Deploy Voyage AI Models Using Google Cloud Model Garden

Available Models

To see which models you can deploy, search for "Voyage" in the Google Cloud Model Garden.

To learn more about Voyage AI models, see Models Overview.

Pricing

Pricing for Voyage AI by MongoDB models in the Google Cloud Model Garden includes:

• Model Usage fee: A cost for using the Voyage AI model container, billed at an hourly rate. The usage fee depends on the specific model and the hardware configuration you choose for deployment. For detailed pricing information, see the pricing section on the model's listing page in the Google Cloud Marketplace.

• Google Cloud underlying instance in your region: The cost of the underlying Google Cloud GPU instance, such as NVIDIA L4, A100, or H100) that is specific to a region, is billed monthly and priced per vCPU. To learn more, see Google Cloud Compute Engine pricing.

All billing charges appear as the use of Vertex AI on your Google Cloud bill.

To view pricing for a specific Voyage AI model:

Quotas

When you deploy Voyage AI models, you consume Vertex AI resources that are subject to quotas. You can view and manage your quotas in the Quotas section of the Google Cloud Console's IAM page. For more information, see View the quotas for your project. In the same page, you can right-click any current quota, click Edit quota, and submit a request to increase your quota if needed.

Prerequisites

To get started using the Voyage AI by MongoDB models through Google Cloud Vertex AI, you must:

• Set up a Google Cloud project and a development environment. For instructions, see Set up your project and development environment.

• Enable the Vertex AI API. For instructions, see Setup.

Hardware Configuration

Each model in the Model Garden lists its recommended hardware configuration. Consult the Google Cloud Model Garden for Vertex AI for each Voyage model's recommended hardware specifications.

For example, for the voyage-4 model, use the following recommended instances that the Vertex AI Model Garden suggests for deployment. These recommendations may change and we recommend that you consult the official Google Cloud Model Garden page for a particular Voyage AI model to see its recommended hardware.

• A2 instances, such as a2-highgpu-1g or a2-ultragpu-1g, with A100 GPUs are the default choice.

• A3 instances, such as a3-highgpu-1g, with H100 GPUs are recommended for higher performance needs.

Supported Regions

The Model Garden lists supported regions for each Voyage AI model. If you need support in another region for any of the models, contact MongoDB support.

Best Practices and Limitations

• Endpoint Type: All Voyage AI models require a dedicated public endpoint type. For more information, see Choose an endpoint type.

• Understand input_type: Query vs. Document: The input_type parameter optimizes embeddings for retrieval tasks. Use "query" for search queries and "document" for content being searched. This optimization improves retrieval accuracy. To learn more about the input_type parameter, see the Embedding and Reranking API Overview.

• Use Different Output Dimensions: Voyage 4 models support multiple output dimensions: 256, 512, 1024 (default), and 2048. Smaller dimensions reduce storage and computation costs, while larger dimensions may provide better accuracy. Choose the dimension that best balances your accuracy requirements with resource constraints.

Locate the Voyage AI Models

To find Voyage AI by MongoDB models in the Model Garden:

Deploy a Voyage AI Model in Vertex AI

To make predictions using a Voyage AI by MongoDB model, you must deploy it to a private dedicated endpoint for online inferences. Deployment associates physical resources with a model for low-latency and high-throughput online predictions. You can deploy multiple models to one endpoint, or the same model to multiple endpoints.

When you deploy a model, consider the following options:

• Endpoint location

• Model container

• Compute resources required to run the model

Once you deploy a model, you can't change these settings. If you need to modify any deployment configuration, you must undeploy the model and redeploy it with the new settings.

Voyage AI models require a dedicated public endpoint. For more information, see Create a public endpoint in the Google Cloud Vertex AI documentation.

To deploy a model in the Google Cloud Vertex AI using the console:

Deploy Using Code

If you selected View Code from the model details page, you can deploy a model programmatically using the Vertex AI SDK. This approach provides full control over deployment configuration through code.

For more information about the Google Cloud Vertex AI SDK, see the Vertex AI SDK for Python documentation.

To deploy a model using code:

from vertexai import model_garden
MODEL_NAME = "mongodb/voyage-4@latest"
model = model_garden.OpenModel(MODEL_NAME)
deploy_options = model.list_deploy_options(concise=True)
print(deploy_options)

# Choose whether to deploy a new model or use an existing endpoint:
deployment_option = "deploy_new"  # ["deploy_new", "use_existing"]
# If using existing endpoint, provide the endpoint ID:
ENDPOINT_ID = ""  # {type:"string"}
if deployment_option == "deploy_new":
    print("Deploying new model...")
    endpoint = model.deploy(
        machine_type="a3-highgpu-1g",
        accelerator_type="NVIDIA_H100_80GB",
        accelerator_count=1,
        accept_eula=True,
        use_dedicated_endpoint=True,
    )
    print(f"Endpoint deployed: {endpoint.display_name}")
    print(f"Endpoint resource name: {endpoint.resource_name}")
else:
    if not ENDPOINT_ID:
        raise ValueError("Please provide an ENDPOINT_ID when using existing endpoint")
    from google.cloud import aiplatform
    print(f"Connecting to existing endpoint: {ENDPOINT_ID}")
    endpoint = aiplatform.Endpoint(
        endpoint_name=f"projects/{PROJECT_ID}/locations/{LOCATION}/endpoints/{ENDPOINT_ID}"
    )
    print(f"Using endpoint: {endpoint.display_name}")
    print(f"Endpoint resource name: {endpoint.resource_name}")

import json
# Multiple texts to embed
texts = [
    "Machine learning enables computers to learn from data.",
    "Natural language processing helps computers understand human language.",
    "Computer vision allows machines to interpret visual information.",
    "Deep learning uses neural networks with multiple layers."
]
# Prepare the batch request and make invoke call
body = {
    "input": texts,
    "output_dimension": 1024,
    "input_type": "document"
}
response = endpoint.invoke(
    request_path="/embeddings",
    body=json.dumps(body).encode("utf-8"),
    headers={"Content-Type": "application/json"}
)
# Extract embeddings
result = response.json()
embeddings = [item["embedding"] for item in result["data"]]
print(f"Number of texts embedded: {len(embeddings)}")
print(f"Embedding dimension: {len(embeddings[0])}")
print(f"\nFirst embedding (first 5 values): {embeddings[0][:5]}")
print(f"Second embedding (first 5 values): {embeddings[1][:5]}")

Undeploy a Model and Delete the Endpoint

To remove a deployed model and its endpoint:

1. Undeploy the model from the endpoint.

2. Optionally delete the endpoint itself.

For detailed instructions, see Undeploy a model and delete the endpoint in the Google Cloud Vertex AI documentation.