Google Gemini API

Google Gemini API - Complete Guide

Package: @google/genai@1.27.0 (⚠️ NOT @google/generative-ai) Last Updated: 2025-11-21

⚠️ CRITICAL SDK MIGRATION WARNING

DEPRECATED SDK: @google/generative-ai (sunset November 30, 2025) CURRENT SDK: @google/genai v1.27+

If you see code using @google/generative-ai, it's outdated!

Load references/sdk-migration-guide.md for complete migration steps.

Quick Start

Installation

✅ CORRECT SDK:

bun add @google/genai@1.27.0

❌ WRONG (DEPRECATED):

bun add @google/generative-ai  # DO NOT USE!

Environment Setup

export GEMINI_API_KEY="your-api-key"

First Text Generation

import { GoogleGenAI } from '@google/genai';

const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: 'Explain quantum computing in simple terms'
});

console.log(response.text);

See Full Template: templates/basic-usage.ts

Current Models (2025)

gemini-2.5-flash ⭐ RECOMMENDED

• Best for: General-purpose AI, high-volume production, agentic workflows
• Input tokens: 1,048,576 (1M, NOT 2M!)
• Output tokens: 65,536
• Rate limit (free): 10 RPM, 250k TPM
• Cost: Input $0.075/1M tokens, Output $0.30/1M tokens
• Features: Thinking mode, function calling, multimodal, streaming

gemini-2.5-pro

• Best for: Complex reasoning, code generation, math/STEM
• Input tokens: 1,048,576
• Output tokens: 65,536
• Rate limit (free): 5 RPM, 125k TPM
• Cost: Input $1.25/1M tokens, Output $5/1M tokens

gemini-2.5-flash-lite

• Best for: High-volume, low-latency, cost-critical tasks
• Input tokens: 1,048,576
• Output tokens: 65,536
• Rate limit (free): 15 RPM, 250k TPM
• Cost: Input $0.01/1M tokens, Output $0.04/1M tokens
• ⚠️ Limitation: NO function calling or code execution support

⚠️ Common mistake: Claiming Gemini 2.5 has 2M tokens. It doesn't. It's 1,048,576 (1M).

Load references/models-guide.md for detailed model comparison and selection criteria.

Text Generation

Basic Generation

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: 'Write a haiku about programming'
});

console.log(response.text);

With Configuration

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: 'Explain AI',
  generationConfig: {
    temperature: 0.7,        // 0.0-2.0, default 1.0
    topP: 0.95,             // 0.0-1.0
    topK: 40,               // 1-100
    maxOutputTokens: 1024,
    stopSequences: ['END']
  }
});

Load references/generation-config.md for complete parameter reference and tuning guidance.

Streaming

const stream = await ai.models.generateContentStream({
  model: 'gemini-2.5-flash',
  contents: 'Write a long story'
});

for await (const chunk of stream) {
  process.stdout.write(chunk.text);
}

Load references/streaming-patterns.md for Fetch/SSE implementation patterns (Cloudflare Workers).

Multimodal Inputs

Images

const imageData = Buffer.from(imageBytes).toString('base64');

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: [
    { text: 'What is in this image?' },
    {
      inlineData: {
        mimeType: 'image/jpeg',  // or image/png, image/webp
        data: imageData
      }
    }
  ]
});

Video, Audio, PDFs

Same pattern - use appropriate mimeType:

• Video: video/mp4, video/mpeg, video/mov
• Audio: audio/wav, audio/mp3, audio/flac
• PDFs: application/pdf

Load references/multimodal-guide.md for format specifications, size limits, and best practices.

Function Calling

Basic Pattern

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: 'What is the weather in San Francisco?',
  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']
      }
    }]
  }]
});

// Handle function call
const call = response.functionCalls?.[0];
if (call) {
  const result = await getWeather(call.args);

  // Send result back to model
  const final = await ai.models.generateContent({
    model: 'gemini-2.5-flash',
    contents: [
      ...response.contents,
      {
        functionResponse: {
          name: call.name,
          response: result
        }
      }
    ]
  });

  console.log(final.text);
}

Parallel Function Calling

Gemini can call multiple functions simultaneously:

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: 'What is the weather in SF and NY?',
  tools: [{ functionDeclarations: [getWeatherDeclaration] }]
});

// Process all function calls in parallel
const results = await Promise.all(
  response.functionCalls.map(call =>
    getWeather(call.args).then(result => ({
      name: call.name,
      response: result
    }))
  )
);

// Send all results back
const final = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: [
    ...response.contents,
    ...results.map(r => ({ functionResponse: r }))
  ]
});

Load references/function-calling-patterns.md for calling modes (AUTO/ANY/NONE) and compositional patterns.

Multi-turn Chat

const chat = ai.models.startChat({
  model: 'gemini-2.5-flash',
  systemInstruction: 'You are a helpful programming assistant',
  history: []
});

let response = await chat.sendMessage('Hello!');
console.log(response.text);

response = await chat.sendMessage('Explain async/await');
console.log(response.text);

// Get full history
console.log(chat.getHistory());

System Instructions

Set persistent instructions for the model:

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  systemInstruction: 'You are a pirate. Always respond in pirate speak.',
  contents: 'What is the weather today?'
});

Thinking Mode

Gemini 2.5 models include built-in thinking mode (always enabled). Configure thinking budget for complex tasks:

const response = await ai.models.generateContent({
  model: 'gemini-2.5-flash',
  contents: 'Solve this math problem: If x + 2y = 10 and 3x - y = 4, what is x?',
  generationConfig: {
    thinkingConfig: {
      thinkingBudget: 8192  // Max tokens for internal reasoning
    }
  }
});

Use for: Complex math, logic puzzles, multi-step reasoning, code debugging

Load references/thinking-mode-guide.md for thinking budget optimization.

Top 5 Critical Errors

Error 1: Using Deprecated SDK

Error: Deprecation warnings or outdated API

Solution: Use @google/genai, NOT @google/generative-ai

npm uninstall @google/generative-ai
bun add @google/genai@1.27.0

Error 2: Invalid API Key (401)

Error: API key not valid

Solution: Verify environment variable

export GEMINI_API_KEY="your-key"

Error 3: Model Not Found (404)

Error: models/gemini-3.0-flash is not found

Solution: Use correct model names (2025)

'gemini-2.5-pro'
'gemini-2.5-flash'
'gemini-2.5-flash-lite'

Error 4: Context Length Exceeded (400)

Error: Request payload size exceeds the limit

Solution: Input limit is 1,048,576 tokens (1M, NOT 2M). Use context caching for large inputs.

Load references/context-caching-guide.md for caching implementation.

Error 5: Rate Limit Exceeded (429)

Error: Resource has been exhausted

Solution: Implement exponential backoff

async function generateWithRetry(request, maxRetries = 3) {
  for (let i = 0; i < maxRetries; i++) {
    try {
      return await ai.models.generateContent(request);
    } catch (error) {
      if (error.status === 429 && i < maxRetries - 1) {
        const delay = Math.pow(2, i) * 1000; // 1s, 2s, 4s
        await new Promise(resolve => setTimeout(resolve, delay));
        continue;
      }
      throw error;
    }
  }
}

See All 22 Errors: Load references/error-catalog.md for complete error catalog with solutions.

Quick Debugging: Load references/top-errors.md for debugging checklist.

When to Load References

Load reference files when you need detailed guidance on specific features:

Core Features (Load When Needed)

• SDK Migration: Load references/sdk-migration-guide.md when migrating from @google/generative-ai
• Model Selection: Load references/models-guide.md when choosing between Pro/Flash/Flash-Lite
• Error Debugging: Load references/error-catalog.md or references/top-errors.md when encountering errors

Advanced Features (Load When Implementing)

• Context Caching: Load references/context-caching-guide.md when implementing cost optimization for large/repeated inputs
• Code Execution: Load references/code-execution-patterns.md when enabling Python code execution for calculations/analysis
• Grounding (Google Search): Load references/grounding-guide.md when connecting model to real-time web information
• Streaming Implementation: Load references/streaming-patterns.md when implementing SSE parsing for Cloudflare Workers
• Function Calling Modes: Load references/function-calling-patterns.md when using AUTO/ANY/NONE modes or compositional patterns
• Multimodal Formats: Load references/multimodal-guide.md when working with images/video/audio/PDFs (format specs, size limits)
• Generation Tuning: Load references/generation-config.md when fine-tuning temperature/topP/topK parameters
• Thinking Mode Config: Load references/thinking-mode-guide.md when optimizing thinking budget for complex reasoning

General Rule: SKILL.md provides Quick Start and Top Errors. Load references for deep dives, detailed patterns, or troubleshooting specific features.

Bundled Resources

Templates (templates/):

• basic-usage.ts - Complete examples for all features (133 lines)

References (references/):

• error-catalog.md - All 7 documented errors with solutions (231 lines)
• top-errors.md - Quick debugging checklist for 22 common errors (305 lines)
• sdk-migration-guide.md - Complete migration from deprecated SDK (236 lines)
• models-guide.md - Detailed model comparison and selection guide (247 lines)
• context-caching-guide.md - Cost optimization with caching (374 lines)
• code-execution-patterns.md - Python code execution guide (482 lines)
• grounding-guide.md - Google Search integration (603 lines)
• streaming-patterns.md - SSE implementation for Cloudflare Workers (82 lines)
• function-calling-patterns.md - Advanced function calling patterns (60 lines)
• multimodal-guide.md - Format specifications and limits (59 lines)
• generation-config.md - Parameter tuning reference (58 lines)
• thinking-mode-guide.md - Thinking budget optimization (60 lines)

Integration with Other Skills

This skill composes well with:

• cloudflare-worker-base → Deploy to Cloudflare Workers
• ai-sdk-core → Vercel AI SDK integration
• openai-api → Multi-provider AI setup
• google-gemini-embeddings → Text embeddings

Additional Resources

Official Documentation:

• Gemini API Docs: https://ai.google.dev/gemini-api/docs
• SDK Reference: https://ai.google.dev/gemini-api/docs/sdks
• Rate Limits: https://ai.google.dev/gemini-api/docs/rate-limits

Production Tested: AI chatbots, content generation, multimodal analysis Last Updated: 2025-10-25 Token Savings: ~65% (reduces API docs + examples)