repository·main·Indexed 22 days ago
https://github.com/google-gemini/deprecated-generative-ai-jsLegacy version of the Google AI JavaScript SDK for the Gemini API. This SDK provides interfaces for content generation, multi-turn chat sessions, embeddings, and token counting. It includes support for tools such as function calling, code execution, and Google Search retrieval, as well as file management via GoogleAIFileManager and context caching via GoogleAICacheManager. This repository is in limited maintenance and will reach end-of-life on November 30, 2025; users are encouraged to migrate to the unified Google Gen AI SDK.
The web sample includes two main demonstrations of the SDK's capabilities:
index.html.chat.html.GenerateContentResponse interface represents the individual response received from calling GenerativeModel.generateContent() or GenerativeModel.generateContentStream(). When using generateContentStream(), the API returns one GenerateContentResponse object for each chunk in the stream until the stream is complete.FunctionDeclarationsTool allows you to define functions the model can call. The model returns a FunctionCall part, which your application should execute, and then you provide the result back via a FunctionResponse part.The ChatSession class enables multi-turn conversations by sending chat messages and automatically storing the history of sent and received messages. This allows the model to maintain context across multiple interactions.
Key behaviors:
sendMessage) or a streaming method (sendMessageStream).SingleRequestOptions will take precedence over the RequestOptions originally provided to GoogleGenerativeAI.getGenerativeModel().You can enforce structured responses by providing a ResponseSchema. The Schema type is a union of several specific schema interfaces:
StringSchema: Includes SimpleStringSchema and EnumStringSchema (for fixed sets of strings).NumberSchema: For floating-point numbers.IntegerSchema: For integers (supports int32 or int64 formats).BooleanSchema: For boolean values.ArraySchema: For lists of items, with optional minItems and maxItems constraints.ObjectSchema: For complex objects with defined properties and required fields.Use the SchemaType enum to specify the type (e.g., SchemaType.OBJECT, SchemaType.STRING).
A Tool allows the Gemini model to access external knowledge or capabilities by defining specific interfaces it can invoke. The Tool type is a union that supports three distinct types of tools:
FunctionDeclarationsTool).CodeExecutionTool).GoogleSearchRetrievalTool).You provide these tools as part of the model configuration to enable the model to perform tasks beyond simple text generation.
export declare type Tool = FunctionDeclarationsTool | CodeExecutionTool | GoogleSearchRetrievalTool;The core data unit in the Gemini API is Content. A Content object consists of a role (e.g., 'user', 'model', 'system', 'function') and an array of Part objects.
Part is a union type that can represent:
TextPart: Plain text.InlineDataPart: Base64 encoded media (blobs).FunctionCallPart: A request from the model to call a function.FunctionResponsePart: The result of a function execution provided to the model.FileDataPart: Reference to a file via URI.ExecutableCodePart: Code to be executed by the model.CodeExecutionResultPart: The output of code execution.When calling generateContentStream(), the API returns a GenerateContentStreamResult object. This object provides two ways to access the generated content:
stream property, which is an AsyncGenerator. This allows you to process individual chunks of the response as they are being generated.response property, which is a Promise. This resolves to the full, aggregated EnhancedGenerateContentResponse once the entire stream has completed.You can use both simultaneously if you need to update a UI incrementally while also having access to the final complete object.
The Tool type allows you to extend model capabilities. A Tool can be one of the following:
FunctionDeclarationsTool): Define FunctionDeclaration objects containing a name, description, and parameters (using FunctionDeclarationSchema) to enable the model to call your local functions.CodeExecutionTool): Enables the model to write and run code (currently supports PYTHON).GoogleSearchRetrievalTool): Enables the model to use Google Search for information retrieval.The Part type is a union type used to represent different types of content within a request or response in the Gemini API. It allows you to compose multi-modal inputs (text, images, files) or handle structured data like function calls and code execution results.
Depending on your use case, a Part can be one of the following:
TextPart: Plain text content.InlineDataPart: Base64 encoded data (e.g., images).FunctionCallPart: A request from the model to call a specific function.FunctionResponsePart: The result of a function call provided back to the model.FileDataPart: Reference to a file already uploaded to the Gemini File API.ExecutableCodePart: Code intended for execution.CodeExecutionResultPart: The output/result of executed code.export type Part = TextPart | InlineDataPart | FunctionCallPart | FunctionResponsePart | FileDataPart | ExecutableCodePart | CodeExecutionResultPart;The StringSchema type is used to describe string-based properties within a JSON schema for structured output. It is a union type that allows you to define either a basic string or a string constrained to a specific set of allowed values (an enum).
To use it, you must choose between:
SimpleStringSchema: For general strings.EnumStringSchema: For strings that must match one of several predefined values.export type StringSchema = SimpleStringSchema | EnumStringSchema;The UsageMetadata interface provides information about the token usage for a specific generation request. It is used to track how many tokens were consumed by the prompt, the generated candidates, and any cached content, allowing you to monitor costs and rate limits.
export interface UsageMetadata