaseApp
await Firebase.initializeApp(
  options: DefaultFirebaseOptions.currentPlatform,
);

// Initialize the Gemini Developer API backend service
// Create a `GenerativeModel` instance with a model that supports your use case
final model = FirebaseAI.googleAI().generativeModel(
  model: 'GEMINI_MODEL_NAME',
  // Provide code execution as a tool that the model can use to generate its response.
  tools: [
    Tool.codeExecution(),
  ],
);

const prompt = 'What is the sum of the first 50 prime numbers? '
    'Generate and run code for the calculation, and make sure you get all 50.';

final response = await model.generateContent([Content.text(prompt)]);

final buffer = StringBuffer();
    for (final part in response.candidates.first.content.parts) {
      if (part is TextPart) {
        buffer.writeln(part.text);
      } else if (part is ExecutableCodePart) {
        buffer.writeln('Executable Code:');
        buffer.writeln('Language: ${part.language}');
        buffer.writeln('Code:');
        buffer.writeln(part.code);
      } else if (part is CodeExecutionResultPart) {
        buffer.writeln('Code Execution Result:');
        buffer.writeln('Outcome: ${part.outcome}');
        buffer.writeln('Output:');
        buffer.writeln(part.output);
      }
    }

Unity


using Firebase;
using Firebase.AI;

// Initialize the Gemini Developer API backend service
var ai = FirebaseAI.GetInstance(FirebaseAI.Backend.GoogleAI());

// Create a `GenerativeModel` instance with a model that supports your use case
var model = ai.GetGenerativeModel(
  modelName: "GEMINI_MODEL_NAME",
  // Provide code execution as a tool that the model can use to generate its response.
  tools: new Tool[] { new Tool(new CodeExecution()) }
);

var prompt = "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50.";

var response = await model.GenerateContentAsync(prompt);

foreach (var part in response.Candidates.First().Content.Parts) {
  if (part is ModelContent.TextPart tp) {
    UnityEngine.Debug.Log($"Text = {tp.Text}");
  } else if (part is ModelContent.ExecutableCodePart esp) {
    UnityEngine.Debug.Log($"Code = {esp.Code}, Language = {esp.Language}");
  } else if (part is ModelContent.CodeExecutionResultPart cerp) {
    UnityEngine.Debug.Log($"Outcome = {cerp.Outcome}, Output = {cerp.Output}");
  }
}

Saiba como escolher um modelo adequado para seu caso de uso e app.

Usar a execução de código no chat

Você também pode usar a execução de código como parte de um chat:

Swift


import FirebaseAILogic

// Initialize the Gemini Developer API backend service
let ai = FirebaseAI.firebaseAI(backend: .googleAI())

// Create a `GenerativeModel` instance with a model that supports your use case
let model = ai.generativeModel(
  modelName: "GEMINI_MODEL_NAME",
  // Provide code execution as a tool that the model can use to generate its response.
  tools: [.codeExecution()]
)

let prompt = """
What is the sum of the first 50 prime numbers?
Generate and run code for the calculation, and make sure you get all 50.
"""
let chat = model.startChat()

let response = try await chat.sendMessage(prompt)

guard let candidate = response.candidates.first else {
  print("No candidates in response.")
  return
}
for part in candidate.content.parts {
  if let textPart = part as? TextPart {
    print("Text = \(textPart.text)")
  } else if let executableCode = part as? ExecutableCodePart {
    print("Code = \(executableCode.code), Language = \(executableCode.language)")
  } else if let executionResult = part as? CodeExecutionResultPart {
    print("Outcome = \(executionResult.outcome), Result = \(executionResult.output ?? "no output")")
  }
}

Kotlin


// Initialize the Gemini Developer API backend service
// Create a `GenerativeModel` instance with a model that supports your use case
val model = Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(
    modelName = "GEMINI_MODEL_NAME",
    // Provide code execution as a tool that the model can use to generate its response.
    tools = listOf(Tool.codeExecution())
)

val prompt =  "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50."
val chat = model.startChat()
val response = chat.sendMessage(prompt)

response.candidates.first().content.parts.forEach {
    if(it is TextPart) {
        println("Text = ${it.text}")
    }
    if(it is ExecutableCodePart) {
        println("Code = ${it.code}, Language = ${it.language}")
    }
    if(it is CodeExecutionResultPart) {
       println("Outcome = ${it.outcome}, Result = ${it.output}")
    }
}

Java


// Initialize the Gemini Developer API backend service
// Create a `GenerativeModel` instance with a model that supports your use case
GenerativeModel ai = FirebaseAI.getInstance(GenerativeBackend.googleAI())
                .generativeModel("GEMINI_MODEL_NAME",
                        null,
                        null,
                        // Provide code execution as a tool that the model can use to generate its response.
                        List.of(Tool.codeExecution()));

// Use the GenerativeModelFutures Java compatibility layer which offers
// support for ListenableFuture and Publisher APIs
GenerativeModelFutures model = GenerativeModelFutures.from(ai);
String text = "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50.";

Content prompt = new Content.Builder()
    .addText(text)
    .build();

ChatFutures chat = model.startChat();
ListenableFuture response = chat.sendMessage(prompt);

Futures.addCallback(response, new FutureCallback() {
   @Override
public void onSuccess(GenerateContentResponse response)   {
    // Access the first candidate's content parts
    List parts = response.getCandidates().get(0).getContent().getParts();
    for (Part part : parts) {
        if (part instanceof TextPart) {
            TextPart textPart = (TextPart) part;
            System.out.println("Text = " + textPart.getText());
        } else if (part instanceof ExecutableCodePart) {
            ExecutableCodePart codePart = (ExecutableCodePart) part;
            System.out.println("Code = " + codePart.getCode() + ", Language = " + codePart.getLanguage());
        } else if (part instanceof CodeExecutionResultPart) {
            CodeExecutionResultPart resultPart = (CodeExecutionResultPart) part;
            System.out.println("Outcome = " + resultPart.getOutcome() + ", Result = " + resultPart.getOutput());
        }
    }
}

    @Override
    public void onFailure(Throwable t) {
        t.printStackTrace();
    }
}, executor);

Web


import { initializeApp } from "firebase/app";
import { getAI, getGenerativeModel, GoogleAIBackend } from "firebase/ai";

// TODO(developer) Replace the following with your app's Firebase configuration
// See: https://firebase.google.com/docs/web/learn-more#config-object
const firebaseConfig = {
  // ...
};

// Initialize FirebaseApp
const firebaseApp = initializeApp(firebaseConfig);

// Initialize the Gemini Developer API backend service
const ai = getAI(firebaseApp, { backend: new GoogleAIBackend() });

// Create a `GenerativeModel` instance with a model that supports your use case
const model = getGenerativeModel(
  ai,
  {
    model: "GEMINI_MODEL_NAME",
    // Provide code execution as a tool that the model can use to generate its response.
    tools: [{ codeExecution: {} }]
  }
);

const prompt =  "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50."
const chat = model.startChat()
const result = await chat.sendMessage(prompt);

const parts = result.response.candidates?.[0].content.parts;

if (parts) {
  parts.forEach((part) => {
    if (part.text) {
        console.log(`Text: ${part.text}`);
    } else if (part.executableCode) {
      console.log(
        `Code: ${part.executableCode.code}, Language: ${part.executableCode.language}`
      );
    } else if (part.codeExecutionResult) {
      console.log(
        `Outcome: ${part.codeExecutionResult.outcome}, Result: ${part.codeExecutionResult.output}`
      );
    }
  });
}

Dart


import 'package:firebase_core/firebase_core.dart';
import 'package:firebase_ai/firebase_ai.dart';
import 'firebase_options.dart';

// Initialize FirebaseApp
await Firebase.initializeApp(
  options: DefaultFirebaseOptions.currentPlatform,
);

// Initialize the Gemini Developer API backend service
// Create a `GenerativeModel` instance with a model that supports your use case
final model = FirebaseAI.googleAI().generativeModel(
  model: 'GEMINI_MODEL_NAME',
  // Provide code execution as a tool that the model can use to generate its response.
  tools: [
    Tool.codeExecution(),
  ],
);

final codeExecutionChat = await model.startChat();

const prompt = 'What is the sum of the first 50 prime numbers? '
    'Generate and run code for the calculation, and make sure you get all 50.';
final response = await codeExecutionChat.sendMessage(Content.text(prompt));

final buffer = StringBuffer();
    for (final part in response.candidates.first.content.parts) {
      if (part is TextPart) {
        buffer.writeln(part.text);
      } else if (part is ExecutableCodePart) {
        buffer.writeln('Executable Code:');
        buffer.writeln('Language: ${part.language}');
        buffer.writeln('Code:');
        buffer.writeln(part.code);
      } else if (part is CodeExecutionResultPart) {
        buffer.writeln('Code Execution Result:');
        buffer.writeln('Outcome: ${part.outcome}');
        buffer.writeln('Output:');
        buffer.writeln(part.output);
      }
    }

Unity


using Firebase;
using Firebase.AI;

// Initialize the Gemini Developer API backend service
var ai = FirebaseAI.GetInstance(FirebaseAI.Backend.GoogleAI());

// Create a `GenerativeModel` instance with a model that supports your use case
var model = ai.GetGenerativeModel(
  modelName: "GEMINI_MODEL_NAME",
  // Provide code execution as a tool that the model can use to generate its response.
  tools: new Tool[] { new Tool(new CodeExecution()) }
);

var prompt = "What is the sum of the first 50 prime numbers? " +
        "Generate and run code for the calculation, and make sure you get all 50.";
var chat = model.StartChat();
var response = await chat.SendMessageAsync(prompt);

foreach (var part in response.Candidates.First().Content.Parts) {
  if (part is ModelContent.TextPart tp) {
    UnityEngine.Debug.Log($"Text = {tp.Text}");
  } else if (part is ModelContent.ExecutableCodePart esp) {
    UnityEngine.Debug.Log($"Code = {esp.Code}, Language = {esp.Language}");
  } else if (part is ModelContent.CodeExecutionResultPart cerp) {
    UnityEngine.Debug.Log($"Outcome = {cerp.Outcome}, Output = {cerp.Output}");
  }
}

Saiba como escolher um modelo adequado para seu caso de uso e app.

Preços

Não há cobrança extra para ativar a execução de código e fornecê-la como uma ferramenta para o modelo. Se o modelo decidir usar a execução de código, você será cobrado pela taxa atual de tokens de entrada e saída com base no modelo do Gemini que estiver usando.

O diagrama a seguir mostra o modelo de faturamento para execução de código:

Diagrama mostrando como os tokens são faturados quando um modelo usa a execução de código. 

Confira um resumo de como os tokens são faturados quando um modelo usa a execução de código:

  • O comando original é faturado uma vez. Os tokens são rotulados como intermediários, que são faturados como tokens de entrada.

  • O código gerado e o resultado do código executado são faturados da seguinte maneira:

    • Quando são usados durante a execução de código, eles são rotulados como tokens intermediários , que são faturados como tokens de entrada.

    • Quando são incluídos como parte da resposta final, eles são faturados como tokens de saída.

  • O resumo final na resposta final é faturado como tokens de saída.

O Gemini API inclui uma contagem de tokens intermediários na resposta da API. Assim, você sabe por que está sendo cobrado por tokens de entrada além do comando inicial.

Observe que o código gerado pode incluir texto e saídas multimodais, como imagens.

Limitações e práticas recomendadas

  • O modelo só pode gerar e executar código em Python. Ele não pode retornar outros artefatos, como arquivos de mídia.

  • A execução de código pode ser executada por no máximo 30 segundos antes do tempo limite.

  • Em alguns casos, a ativação da execução de código pode levar a regressões em outras áreas da saída do modelo (por exemplo, escrever uma história).

  • A ferramenta de execução de código não aceita URIs de arquivo como entrada/saída. No entanto, a ferramenta de execução de código aceita entrada de arquivo e saída de gráfico como bytes inline. Ao usar esses recursos de entrada e saída, você pode fazer upload de arquivos CSV e de texto, fazer perguntas sobre os arquivos e gerar gráficos do Matplotlib como parte do resultado da execução de código. Os tipos MIME aceitos para bytes inline são .cpp, .csv, .java, .jpeg, .js, .png, .py, .ts e .xml.

Bibliotecas permitidas

O ambiente de execução de código inclui as seguintes bibliotecas. Não é possível instalar suas próprias bibliotecas.


Enviar feedback sobre sua experiência com Firebase AI Logic