@anvia/gemini

Gemini and Vertex AI provider adapter for Anvia.

Use this package when you want Anvia agents, extractors, pipelines, embeddings, image generation, or transcription to run on Google's Gemini APIs.

Installation

pnpm add @anvia/gemini @anvia/core

In this monorepo, the package is available through the workspace:

pnpm --filter @anvia/gemini build

Usage

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

const client = new GeminiClient({
  apiKey,
});

const model = client.completionModel({ modelId: "gemini-3.7-flash" });

const agent = new Agent({
  id: "assistant",
  model: model,
  instructions: "Answer clearly and concisely.",
});

const result = await agent.generate({ prompt: "Summarize Anvia in one sentence." });
if (result.type === "response") console.log(result.output);

Vertex AI

Use the exclusive Vertex AI configuration with a Google Cloud project and location:

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

const client = new GeminiClient({
  vertexAi: {
    projectId: "my-gcp-project",
    location: "us-central1",
  },
});

const model = client.completionModel({ modelId: "gemini-3.7-flash" });

The Google SDK uses Application Default Credentials. For a service-account JSON file, set GOOGLE_APPLICATION_CREDENTIALS before starting the application:

export GOOGLE_APPLICATION_CREDENTIALS="/absolute/path/service-account.json"

Trusted credential objects can also be passed explicitly:

const client = new GeminiClient({
  vertexAi: {
    projectId: "my-gcp-project",
    location: "us-central1",
    googleAuthOptions: {
      credentials: serviceAccountJson,
    },
  },
});

Validate externally supplied credential configurations before using them. Never commit credential JSON or service-account keys.

Embeddings

const embeddings = client.embeddingModel({ modelId: "gemini-embedding-2" });
const vectors = await embeddings.embedTexts(["Anvia is a TypeScript AI runtime."]);

Image Generation

Choose the image API explicitly. Gemini native image models use generateContent; Imagen uses generateImages.

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

const client = new GeminiClient({ apiKey });

const nativeImageModel = client.imageGenerationModel({
  api: "generateContent",
  modelId: GEMINI_2_5_FLASH_IMAGE,
});
const imagenModel = client.imageGenerationModel({
  api: "generateImages",
  modelId: IMAGEN_4_GENERATE,
});

Exports

• GeminiClient
• structural completion, embedding, image, and transcription handle types
• Gemini model-ID types
• GEMINI_2_5_FLASH_IMAGE
• GEMINI_3_1_FLASH_IMAGE
• GEMINI_3_1_FLASH_LITE_IMAGE
• GEMINI_3_PRO_IMAGE
• IMAGEN_4_GENERATE
• gemini

Development

pnpm --filter @anvia/gemini typecheck
pnpm --filter @anvia/gemini test
pnpm --filter @anvia/gemini build

Package-local typecheck and build scripts build @anvia/core first so core subpath types are available in a fresh worktree.