The model’s response to a generate content request.

A struct representing a possible reply to a content generation prompt. Each content generation prompt may produce multiple candidate responses.

A collection of source attributions for a piece of content.

A struct describing a source attribution.

A value enumerating possible reasons for a model to terminate a content generation request.

A metadata struct containing any feedback the model had on the prompt it was provided.

Metadata returned to the client when grounding is enabled.

Represents a specific segment within a ModelContent struct, often used to pinpoint the exact location of text or data that grounding information refers to.

A struct defining model parameters to be used when sending generative AI requests to the backend model.

Configuration parameters for sending requests to the backend.

Represents token counting info for a single modality.

Content part modality.

A type describing data in media formats interpretable by an AI model. Each generative AI request or response contains an Array of ModelContents, and each ModelContent value may comprise multiple heterogeneous Parts.

A type defining potentially harmful media categories and their model-assigned ratings. A value of this type may be assigned to a category for every model-generated response, not just responses that exceed a certain threshold.

A type used to specify a threshold for harmful content, beyond which the model will return a fallback response instead of generated content.

See safety settings for Gemini models for more details.

Categories describing the potential harm a piece of content may pose.

Structured representation of a function declaration.

This FunctionDeclaration is a representation of a block of code that can be used as a Tool by the model and executed by the client.

A tool that allows the generative model to connect to Google Search to access and incorporate up-to-date information from the web into its responses.

A helper tool that the model may use when generating responses.

A Tool is a piece of code that enables the system to interact with external systems to perform an action, or set of actions, outside of knowledge and scope of the model.

Configuration for specifying function calling behavior.

Tool configuration for any Tool specified in the request.

The model’s response to a count tokens request.

Represents available backend APIs for the Firebase AI SDK.

An aspect ratio for images generated by Imagen.

To specify an aspect ratio for generated images, set aspectRatio in your ImagenGenerationConfig. See the Cloud documentation for more details and examples of the supported aspect ratios.

Configuration options for generating images with Imagen.

See Parameters for Imagen models to learn about parameters available for use with Imagen models, including how to configure them.

A response from a request to generate images with Imagen.

The type placeholder T is an image type; this is currently always an ImagenInlineImage.

This type is returned from:

• generateImages(prompt:) where T is ImagenInlineImage

An image format for images generated by Imagen.

To specify an image format for generated images, set imageFormat in your ImagenGenerationConfig. See the Cloud documentation for more details.

An error that occurs when image generation fails due to all generated images being blocked.

The images may have been blocked due to the specified ImagenSafetyFilterLevel, the ImagenPersonFilterLevel, or filtering included in the model. These filter levels may be adjusted in your ImagenSafetySettings. See the Responsible AI and usage guidelines for Imagen for more details.

An image generated by Imagen, represented as inline data.

A filter level controlling whether generation of images containing people or faces is allowed.

See the personGeneration documentation for more details.

A filter level controlling how aggressively to filter sensitive content.

Text prompts provided as inputs and images (generated or uploaded) through Imagen on Vertex AI are assessed against a list of safety filters, which include ‘harmful categories’ (for example, violence, sexual, derogatory, and toxic). This filter level controls how aggressively to filter out potentially harmful content from responses. See the safetySetting documentation and the Responsible AI and usage guidelines for more details.

Settings for controlling the aggressiveness of filtering out sensitive content.

See the Responsible AI and usage guidelines for more details.

Configuration options for audio transcriptions when communicating with a model that supports the Gemini Live API.

While there are not currently any options, this will likely change in the future. For now, just providing an instance of this struct will enable audio transcriptions for the corresponding input or output fields.

Text transcription of some audio form during a live interaction with the model.

Configuration options for live content generation.

Incremental server update generated by the model in response to client messages.

Content is generated as quickly as possible, and not in realtime. Clients may choose to buffer and play it out in realtime.

Server will not be able to service client soon.

To learn more about session limits, see the docs on Maximum session duration.

Update from the server, generated from the model in response to client messages.

Request for the client to execute the provided functionCalls.

The client should return matching FunctionResponsePart, where the functionId fields correspond to individual FunctionCallParts.

Notification for the client to cancel a previous function call from LiveServerToolCall.

The client does not need to send FunctionResponseParts for the cancelled FunctionCallParts.

The model sent a message that the SDK failed to parse.

This may indicate that the SDK version needs updating, a model is too old for the current SDK version, or that the model is just not supported.

Check the NSUnderlyingErrorKey entry in errorUserInfo for the error that caused this.

The live session was closed, because the network connection was lost.

Check the NSUnderlyingErrorKey entry in errorUserInfo for the error that caused this.

The live session was closed, but not for a reason the SDK expected.

Check the NSUnderlyingErrorKey entry in errorUserInfo for the error that caused this.

The model refused our request to setup a live session.

This can occur due to the model not supporting the requested response modalities, the project not having access to the model, the model being invalid, or some internal error.

Check the NSUnderlyingErrorKey entry in errorUserInfo for the error that caused this.

Configuration for controlling the voice of the model during conversation.

A text part containing a string value.

A data part that is provided inline in requests.

Data provided as an inline data part is encoded as base64 and included directly (inline) in the request. For large files, see FileDataPart which references content by URI instead of including the data in the request.

File data stored in Cloud Storage for Firebase, referenced by URI.

A predicted function call returned from the model.

Result output from a function call.

Contains a string representing the FunctionDeclaration.name and a structured JSON object containing any output from the function is used as context to the model. This should contain the result of a FunctionCallPart made based on model prediction.

A part containing code that was executed by the model.

The result of executing code.

Represents the different types, or modalities, of data that a model can produce as output.

To configure the desired output modalities for model requests, set the responseModalities parameter when initializing a GenerationConfig. See the multimodal responses documentation for more details.

Configuration for controlling the “thinking” behavior of compatible Gemini models.

Certain models, like Gemini 2.5 Flash and Pro, utilize a thinking process before generating a response. This allows them to reason through complex problems and plan a more coherent and accurate answer.

A tool that allows the model to execute code.

This tool can be used to solve complex problems, for example, by generating and executing Python code to solve a math problem.

Metadata related to the urlContext() tool.

Metadata for a single URL retrieved by the urlContext() tool.