Gemini provider

@anvia/gemini adapts Google’s @google/genai SDK to Anvia contracts. Use it for Gemini API or Vertex AI completions, embeddings, image generation, transcription, and model listing.

Install

pnpm add @anvia/core @anvia/gemini

Create a Gemini API client with an API key:

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

export const gemini = new GeminiClient({
  apiKey: process.env.GEMINI_API_KEY,
});

Or create a Vertex AI client:

const vertexGemini = new GeminiClient({
  vertexai: true,
  project: process.env.GOOGLE_CLOUD_PROJECT,
  location: "us-central1",
});

GeminiClient requires either apiKey for Gemini API mode or vertexai: true with project and location for Vertex AI mode. You can also pass an already-created GoogleGenAI client.

Completion Models

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

const gemini = new GeminiClient({
  apiKey: process.env.GEMINI_API_KEY,
});

const model = gemini.completionModel("gemini-2.5-flash");

export const agent = new AgentBuilder("assistant", model)
  .instructions("Answer clearly and concisely.")
  .build();

GeminiCompletionModel supports streaming, tools, tool choice, image input, document input, output schemas, and reasoning content at the Anvia contract level.

Gemini-specific generation config belongs in completion additionalParams.config:

import { createCompletion } from "@anvia/core";

const response = await createCompletion(model, {
  input: "Draft a short release note.",
  params: {
    config: {
      topP: 0.8,
    },
  },
});

Embeddings

const embeddings = gemini.embeddingModel("gemini-embedding-001", {
  taskType: "RETRIEVAL_DOCUMENT",
  dimensions: 768,
  maxBatchSize: 100,
});

const vectors = await embeddings.embedTexts(["Anvia is a TypeScript AI runtime."]);

GeminiEmbeddingModelOptions supports dimensions, maxBatchSize, taskType, and title.

Image Generation

Gemini exposes two image model factories:

import {
  GEMINI_2_5_FLASH_IMAGE,
  IMAGEN_4_GENERATE,
  GeminiClient,
} from "@anvia/gemini";

const gemini = new GeminiClient({
  apiKey: process.env.GEMINI_API_KEY,
});

const nativeImageModel = gemini.imageGenerationModel(GEMINI_2_5_FLASH_IMAGE);
const imagenModel = gemini.imagenGenerationModel(IMAGEN_4_GENERATE);

imageGenerationModel(...) uses Gemini native image generation through models.generateContent. imagenGenerationModel(...) uses Imagen through models.generateImages.

Both map the core width and height request to an aspect ratio. Use additionalParams.config for provider-specific image config such as output count or aspect behavior.

Transcription

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

Gemini transcription sends audio bytes as inline data to models.generateContent. The adapter infers a MIME type from the request filename and applies a transcription system instruction.

Model Listing

const models = await gemini.listModels();

Gemini model listing normalizes model ids from Gemini API responses and Vertex AI compatible responses when available.

Exports

The root package exports GeminiClient, GeminiCompletionModel, GeminiEmbeddingModel, GeminiImageGenerationModel, GeminiImagenGenerationModel, GeminiTranscriptionModel, GeminiEmbeddingModelOptions, GeminiEmbeddingTaskType, constants such as GEMINI_2_5_FLASH_IMAGE, GEMINI_3_PRO_IMAGE_PREVIEW, IMAGEN_4_GENERATE, and the gemini namespace.