Function Calling with Gemini

Gemini supports function calling to connect AI responses with external tools and APIs.

Defining Functions

const response = await ai.models.generateContent({
  model: "gemini-2.5-flash",
  contents: "What's the weather in Tokyo?",
  config: {
    tools: [{
      functionDeclarations: [{
        name: "getWeather",
        description: "Get current weather for a location",
        parameters: {
          type: "object",
          properties: {
            location: { type: "string", description: "City name" },
            unit: { type: "string", enum: ["celsius", "fahrenheit"] }
          },
          required: ["location"]
        }
      }]
    }],
  },
});

Structured Output

Force Gemini to output valid JSON with a schema:

const response = await ai.models.generateContent({
  model: "gemini-2.5-flash",
  contents: "List 5 programming languages with their use cases",
  config: {
    responseMimeType: "application/json",
    responseSchema: {
      type: "array",
      items: {
        type: "object",
        properties: {
          name: { type: "string" },
          useCase: { type: "string" }
        },
        required: ["name", "useCase"]
      }
    },
  },
});

Advanced Patterns

• Parallel function calls — Gemini can request multiple functions at once
• Function call modes — AUTO, ANY, or NONE to control tool behavior
• Nested schemas — complex object hierarchies for structured output
• Enum constraints — restrict outputs to predefined values

Common Pitfall: Relying Solely on responseMimeType Without a Schema

Setting responseMimeType: "application/json" asks Gemini to produce valid JSON, but without a responseSchema the model decides the JSON structure itself — which varies between calls. Always provide a schema for predictable output:

// Unpredictable: model chooses structure
config: { responseMimeType: "application/json" }

// Predictable: model follows your schema exactly
config: {
  responseMimeType: "application/json",
  responseSchema: {
    type: "array",
    items: {
      type: "object",
      properties: {
        name: { type: "string" },
        useCase: { type: "string" },
      },
      required: ["name", "useCase"],
    },
  },
}

Best Practice: Handle the NONE Stop Reason

When Gemini cannot execute any of the provided functions (e.g., because none match the user's request), it may return with a finishReason of "STOP" and no tool calls. Always check whether the model produced tool calls before trying to process them:

const candidate = response.candidates?.[0];
const hasToolCalls = candidate?.content?.parts?.some(p => p.functionCall);
if (!hasToolCalls) {
  // Model answered directly or could not find an appropriate tool
  return candidate?.content?.parts?.[0]?.text ?? "No response";
}

Named Product Reference: Genkit by Firebase

Firebase Genkit is Google's open-source framework for building AI-powered apps with Gemini. It provides a plugin system for LLM providers, vector stores, and tracing — similar in concept to LangChain but with tighter Gemini and Firebase integration. Genkit's built-in flow tracing UI shows the inputs and outputs of each step, making it easier to debug complex function-calling pipelines.