Google GenAI (Gemini)

The Google GenAI plugin provides access to Google’s Gemini models, text embeddings, Imagen image generation, text-to-speech, and Veo video generation.

Installation

<dependency>
    <groupId>com.google.genkit</groupId>
    <artifactId>genkit-plugin-google-genai</artifactId>
    <version>1.0.0-SNAPSHOT</version>
</dependency>

Configuration

export GOOGLE_GENAI_API_KEY=your-api-key

Get an API key from Google AI Studio.

Vertex AI

To use Vertex AI instead of the Google AI Developer API:

export GOOGLE_GENAI_USE_VERTEXAI=true
export GOOGLE_CLOUD_PROJECT=my-project
export GOOGLE_CLOUD_LOCATION=us-central1  # optional, defaults to us-central1

Usage

import com.google.genkit.plugins.googlegenai.GoogleGenAIPlugin;

Genkit genkit = Genkit.builder()
    .plugin(GoogleGenAIPlugin.create(System.getenv("GOOGLE_GENAI_API_KEY")))
    .build();

ModelResponse response = genkit.generate(
    GenerateOptions.builder()
        .model("googleai/gemini-2.5-flash")
        .prompt("Tell me about AI")
        .build());

Embeddings

Generate text embeddings for RAG, semantic search, and similarity tasks:

import com.google.genkit.ai.EmbedResponse;
import com.google.genkit.ai.Document;

List<Document> documents = List.of(
    Document.fromText("Genkit is a framework for building AI apps"),
    Document.fromText("Firebase provides cloud services")
);

EmbedResponse response = genkit.embed("googleai/gemini-embedding-001", documents);

// Access embedding vectors
float[] vector = response.getEmbeddings().get(0).getValues();
// vector.length == 768

Embeddings are used automatically by vector store plugins (Firebase, Pinecone, pgvector, etc.) when you configure an embedder name. You can also use them directly for custom similarity search.

Embedding options

You can pass task-specific options to optimize embedding quality:

Map<String, Object> embedOptions = Map.of(
    "taskType", "RETRIEVAL_DOCUMENT",  // or "RETRIEVAL_QUERY", "SEMANTIC_SIMILARITY"
    "title", "Document title",
    "outputDimensionality", 256  // reduce dimensions if needed
);

Text-to-Speech (TTS)

Generate natural-sounding speech from text using Gemini TTS models (gemini-3.1-flash-tts-preview, gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts):

Map<String, Object> ttsOptions = Map.of("voiceName", "Zephyr");
GenerationConfig config = GenerationConfig.builder()
    .custom(ttsOptions)
    .build();

ModelResponse response = genkit.generate(
    GenerateOptions.builder()
        .model("googleai/gemini-2.5-flash-preview-tts")
        .prompt("Hello! Welcome to Genkit Java.")
        .config(config)
        .build());

// The response contains audio as a media part (WAV format, base64-encoded)
String audioDataUrl = response.getMessage().getParts().get(0).getMedia().getUrl();
// "data:audio/wav;base64,..."

The plugin automatically frames your prompt with a synthesis preamble so the TTS model treats it as a transcript to voice. Without this, TTS models reject “vague” prompts with a 400 error (Model tried to generate text, but it should only be used for TTS). To supply your own framing (e.g. "Say cheerfully: ..."), set a ttsInstruction custom option — a preamble string, or an empty string to send the prompt verbatim:

Map<String, Object> ttsOptions = Map.of("voiceName", "Zephyr", "ttsInstruction", "");

Saving audio to a file

String dataUrl = response.getMessage().getParts().get(0).getMedia().getUrl();
String base64Data = dataUrl.substring(dataUrl.indexOf(",") + 1);
byte[] audioBytes = Base64.getDecoder().decode(base64Data);
Files.write(Path.of("output.wav"), audioBytes);

Video Generation (Veo)

Generate videos from text prompts or images using Google’s Veo models:

Map<String, Object> veoOptions = Map.of(
    "numberOfVideos", 1,
    "durationSeconds", 8,
    "aspectRatio", "16:9",
    "timeoutMs", 600000  // 10 min — video generation can take a while
);
GenerationConfig config = GenerationConfig.builder()
    .custom(veoOptions)
    .build();

ModelResponse response = genkit.generate(
    GenerateOptions.builder()
        .model("googleai/veo-3.1-generate-preview")
        .prompt("A serene Japanese garden with cherry blossoms falling")
        .config(config)
        .build());

// The response contains video as a media part (base64-encoded)
String videoDataUrl = response.getMessage().getParts().get(0).getMedia().getUrl();

Note

Video generation is an asynchronous operation. The plugin automatically polls until the video is ready. This can take several minutes depending on the model and settings.

Image Generation (Imagen)

Generate images with Imagen:

Map<String, Object> imagenOptions = Map.of(
    "numberOfImages", 1,
    "aspectRatio", "1:1"
);
GenerationConfig config = GenerationConfig.builder()
    .custom(imagenOptions)
    .build();

ModelResponse response = genkit.generate(
    GenerateOptions.builder()
        .model("googleai/imagen-4.0-fast-generate-001")
        .prompt("A cat wearing a space suit")
        .config(config)
        .build());

Video generation and editing (Gemini Omni)

Gemini Omni (googleai/gemini-omni-flash-preview) generates and iteratively edits video through the Gemini Interactions API. It supports conversational editing — each turn can build on the previous result while preserving elements you did not mention.

// First turn — generate a video
ModelResponse first = genkit.generate(
    GenerateOptions.builder()
        .model("googleai/gemini-omni-flash-preview")
        .prompt("A marble rolling down a chain-reaction track")
        .build());

// The returned interaction id lets you continue editing
String interactionId = (String) first.getCustom().get("interactionId");

// Follow-up turn — edit the previous result
ModelResponse edited = genkit.generate(
    GenerateOptions.builder()
        .model("googleai/gemini-omni-flash-preview")
        .prompt("Brighten the background and add a slow push-in on the logo")
        .config(GenerationConfig.builder()
            .custom(Map.of("previousInteractionId", interactionId))
            .build())
        .build());

// The generated video is returned as a media part (data URL by default)
String videoUrl = edited.getMessage().getParts().get(0).getMedia().getUrl();

Config options: previousInteractionId (continue/edit a prior interaction), aspectRatio (e.g. "16:9"), duration (e.g. "10s"), delivery ("inline" (default, base64) | "uri"), task ("text_to_video" | "image_to_video"), thinkingLevel, and maxOutputTokens. Only the Gemini Developer API (API key) is supported — not Vertex AI. The Interactions API is in preview.

Sample

See the google-genai sample for complete examples of text generation, tool calling, embeddings, image generation, TTS, and video generation.