Generating responses with Gemini API (gemini-2.5-flash-lite)

Structure

LLM Usage

We are using gemini-2.5-flash-lite to generate two types of responses:

1. Human readable text
2. JSON

Classes

The handling is structured into multiple different classes such as:

1. GoogleAIStudioLLM
2. NetworkClient
3. StreamParser
4. StreamProzessor

Other required objects are:

1. RequestBuilder
2. SpeechManager

GoogleAIStudioLLM.kt

This class features the two most important functions:

1. generate(command: String, structured: Boolean) which returns a String
2. generateStream(command: String, onChunk: (String) -> Unit, onComplete: () -> Unit, onError: (Exception) -> Unit)

The generate function is crucial for JSON communication as its Boolean "structured" argument allows us to switch between requests in JSON and human readable text.

NetworkClient.kt

Creates a connection to the Gemini API and handles unspecific errors and requests.

RequestBuilder.kt

The general request body which is used for any LLM communication is definied here.

SpeechManager.kt

Allows to force stop streamed text.

Used when the user doubletaps the button on EyeAIVision or clicks the Speechrecognition Button in the App twice. This interrupts any spoken text instantly.

StreamParser.kt

Parses every chunk of the stream.

StreamProcessor.kt

Buffers every chunk before passing it to the parser