Anvia Docs

@anvia/gemini adapts the Google GenAI SDK into Anvia's provider-neutral interfaces. It supports completions, embeddings, image generation (both native and Imagen), and transcription.

Install

pnpm add @anvia/gemini

The Google GenAI SDK (@google/genai) is a transitive dependency.

Quick Start

import { AgentBuilder } from "@anvia/core";
import { GeminiClient } from "@anvia/gemini";

const client = new GeminiClient({ apiKey });
const model = client.completionModel("gemini-2.5-flash");

const agent = new AgentBuilder("support", model)
  .instructions("Answer clearly.")
  .build();

const response = await agent.prompt("What is Anvia?").send();
console.log(response.output);

Client Configuration

// Google AI API
const client = new GeminiClient({ apiKey });

// Vertex AI
const client = new GeminiClient({
  vertexai: true,
  project: "my-gcp-project",
  location: "us-central1",
});

// Pre-configured SDK instance
const client = new GeminiClient({ client: existingGoogleGenAIInstance });

Option	Purpose
`apiKey`	Google AI API key (required for Google AI)
`vertexai`	Set to `true` for Vertex AI
`project`	GCP project ID (required for Vertex AI)
`location`	GCP region (required for Vertex AI)
`client`	Pre-configured `GoogleGenAI` instance

Provider clients only use explicit constructor values. They do not read environment variables.

Completion Models

// Default: gemini-2.5-flash
const model = client.completionModel();

// Specific model
const model = client.completionModel("gemini-3-pro");

Supports streaming and non-streaming through the standard StreamingCompletionModel interface.

Embedding Models

const embeddingModel = client.embeddingModel("gemini-embedding-001");

const embeddings = await embeddingModel.embedTexts(["hello", "world"]);

Option	Default	Purpose
`taskType`	--	Embedding task type (e.g., `"RETRIEVAL_QUERY"`, `"RETRIEVAL_DOCUMENT"`)
`maxBatchSize`	model default	Texts per API call

Task Types

type GeminiEmbeddingTaskType =
  | "RETRIEVAL_QUERY"
  | "RETRIEVAL_DOCUMENT"
  | "SEMANTIC_SIMILARITY"
  | "CLASSIFICATION"
  | "CLUSTERING"
  | "QUESTION_ANSWERING"
  | "FACT_VERIFICATION";

Image Generation

Native Gemini Image Generation

const imageModel = client.imageGenerationModel("gemini-2.5-flash-preview-image");

const result = await imageModel.generate({
  prompt: "A diagram of an agent runtime",
});

// result.images[0] is Uint8Array (PNG)

Model constants: GEMINI_2_5_FLASH_IMAGE, GEMINI_3_PRO_IMAGE_PREVIEW.

Imagen Generation

const imagenModel = client.imagenGenerationModel("imagen-4-generate");

const result = await imagenModel.generate({
  prompt: "A watercolor painting of a mountain",
});

Model constant: IMAGEN_4_GENERATE.

Transcription

const transcriptionModel = client.transcriptionModel("gemini-2.5-flash");

const result = await transcriptionModel.transcribe({
  audio: audioBuffer,
});

console.log(result.text);

Uses Gemini's multimodal input to process audio.

Model Listing

const models = await client.listModels();
console.log(models.data.map((m) => m.id));

Fetches available Gemini models and returns a normalized ModelList.

Error Handling

Constructor throws when required credentials are missing
listModels() throws ModelListingError when the provider request fails
Model operations throw Google GenAI SDK errors directly

Models for model name conventions
Gemini Reference for full API types
Getting Started for a first agent walkthrough