google-gemini-api

What it does

The skill provides a comprehensive guide for using Google Gemini API with the correct current SDK (@google/genai v1.27+). It covers text generation (including streaming), multimodal inputs (images, video, audio, PDFs), function calling (including parallel execution), thinking mode, system instructions, and multi-turn chat. It documents current Gemini 2.5 models (Pro, Flash, Flash-Lite) with their context windows and features, and includes a migration emphasis away from the deprecated SDK. It also outlines how to run both Node.js SDK and fetch-based approaches, and demonstrates setup for Cloudflare Workers, as well as hosting considerations and common pitfalls like context window errors. Commands and code samples use the exact endpoints and method names described in the guidance, and it specifies required API keys and model endpoints for generation and streaming.

How it works

• Installation and environment setup require the CORRECT SDK: npm install @google/genai@1.27.0 and exporting GEMINI_API_KEY.
• Basic text generation is shown both with the Node.js SDK (import { GoogleGenAI } from '@google/genai') and with fetch, using model: 'gemini-2.5-flash' and the appropriate generateContent endpoint; streaming is demonstrated via generateContentStream (SDK) and streamGenerateContent (fetch).
• Multimodal inputs include images, video, audio, and PDFs, each supported via contents with parts including inlineData (base64) and mimeType; specific formats and size limits are enumerated (e.g., images up to 20MB, videos up to 2 minutes inline, PDFs up to 30MB).
• Function calling is supported with a sample function declaration (name, description, parametersJsonSchema) and a request configuring tools to enable function execution; it shows how to capture the functionCall and then return results to the model.
• Thinking mode is highlighted as the default-on feature across Gemini-2.5 variants, with parallel and compositional function calling supported.
• The guide distinguishes Node.js SDK benefits and fetch-based approaches for environments like Cloudflare Workers, browsers, and edge runtimes, including tradeoffs like SSE parsing and error handling.
• Phase 1 and Phase 2 features are listed (e.g., context caching, code execution, grounding with Google Search) and the context window size is clarified as 1,048,576 tokens for Gemini 2.5 models, not 2M.
• It includes model-specific capabilities (Pro, Flash, Flash-Lite) and confirms their shared features (thinking mode, function calling, multimodal, streaming, system instructions) and token limits.

When to use it

Use when integrating Gemini API, implementing multimodal AI applications, using thinking mode for complex reasoning, function calling with parallel execution, streaming responses, deploying to Cloudflare Workers, building chat applications, or encountering SDK deprecation warnings, context window errors, model not found errors, function calling failures, or multimodal format errors.

What it can touch

• Tools: Function calling with declared tools (example includes a get_current_weather function). The skill demonstrates using tools via the SDK by attaching functionDeclarations and then handling the functionCall in the response.

Caveats

• License: MIT
• It repeatedly cautions to use the correct current SDK (@google/genai) and warns against deprecated @google/generative-ai; migration guidance is included.
• It specifies precise context window and token limits for Gemini 2.5 models (1,048,576 input tokens; 65,536 output tokens) and notes that 2M-token context is incorrect for these models.
• It notes that PDFs and multimodal inputs have size limits and format requirements (image, video, audio, PDF).