Abstract:

This document is a technical migration guide for developers transitioning from the legacy google-generativeai library to the new, streamlined Google GenAI SDK (google-genai). The guide highlights the core architectural change: the introduction of a centralized Client object that serves as the single entry point for all API interactions, replacing previous ad-hoc methods. It provides comprehensive "before-and-after" code examples in Python, demonstrating key changes across various functionalities. These include simplified authentication, a more consistent structure for content generation, configuration, and safety settings. The guide also details updated workflows for advanced features like chat sessions, function calling, code execution, file management, context caching, and model tuning, emphasizing an improved and more uniform developer experience.

Migrating to the New Google GenAI SDK: Key Changes and Updates

• (0:00) Core Architectural Shift: The most significant change is the move from interacting with individual model objects to using a central genai.Client() object. This client now serves as the single entry point for all API services like models, chats, files, and tuning.
• (0:00) Simplified Installation & Authentication: The new SDK is installed with pip install google-genai. Authentication is streamlined, as the client can automatically use an API key from the GEMINI_API_KEY environment variable, or it can be passed explicitly during client initialization.
• (0:00) Consistent API Calls: Previously separate functions like genai.upload_file or model.generate_content are now unified as methods on the client object (e.g., client.files.upload(), client.models.generate_content()). This creates a more predictable and uniform structure.
• (0:00) Centralized Configuration: Optional parameters for generation (like temperature, max_output_tokens, safety_settings, and response_mime_type) are now passed together in a single config argument, which can be a dictionary or a types.GenerateContentConfig object.
• (0:00) Updated Chat and State Management: Chat sessions are now created via client.chats.create(), making the process more explicit and separating it from the model definition.
• (0:00) Enhanced Tools and Function Calling: Automatic function calling is now the default behavior in generate_content. Tools such as code_execution and google_search are enabled within the config argument, providing a unified way to add capabilities to the model.
• (0:00) Improved File and Cache Management: The entire lifecycle for files (upload, list, get, delete) and context caches is now managed through dedicated services on the client object (client.files and client.caches), improving clarity over the previous global functions.
• (0:00) Streamlined Model Tuning: The new SDK simplifies model tuning with the client.tunings.tune() method, which launches the tuning job and polls for completion, reducing the manual steps required previously.