Starting April 29, 2025, Gemini 1.5 Pro and Gemini 1.5 Flash models are not available in projects that have no prior usage of these models, including new projects. For details, see Model versions and lifecycle.

Get multimodal embeddings

This guide shows you how to get multimodal embeddings for text, image, and video data using the multimodalembedding model. This page covers the following topics:

Supported models: Learn about the model available for generating multimodal embeddings.
Best practices: Get tips for using the multimodal embeddings model effectively.
Before you begin: Set up your project and environment.
Send an embedding request (image and text): See how to get embeddings for image and text data.
Send an embedding request (video, image, or text): See how to get embeddings for video data, including different modes for analysis.
Specify lower-dimension embeddings: Learn how to request embeddings with different dimensions to optimize for latency, storage, or quality.
API usage and limits: Understand API quotas, available locations, and how to handle errors.

The following diagram summarizes the overall workflow for getting multimodal embeddings:

The multimodal embeddings model generates 1408-dimension vectors based on the input you provide, which can include a combination of image, text, and video data. You can then use the embedding vectors for subsequent tasks like image classification or video content moderation.

The image embedding vector and text embedding vector are in the same semantic space with the same dimensionality. Consequently, you can use these vectors interchangeably for use cases like searching for an image with text or searching for a video with an image.

For text-only embedding use cases, we recommend using the Vertex AI text-embeddings API instead. For example, the text-embeddings API might be better for text-based semantic search, clustering, long-form document analysis, and other text retrieval or question-answering use cases. For more information, see Get text embeddings.

Supported models

To get multimodal embeddings, use the following model:

multimodalembedding

Best practices

Consider the following best practices when you use the multimodal embeddings model:

Text in images: The model can distinguish text in images, similar to optical character recognition (OCR). If you need to distinguish between a description of the image content and the text within an image, use prompt engineering to specify your target content. For example, instead of "cat", specify "picture of a cat" or "the text 'cat'", depending on your use case.

the text 'cat'
picture of a cat	^{Image credit: Manja Vitolic on Unsplash.}

Embedding similarities: The dot product of embeddings isn't a calibrated probability. The dot product is a similarity metric and might have different score distributions for different use cases. Consequently, avoid using a fixed value threshold to measure quality. Instead, use ranking approaches for retrieval, or use sigmoid for classification.

Before you begin

Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.

In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

Go to project selector

Verify that billing is enabled for your Google Cloud project.

Enable the Vertex AI API.

Enable the API

In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

Go to project selector

Verify that billing is enabled for your Google Cloud project.

Enable the Vertex AI API.

Enable the API

Set up authentication for your environment.

Select the tab for how you plan to use the samples on this page:
Java

To use the Java samples on this page in a local development environment, install and initialize the gcloud CLI, and then set up Application Default Credentials with your user credentials.
1. Install the Google Cloud CLI.
2. If you're using an external identity provider (IdP), you must first sign in to the gcloud CLI with your federated identity.
3. After initializing the gcloud CLI, update it and install the required components:
  gcloud components update gcloud components install beta
For more information, see Set up ADC for a local development environment in the Google Cloud authentication documentation.
Node.js

To use the Node.js samples on this page in a local development environment, install and initialize the gcloud CLI, and then set up Application Default Credentials with your user credentials.
1. Install the Google Cloud CLI.
2. If you're using an external identity provider (IdP), you must first sign in to the gcloud CLI with your federated identity.
3. After initializing the gcloud CLI, update it and install the required components:
  gcloud components update gcloud components install beta
For more information, see Set up ADC for a local development environment in the Google Cloud authentication documentation.
Python

To use the Python samples on this page in a local development environment, install and initialize the gcloud CLI, and then set up Application Default Credentials with your user credentials.
1. Install the Google Cloud CLI.
2. If you're using an external identity provider (IdP), you must first sign in to the gcloud CLI with your federated identity.
3. After initializing the gcloud CLI, update it and install the required components:
  gcloud components update gcloud components install beta
For more information, see Set up ADC for a local development environment in the Google Cloud authentication documentation.
REST

To use the REST API samples on this page in a local development environment, you use the credentials you provide to the gcloud CLI.
1. Install the Google Cloud CLI.
2. If you're using an external identity provider (IdP), you must first sign in to the gcloud CLI with your federated identity.
3. After initializing the gcloud CLI, update it and install the required components:
  gcloud components update gcloud components install beta
For more information, see Authenticate for using REST in the Google Cloud authentication documentation.
To use the Python SDK, follow the instructions at Install the Vertex AI SDK for Python. For more information, see the Vertex AI SDK for Python API reference documentation.
Optional: Review pricing for this feature. Pricing for embeddings depends on the data type (image, text, or video) and, for video, the embedding mode you use (Essential, Standard, or Plus).

Send an embedding request (image and text)

Use the following code samples to send an embedding request with image and text data. The samples show how to send a request with both data types, but you can also send requests with only image data or only text data.

REST

For more information about multimodalembedding model requests, see the multimodalembedding model API reference.

Before using any of the request data, make the following replacements:

LOCATION: Your project's region. For example, us-central1, europe-west2, or asia-northeast3. For a list of available regions, see Generative AI on Vertex AI locations.
PROJECT_ID: Your Google Cloud project ID.
TEXT: The target text to get embeddings for. For example, a cat.
B64_ENCODED_IMG: The target image to get embeddings for. The image must be specified as a base64-encoded byte string.

HTTP method and URL:

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict

Request JSON body:

{
  "instances": [
    {
      "text": "TEXT",
      "image": {
        "bytesBase64Encoded": "B64_ENCODED_IMG"
      }
    }
  ]
}

To send your request, choose one of these options:

curl

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login , or by using Cloud Shell, which automatically logs you into the gcloud CLI . You can check the currently active account by running gcloud auth list.

Save the request body in a file named request.json, and execute the following command:

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict"

PowerShell

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login . You can check the currently active account by running gcloud auth list.

Save the request body in a file named request.json, and execute the following command:

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict" | Select-Object -Expand Content

The embedding the model returns is a 1408 float vector. The following sample response is shortened for space.

{
  "predictions": [
    {
      "textEmbedding": [
        0.010477379,
        -0.00399621,
        0.00576670747,
        [...]
        -0.00823613815,
        -0.0169572588,
        -0.00472954148
      ],
      "imageEmbedding": [
        0.00262696808,
        -0.00198890246,
        0.0152047109,
        -0.0103145819,
        [...]
        0.0324628279,
        0.0284924973,
        0.011650892,
        -0.00452344026
      ]
    }
  ],
  "deployedModelId": "DEPLOYED_MODEL_ID"
}

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation.

import vertexai
from vertexai.vision_models import Image, MultiModalEmbeddingModel

# TODO(developer): Update & uncomment line below
# PROJECT_ID = "your-project-id"
vertexai.init(project=PROJECT_ID, location="us-central1")

model = MultiModalEmbeddingModel.from_pretrained("multimodalembedding@001")
image = Image.load_from_file(
    "gs://cloud-samples-data/vertex-ai/llm/prompts/landmark1.png"
)

embeddings = model.get_embeddings(
    image=image,
    contextual_text="Colosseum",
    dimension=1408,
)
print(f"Image Embedding: {embeddings.image_embedding}")
print(f"Text Embedding: {embeddings.text_embedding}")
# Example response:
# Image Embedding: [-0.0123147098, 0.0727171078, ...]
# Text Embedding: [0.00230263756, 0.0278981831, ...]

Node.js

Before trying this sample, follow the Node.js setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Node.js API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

/**
 * TODO(developer): Uncomment these variables before running the sample.\
 * (Not necessary if passing values as arguments)
 */
// const project = 'YOUR_PROJECT_ID';
// const location = 'YOUR_PROJECT_LOCATION';
// const baseImagePath = 'YOUR_BASE_IMAGE_PATH';
// const textPrompt = 'YOUR_TEXT_PROMPT';
const aiplatform = require('@google-cloud/aiplatform');

// Imports the Google Cloud Prediction service client
const {PredictionServiceClient} = aiplatform.v1;

// Import the helper module for converting arbitrary protobuf.Value objects.
const {helpers} = aiplatform;

// Specifies the location of the api endpoint
const clientOptions = {
  apiEndpoint: 'us-central1-aiplatform.googleapis.com',
};
const publisher = 'google';
const model = 'multimodalembedding@001';

// Instantiates a client
const predictionServiceClient = new PredictionServiceClient(clientOptions);

async function predictImageFromImageAndText() {
  // Configure the parent resource
  const endpoint = `projects/${project}/locations/${location}/publishers/${publisher}/models/${model}`;

  const fs = require('fs');
  const imageFile = fs.readFileSync(baseImagePath);

  // Convert the image data to a Buffer and base64 encode it.
  const encodedImage = Buffer.from(imageFile).toString('base64');

  const prompt = {
    text: textPrompt,
    image: {
      bytesBase64Encoded: encodedImage,
    },
  };
  const instanceValue = helpers.toValue(prompt);
  const instances = [instanceValue];

  const parameter = {
    sampleCount: 1,
  };
  const parameters = helpers.toValue(parameter);

  const request = {
    endpoint,
    instances,
    parameters,
  };

  // Predict request
  const [response] = await predictionServiceClient.predict(request);
  console.log('Get image embedding response');
  const predictions = response.predictions;
  console.log('\tPredictions :');
  for (const prediction of predictions) {
    console.log(`\t\tPrediction : ${JSON.stringify(prediction)}`);
  }
}

await predictImageFromImageAndText();

Java

Before trying this sample, follow the Java setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Java API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.


import com.google.cloud.aiplatform.v1beta1.EndpointName;
import com.google.cloud.aiplatform.v1beta1.PredictResponse;
import com.google.cloud.aiplatform.v1beta1.PredictionServiceClient;
import com.google.cloud.aiplatform.v1beta1.PredictionServiceSettings;
import com.google.gson.Gson;
import com.google.gson.JsonObject;
import com.google.protobuf.InvalidProtocolBufferException;
import com.google.protobuf.Value;
import com.google.protobuf.util.JsonFormat;
import java.io.IOException;
import java.nio.charset.StandardCharsets;
import java.nio.file.Files;
import java.nio.file.Paths;
import java.util.ArrayList;
import java.util.Base64;
import java.util.HashMap;
import java.util.List;
import java.util.Map;

public class PredictImageFromImageAndTextSample {

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace this variable before running the sample.
    String project = "YOUR_PROJECT_ID";
    String textPrompt = "YOUR_TEXT_PROMPT";
    String baseImagePath = "YOUR_BASE_IMAGE_PATH";

    // Learn how to use text prompts to update an image:
    // https://cloud.google.com/vertex-ai/docs/generative-ai/image/edit-images
    Map<String, Object> parameters = new HashMap<String, Object>();
    parameters.put("sampleCount", 1);

    String location = "us-central1";
    String publisher = "google";
    String model = "multimodalembedding@001";

    predictImageFromImageAndText(
        project, location, publisher, model, textPrompt, baseImagePath, parameters);
  }

  // Update images using text prompts
  public static void predictImageFromImageAndText(
      String project,
      String location,
      String publisher,
      String model,
      String textPrompt,
      String baseImagePath,
      Map<String, Object> parameters)
      throws IOException {
    final String endpoint = String.format("%s-aiplatform.googleapis.com:443", location);
    final PredictionServiceSettings predictionServiceSettings =
        PredictionServiceSettings.newBuilder().setEndpoint(endpoint).build();

    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests.
    try (PredictionServiceClient predictionServiceClient =
        PredictionServiceClient.create(predictionServiceSettings)) {
      final EndpointName endpointName =
          EndpointName.ofProjectLocationPublisherModelName(project, location, publisher, model);

      // Convert the image to Base64
      byte[] imageData = Base64.getEncoder().encode(Files.readAllBytes(Paths.get(baseImagePath)));
      String encodedImage = new String(imageData, StandardCharsets.UTF_8);

      JsonObject jsonInstance = new JsonObject();
      jsonInstance.addProperty("text", textPrompt);
      JsonObject jsonImage = new JsonObject();
      jsonImage.addProperty("bytesBase64Encoded", encodedImage);
      jsonInstance.add("image", jsonImage);

      Value instanceValue = stringToValue(jsonInstance.toString());
      List<Value> instances = new ArrayList<>();
      instances.add(instanceValue);

      Gson gson = new Gson();
      String gsonString = gson.toJson(parameters);
      Value parameterValue = stringToValue(gsonString);

      PredictResponse predictResponse =
          predictionServiceClient.predict(endpointName, instances, parameterValue);
      System.out.println("Predict Response");
      System.out.println(predictResponse);
      for (Value prediction : predictResponse.getPredictionsList()) {
        System.out.format("\tPrediction: %s\n", prediction);
      }
    }
  }

  // Convert a Json string to a protobuf.Value
  static Value stringToValue(String value) throws InvalidProtocolBufferException {
    Value.Builder builder = Value.newBuilder();
    JsonFormat.parser().merge(value, builder);
    return builder.build();
  }
}

Go

Before trying this sample, follow the Go setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Go API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

import (
	"context"
	"encoding/json"
	"fmt"
	"io"

	aiplatform "cloud.google.com/go/aiplatform/apiv1beta1"
	aiplatformpb "cloud.google.com/go/aiplatform/apiv1beta1/aiplatformpb"
	"google.golang.org/api/option"
	"google.golang.org/protobuf/encoding/protojson"
	"google.golang.org/protobuf/types/known/structpb"
)

// generateForTextAndImage shows how to use the multimodal model to generate embeddings for
// text and image inputs.
func generateForTextAndImage(w io.Writer, project, location string) error {
	// location = "us-central1"
	ctx := context.Background()
	apiEndpoint := fmt.Sprintf("%s-aiplatform.googleapis.com:443", location)
	client, err := aiplatform.NewPredictionClient(ctx, option.WithEndpoint(apiEndpoint))
	if err != nil {
		return fmt.Errorf("failed to construct API client: %w", err)
	}
	defer client.Close()

	model := "multimodalembedding@001"
	endpoint := fmt.Sprintf("projects/%s/locations/%s/publishers/google/models/%s", project, location, model)

	// This is the input to the model's prediction call. For schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#request_body
	instance, err := structpb.NewValue(map[string]any{
		"image": map[string]any{
			// Image input can be provided either as a Google Cloud Storage URI or as
			// base64-encoded bytes using the "bytesBase64Encoded" field.
			"gcsUri": "gs://cloud-samples-data/vertex-ai/llm/prompts/landmark1.png",
		},
		"text": "Colosseum",
	})
	if err != nil {
		return fmt.Errorf("failed to construct request payload: %w", err)
	}

	req := &aiplatformpb.PredictRequest{
		Endpoint: endpoint,
		// The model supports only 1 instance per request.
		Instances: []*structpb.Value{instance},
	}

	resp, err := client.Predict(ctx, req)
	if err != nil {
		return fmt.Errorf("failed to generate embeddings: %w", err)
	}

	instanceEmbeddingsJson, err := protojson.Marshal(resp.GetPredictions()[0])
	if err != nil {
		return fmt.Errorf("failed to convert protobuf value to JSON: %w", err)
	}
	// For response schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#response-body
	var instanceEmbeddings struct {
		ImageEmbeddings []float32 `json:"imageEmbedding"`
		TextEmbeddings  []float32 `json:"textEmbedding"`
	}
	if err := json.Unmarshal(instanceEmbeddingsJson, &instanceEmbeddings); err != nil {
		return fmt.Errorf("failed to unmarshal JSON: %w", err)
	}

	imageEmbedding := instanceEmbeddings.ImageEmbeddings
	textEmbedding := instanceEmbeddings.TextEmbeddings

	fmt.Fprintf(w, "Text embedding (length=%d): %v\n", len(textEmbedding), textEmbedding)
	fmt.Fprintf(w, "Image embedding (length=%d): %v\n", len(imageEmbedding), imageEmbedding)
	// Example response:
	// Text embedding (length=1408): [0.0023026613 0.027898183 -0.011858357 ... ]
	// Image embedding (length=1408): [-0.012314269 0.07271844 0.00020170923 ... ]

	return nil
}

Send an embedding request (video, image, or text)

You can send an embedding request with an input video alone, or with a combination of video, image, and text data.

Video embedding modes

There are three modes you can use with video embeddings: Essential, Standard, or Plus. The mode corresponds to the density of the generated embeddings, which you specify with the intervalSec field in the request. The model generates one embedding for each video interval of the length that you specify. The minimum video interval length is 4 seconds. Interval lengths greater than 120 seconds might negatively affect the quality of the generated embeddings.

Pricing for video embedding depends on the mode you use. For more information, see pricing.

The following table summarizes the three modes you can use for video embeddings:

Mode	Description	Use Case	Embedding Interval (`intervalSec`)
Essential	Generates the fewest embeddings per minute of video.	Best for coarse-grained analysis where cost-efficiency is a priority and high temporal detail is not required.	>= 15 seconds
Standard	Offers a balance between embedding density and cost.	Suitable for general-purpose video analysis, such as topic segmentation or standard search and retrieval.	8 to < 15 seconds
Plus	Generates the most embeddings per minute of video, providing the highest temporal granularity.	Ideal for detailed, fine-grained analysis, such as detecting short events, action recognition, or content moderation.	4 to < 8 seconds

Video embeddings best practices

Consider the following when you send video embedding requests:

To generate a single embedding for the first two minutes of an input video of any length, use the following videoSegmentConfig setting:

request.json:
```
// other request body content
"videoSegmentConfig": {
  "intervalSec": 120
}
// other request body content
```

To generate embeddings for a video longer than two minutes, you can send multiple requests that specify the start and end times in the videoSegmentConfig:

request1.json:

// other request body content
"videoSegmentConfig": {
  "startOffsetSec": 0,
  "endOffsetSec": 120
}
// other request body content

request2.json:

// other request body content
"videoSegmentConfig": {
  "startOffsetSec": 120,
  "endOffsetSec": 240
}
// other request body content

Get video embeddings

Use the following sample to get embeddings for video content alone.

REST

For more information about multimodalembedding model requests, see the multimodalembedding model API reference.

The following example uses a video located in Cloud Storage. You can also use the video.bytesBase64Encoded field to provide a base64-encoded string representation of the video.

Before using any of the request data, make the following replacements:

LOCATION: Your project's region. For example, us-central1, europe-west2, or asia-northeast3. For a list of available regions, see Generative AI on Vertex AI locations.
PROJECT_ID: Your Google Cloud project ID.
VIDEO_URI: The Cloud Storage URI of the target video to get embeddings for. For example, gs://my-bucket/embeddings/supermarket-video.mp4.
You can also provide the video as a base64-encoded byte string:
```
[...]
"video": {
  "bytesBase64Encoded": "B64_ENCODED_VIDEO"
}
[...]
```
videoSegmentConfig (START_SECOND, END_SECOND, INTERVAL_SECONDS). Optional. The specific video segments (in seconds) the embeddings are generated for.
The value you set for videoSegmentConfig.intervalSec affects the pricing tier you are charged at. For more information, see the video embedding modes section and pricing page.

For example:
```
[...]
"videoSegmentConfig": {
  "startOffsetSec": 10,
  "endOffsetSec": 60,
  "intervalSec": 10
}
[...]
```
Using this config specifies video data from 10 seconds to 60 seconds and generates embeddings for the following 10 second video intervals: [10, 20), [20, 30), [30, 40), [40, 50), [50, 60). This video interval ("intervalSec": 10) falls in the Standard video embedding mode, and the user is charged at the Standard mode pricing rate.

If you omit videoSegmentConfig, the service uses the following default values: "videoSegmentConfig": { "startOffsetSec": 0, "endOffsetSec": 120, "intervalSec": 16 }. This video interval ("intervalSec": 16) falls in the Essential video embedding mode, and the user is charged at the Essential mode pricing rate.

HTTP method and URL:

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict

Request JSON body:

{
  "instances": [
    {
      "video": {
        "gcsUri": "VIDEO_URI",
        "videoSegmentConfig": {
          "startOffsetSec": START_SECOND,
          "endOffsetSec": END_SECOND,
          "intervalSec": INTERVAL_SECONDS
        }
      }
    }
  ]
}

To send your request, choose one of these options:

curl

Save the request body in a file named request.json, and execute the following command:

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict"

PowerShell

Save the request body in a file named request.json, and execute the following command:

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict" | Select-Object -Expand Content

The embedding the model returns is a 1408 float vector. The following sample responses are shortened for space.

Response (7 second video, no videoSegmentConfig specified):

{
  "predictions": [
    {
      "videoEmbeddings": [
        {
          "endOffsetSec": 7,
          "embedding": [
            -0.0045467657,
            0.0258095954,
            0.0146885719,
            0.00945400633,
            [...]
            -0.0023291884,
            -0.00493789,
            0.00975185353,
            0.0168156829
          ],
          "startOffsetSec": 0
        }
      ]
    }
  ],
  "deployedModelId": "DEPLOYED_MODEL_ID"
}

Response (59 second video, with the following video segment config: "videoSegmentConfig": { "startOffsetSec": 0, "endOffsetSec": 60, "intervalSec": 10 }):

{
  "predictions": [
    {
      "videoEmbeddings": [
        {
          "endOffsetSec": 10,
          "startOffsetSec": 0,
          "embedding": [
            -0.00683252793,
            0.0390476175,
            [...]
            0.00657121744,
            0.013023301
          ]
        },
        {
          "startOffsetSec": 10,
          "endOffsetSec": 20,
          "embedding": [
            -0.0104404651,
            0.0357737206,
            [...]
            0.00509833824,
            0.0131902946
          ]
        },
        {
          "startOffsetSec": 20,
          "embedding": [
            -0.0113538112,
            0.0305239167,
            [...]
            -0.00195809244,
            0.00941874553
          ],
          "endOffsetSec": 30
        },
        {
          "embedding": [
            -0.00299320649,
            0.0322436653,
            [...]
            -0.00993082579,
            0.00968887936
          ],
          "startOffsetSec": 30,
          "endOffsetSec": 40
        },
        {
          "endOffsetSec": 50,
          "startOffsetSec": 40,
          "embedding": [
            -0.00591270532,
            0.0368893594,
            [...]
            -0.00219071587,
            0.0042470959
          ]
        },
        {
          "embedding": [
            -0.00458270218,
            0.0368121453,
            [...]
            -0.00317760976,
            0.00595594104
          ],
          "endOffsetSec": 59,
          "startOffsetSec": 50
        }
      ]
    }
  ],
  "deployedModelId": "DEPLOYED_MODEL_ID"
}

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation.

import vertexai

from vertexai.vision_models import MultiModalEmbeddingModel, Video
from vertexai.vision_models import VideoSegmentConfig

# TODO(developer): Update & uncomment line below
# PROJECT_ID = "your-project-id"
vertexai.init(project=PROJECT_ID, location="us-central1")

model = MultiModalEmbeddingModel.from_pretrained("multimodalembedding@001")

embeddings = model.get_embeddings(
    video=Video.load_from_file(
        "gs://cloud-samples-data/vertex-ai-vision/highway_vehicles.mp4"
    ),
    video_segment_config=VideoSegmentConfig(end_offset_sec=1),
)

# Video Embeddings are segmented based on the video_segment_config.
print("Video Embeddings:")
for video_embedding in embeddings.video_embeddings:
    print(
        f"Video Segment: {video_embedding.start_offset_sec} - {video_embedding.end_offset_sec}"
    )
    print(f"Embedding: {video_embedding.embedding}")

# Example response:
# Video Embeddings:
# Video Segment: 0.0 - 1.0
# Embedding: [-0.0206376351, 0.0123456789, ...]

Go

Before trying this sample, follow the Go setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Go API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

import (
	"context"
	"encoding/json"
	"fmt"
	"io"
	"time"

	aiplatform "cloud.google.com/go/aiplatform/apiv1beta1"
	aiplatformpb "cloud.google.com/go/aiplatform/apiv1beta1/aiplatformpb"
	"google.golang.org/api/option"
	"google.golang.org/protobuf/encoding/protojson"
	"google.golang.org/protobuf/types/known/structpb"
)

// generateForVideo shows how to use the multimodal model to generate embeddings for video input.
func generateForVideo(w io.Writer, project, location string) error {
	// location = "us-central1"

	// The default context timeout may be not enough to process a video input.
	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
	defer cancel()

	apiEndpoint := fmt.Sprintf("%s-aiplatform.googleapis.com:443", location)
	client, err := aiplatform.NewPredictionClient(ctx, option.WithEndpoint(apiEndpoint))
	if err != nil {
		return fmt.Errorf("failed to construct API client: %w", err)
	}
	defer client.Close()

	model := "multimodalembedding@001"
	endpoint := fmt.Sprintf("projects/%s/locations/%s/publishers/google/models/%s", project, location, model)

	// This is the input to the model's prediction call. For schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#request_body
	instances, err := structpb.NewValue(map[string]any{
		"video": map[string]any{
			// Video input can be provided either as a Google Cloud Storage URI or as base64-encoded
			// bytes using the "bytesBase64Encoded" field.
			"gcsUri": "gs://cloud-samples-data/vertex-ai-vision/highway_vehicles.mp4",
			"videoSegmentConfig": map[string]any{
				"startOffsetSec": 1,
				"endOffsetSec":   5,
			},
		},
	})
	if err != nil {
		return fmt.Errorf("failed to construct request payload: %w", err)
	}

	req := &aiplatformpb.PredictRequest{
		Endpoint: endpoint,
		// The model supports only 1 instance per request.
		Instances: []*structpb.Value{instances},
	}
	resp, err := client.Predict(ctx, req)
	if err != nil {
		return fmt.Errorf("failed to generate embeddings: %w", err)
	}

	instanceEmbeddingsJson, err := protojson.Marshal(resp.GetPredictions()[0])
	if err != nil {
		return fmt.Errorf("failed to convert protobuf value to JSON: %w", err)
	}
	// For response schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#response-body
	var instanceEmbeddings struct {
		VideoEmbeddings []struct {
			Embedding      []float32 `json:"embedding"`
			StartOffsetSec float64   `json:"startOffsetSec"`
			EndOffsetSec   float64   `json:"endOffsetSec"`
		} `json:"videoEmbeddings"`
	}
	if err := json.Unmarshal(instanceEmbeddingsJson, &instanceEmbeddings); err != nil {
		return fmt.Errorf("failed to unmarshal json: %w", err)
	}
	// Get the embedding for our single video segment (`.videoEmbeddings` object has one entry per
	// each processed segment).
	videoEmbedding := instanceEmbeddings.VideoEmbeddings[0]

	fmt.Fprintf(w, "Video embedding (seconds: %.f-%.f; length=%d): %v\n",
		videoEmbedding.StartOffsetSec,
		videoEmbedding.EndOffsetSec,
		len(videoEmbedding.Embedding),
		videoEmbedding.Embedding,
	)
	// Example response:
	// Video embedding (seconds: 1-5; length=1408): [-0.016427778 0.032878537 -0.030755188 ... ]

	return nil
}

Get image, text, and video embeddings

Use the following sample to get embeddings for video, text, and image content.

REST

For more information about multimodalembedding model requests, see the multimodalembedding model API reference.

The following example uses image, text, and video data. You can use any combination of these data types in your request body.

Additionally, this sample uses a video located in Cloud Storage. You can also use the video.bytesBase64Encoded field to provide a base64-encoded string representation of the video.

Before using any of the request data, make the following replacements:

LOCATION: Your project's region. For example, us-central1, europe-west2, or asia-northeast3. For a list of available regions, see Generative AI on Vertex AI locations.
PROJECT_ID: Your Google Cloud project ID.
TEXT: The target text to get embeddings for. For example, a cat.
IMAGE_URI: The Cloud Storage URI of the target image to get embeddings for. For example, gs://my-bucket/embeddings/supermarket-img.png.
You can also provide the image as a base64-encoded byte string:
```
[...]
"image": {
  "bytesBase64Encoded": "B64_ENCODED_IMAGE"
}
[...]
```
VIDEO_URI: The Cloud Storage URI of the target video to get embeddings for. For example, gs://my-bucket/embeddings/supermarket-video.mp4.
You can also provide the video as a base64-encoded byte string:
```
[...]
"video": {
  "bytesBase64Encoded": "B64_ENCODED_VIDEO"
}
[...]
```
videoSegmentConfig (START_SECOND, END_SECOND, INTERVAL_SECONDS). Optional. The specific video segments (in seconds) the embeddings are generated for.
The value you set for videoSegmentConfig.intervalSec affects the pricing tier you are charged at. For more information, see the video embedding modes section and pricing page.

For example:
```
[...]
"videoSegmentConfig": {
  "startOffsetSec": 10,
  "endOffsetSec": 60,
  "intervalSec": 10
}
[...]
```
Using this config specifies video data from 10 seconds to 60 seconds and generates embeddings for the following 10 second video intervals: [10, 20), [20, 30), [30, 40), [40, 50), [50, 60). This video interval ("intervalSec": 10) falls in the Standard video embedding mode, and the user is charged at the Standard mode pricing rate.

If you omit videoSegmentConfig, the service uses the following default values: "videoSegmentConfig": { "startOffsetSec": 0, "endOffsetSec": 120, "intervalSec": 16 }. This video interval ("intervalSec": 16) falls in the Essential video embedding mode, and the user is charged at the Essential mode pricing rate.

HTTP method and URL:

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict

Request JSON body:

{
  "instances": [
    {
      "text": "TEXT",
      "image": {
        "gcsUri": "IMAGE_URI"
      },
      "video": {
        "gcsUri": "VIDEO_URI",
        "videoSegmentConfig": {
          "startOffsetSec": START_SECOND,
          "endOffsetSec": END_SECOND,
          "intervalSec": INTERVAL_SECONDS
        }
      }
    }
  ]
}

To send your request, choose one of these options:

curl

Save the request body in a file named request.json, and execute the following command:

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict"

PowerShell

Save the request body in a file named request.json, and execute the following command:

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict" | Select-Object -Expand Content

The embedding the model returns is a 1408 float vector. The following sample response is shortened for space.

{
  "predictions": [
    {
      "textEmbedding": [
        0.0105433334,
        -0.00302835181,
        0.00656806398,
        0.00603460241,
        [...]
        0.00445805816,
        0.0139605571,
        -0.00170318608,
        -0.00490092579
      ],
      "videoEmbeddings": [
        {
          "startOffsetSec": 0,
          "endOffsetSec": 7,
          "embedding": [
            -0.00673126569,
            0.0248149596,
            0.0128901172,
            0.0107588246,
            [...]
            -0.00180952181,
            -0.0054573305,
            0.0117037306,
            0.0169312079
          ]
        }
      ],
      "imageEmbedding": [
        -0.00728622358,
        0.031021487,
        -0.00206603738,
        0.0273937676,
        [...]
        -0.00204976718,
        0.00321615417,
        0.0121978866,
        0.0193375275
      ]
    }
  ],
  "deployedModelId": "DEPLOYED_MODEL_ID"
}

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation.

import vertexai

from vertexai.vision_models import Image, MultiModalEmbeddingModel, Video
from vertexai.vision_models import VideoSegmentConfig

# TODO(developer): Update & uncomment line below
# PROJECT_ID = "your-project-id"
vertexai.init(project=PROJECT_ID, location="us-central1")

model = MultiModalEmbeddingModel.from_pretrained("multimodalembedding@001")

image = Image.load_from_file(
    "gs://cloud-samples-data/vertex-ai/llm/prompts/landmark1.png"
)
video = Video.load_from_file(
    "gs://cloud-samples-data/vertex-ai-vision/highway_vehicles.mp4"
)

embeddings = model.get_embeddings(
    image=image,
    video=video,
    video_segment_config=VideoSegmentConfig(end_offset_sec=1),
    contextual_text="Cars on Highway",
)

print(f"Image Embedding: {embeddings.image_embedding}")

# Video Embeddings are segmented based on the video_segment_config.
print("Video Embeddings:")
for video_embedding in embeddings.video_embeddings:
    print(
        f"Video Segment: {video_embedding.start_offset_sec} - {video_embedding.end_offset_sec}"
    )
    print(f"Embedding: {video_embedding.embedding}")

print(f"Text Embedding: {embeddings.text_embedding}")
# Example response:
# Image Embedding: [-0.0123144267, 0.0727186054, 0.000201397663, ...]
# Video Embeddings:
# Video Segment: 0.0 - 1.0
# Embedding: [-0.0206376351, 0.0345234685, ...]
# Text Embedding: [-0.0207006838, -0.00251058186, ...]

Go

Before trying this sample, follow the Go setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Go API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

import (
	"context"
	"encoding/json"
	"fmt"
	"io"
	"time"

	aiplatform "cloud.google.com/go/aiplatform/apiv1beta1"
	aiplatformpb "cloud.google.com/go/aiplatform/apiv1beta1/aiplatformpb"
	"google.golang.org/api/option"
	"google.golang.org/protobuf/encoding/protojson"
	"google.golang.org/protobuf/types/known/structpb"
)

// generateForImageTextAndVideo shows how to use the multimodal model to generate embeddings for
// image, text and video data.
func generateForImageTextAndVideo(w io.Writer, project, location string) error {
	// location = "us-central1"

	// The default context timeout may be not enough to process a video input.
	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
	defer cancel()

	apiEndpoint := fmt.Sprintf("%s-aiplatform.googleapis.com:443", location)
	client, err := aiplatform.NewPredictionClient(ctx, option.WithEndpoint(apiEndpoint))
	if err != nil {
		return fmt.Errorf("failed to construct API client: %w", err)
	}
	defer client.Close()

	model := "multimodalembedding@001"
	endpoint := fmt.Sprintf("projects/%s/locations/%s/publishers/google/models/%s", project, location, model)

	// This is the input to the model's prediction call. For schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#request_body
	instance, err := structpb.NewValue(map[string]any{
		"text": "Domestic cats in natural conditions",
		"image": map[string]any{
			// Image and video inputs can be provided either as a Google Cloud Storage URI or as
			// base64-encoded bytes using the "bytesBase64Encoded" field.
			"gcsUri": "gs://cloud-samples-data/generative-ai/image/320px-Felis_catus-cat_on_snow.jpg",
		},
		"video": map[string]any{
			"gcsUri": "gs://cloud-samples-data/video/cat.mp4",
		},
	})
	if err != nil {
		return fmt.Errorf("failed to construct request payload: %w", err)
	}

	req := &aiplatformpb.PredictRequest{
		Endpoint: endpoint,
		// The model supports only 1 instance per request.
		Instances: []*structpb.Value{instance},
	}

	resp, err := client.Predict(ctx, req)
	if err != nil {
		return fmt.Errorf("failed to generate embeddings: %w", err)
	}

	instanceEmbeddingsJson, err := protojson.Marshal(resp.GetPredictions()[0])
	if err != nil {
		return fmt.Errorf("failed to convert protobuf value to JSON: %w", err)
	}
	// For response schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#response-body
	var instanceEmbeddings struct {
		ImageEmbeddings []float32 `json:"imageEmbedding"`
		TextEmbeddings  []float32 `json:"textEmbedding"`
		VideoEmbeddings []struct {
			Embedding      []float32 `json:"embedding"`
			StartOffsetSec float64   `json:"startOffsetSec"`
			EndOffsetSec   float64   `json:"endOffsetSec"`
		} `json:"videoEmbeddings"`
	}
	if err := json.Unmarshal(instanceEmbeddingsJson, &instanceEmbeddings); err != nil {
		return fmt.Errorf("failed to unmarshal JSON: %w", err)
	}

	imageEmbedding := instanceEmbeddings.ImageEmbeddings
	textEmbedding := instanceEmbeddings.TextEmbeddings
	// Get the embedding for our single video segment (`.videoEmbeddings` object has one entry per
	// each processed segment).
	videoEmbedding := instanceEmbeddings.VideoEmbeddings[0].Embedding

	fmt.Fprintf(w, "Image embedding (length=%d): %v\n", len(imageEmbedding), imageEmbedding)
	fmt.Fprintf(w, "Text embedding (length=%d): %v\n", len(textEmbedding), textEmbedding)
	fmt.Fprintf(w, "Video embedding (length=%d): %v\n", len(videoEmbedding), videoEmbedding)
	// Example response:
	// Image embedding (length=1408): [-0.01558477 0.0258355 0.016342038 ... ]
	// Text embedding (length=1408): [-0.005894961 0.008349559 0.015355394 ... ]
	// Video embedding (length=1408): [-0.018867437 0.013997682 0.0012682161 ... ]

	return nil
}

Specify lower-dimension embeddings

By default, an embedding request returns a 1408-dimension float vector for a data type. You can also specify lower-dimension embeddings (128, 256, or 512 float vectors) for text and image data. This option lets you optimize for latency and storage or quality based on how you plan to use the embeddings. Lower-dimension embeddings require less storage and provide lower latency for subsequent embedding tasks (like search or recommendation), while higher-dimension embeddings offer greater accuracy for the same tasks.

Dimension	Description	Pros	Cons	Use Case
128	Lowest dimension embedding.	Fastest performance, lowest storage and memory cost.	Lowest accuracy and granularity.	Applications where speed and cost are critical, and lower accuracy is acceptable (e.g., large-scale, coarse-grained search).
256	A balance between performance and quality.	Good performance with moderate storage costs.	Less accurate than higher-dimension embeddings.	General-purpose applications requiring a balance of speed and accuracy.
512	Higher-quality embedding.	Improved accuracy and detail representation.	Slower performance and higher storage costs than lower dimensions.	Tasks requiring higher accuracy, such as fine-grained similarity search or classification.
1408 (default)	Highest dimension embedding.	Highest accuracy and most detailed semantic representation.	Highest latency and storage/memory requirements.	High-stakes applications where quality is paramount and performance costs are acceptable.

REST

For more information about multimodalembedding model requests, see the multimodalembedding model API reference.

To request a lower-dimension embedding, add the parameters.dimension field to your request. Set the value to 128, 256, 512, or 1408. The response then includes an embedding of the specified dimension.

Before using any of the request data, make the following replacements:

LOCATION: Your project's region. For example, us-central1, europe-west2, or asia-northeast3. For a list of available regions, see Generative AI on Vertex AI locations.
PROJECT_ID: Your Google Cloud project ID.
IMAGE_URI: The Cloud Storage URI of the target image to get embeddings for. For example, gs://my-bucket/embeddings/supermarket-img.png.
You can also provide the image as a base64-encoded byte string:
```
[...]
"image": {
  "bytesBase64Encoded": "B64_ENCODED_IMAGE"
}
[...]
```
TEXT: The target text to get embeddings for. For example, a cat.
EMBEDDING_DIMENSION: The number of embedding dimensions. Lower values offer decreased latency when using these embeddings for subsequent tasks, while higher values offer better accuracy. Available values: 128, 256, 512, and 1408 (default).

HTTP method and URL:

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict

Request JSON body:

{
  "instances": [
    {
      "image": {
        "gcsUri": "IMAGE_URI"
      },
      "text": "TEXT"
    }
  ],
  "parameters": {
    "dimension": EMBEDDING_DIMENSION
  }
}

To send your request, choose one of these options:

curl

Save the request body in a file named request.json, and execute the following command:

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict"

PowerShell

Save the request body in a file named request.json, and execute the following command:

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/multimodalembedding@001:predict" | Select-Object -Expand Content

The embedding the model returns a float vector of the dimension you specify. The following sample responses are shortened for space.

128 dimensions:

{
  "predictions": [
    {
      "imageEmbedding": [
        0.0279239565,
        [...128 dimension vector...]
        0.00403284049
      ],
      "textEmbedding": [
        0.202921599,
        [...128 dimension vector...]
        -0.0365431122
      ]
    }
  ],
  "deployedModelId": "DEPLOYED_MODEL_ID"
}

256 dimensions:

{
  "predictions": [
    {
      "imageEmbedding": [
        0.248620048,
        [...256 dimension vector...]
        -0.0646447465
      ],
      "textEmbedding": [
        0.0757875815,
        [...256 dimension vector...]
        -0.02749932
      ]
    }
  ],
  "deployedModelId": "DEPLOYED_MODEL_ID"
}

512 dimensions:

{
  "predictions": [
    {
      "imageEmbedding": [
        -0.0523675755,
        [...512 dimension vector...]
        -0.0444030389
      ],
      "textEmbedding": [
        -0.0592851527,
        [...512 dimension vector...]
        0.0350437127
      ]
    }
  ],
  "deployedModelId": "DEPLOYED_MODEL_ID"
}

Python

To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation.

import vertexai

from vertexai.vision_models import Image, MultiModalEmbeddingModel

# TODO(developer): Update & uncomment line below
# PROJECT_ID = "your-project-id"
vertexai.init(project=PROJECT_ID, location="us-central1")

# TODO(developer): Try different dimenions: 128, 256, 512, 1408
embedding_dimension = 128

model = MultiModalEmbeddingModel.from_pretrained("multimodalembedding@001")
image = Image.load_from_file(
    "gs://cloud-samples-data/vertex-ai/llm/prompts/landmark1.png"
)

embeddings = model.get_embeddings(
    image=image,
    contextual_text="Colosseum",
    dimension=embedding_dimension,
)

print(f"Image Embedding: {embeddings.image_embedding}")
print(f"Text Embedding: {embeddings.text_embedding}")

# Example response:
# Image Embedding: [0.0622573346, -0.0406507477, 0.0260440577, ...]
# Text Embedding: [0.27469793, -0.146258667, 0.0222803634, ...]

Go

Before trying this sample, follow the Go setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Go API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

import (
	"context"
	"encoding/json"
	"fmt"
	"io"

	aiplatform "cloud.google.com/go/aiplatform/apiv1beta1"
	aiplatformpb "cloud.google.com/go/aiplatform/apiv1beta1/aiplatformpb"
	"google.golang.org/api/option"
	"google.golang.org/protobuf/encoding/protojson"
	"google.golang.org/protobuf/types/known/structpb"
)

// generateWithLowerDimension shows how to generate lower-dimensional embeddings for text and image inputs.
func generateWithLowerDimension(w io.Writer, project, location string) error {
	// location = "us-central1"
	ctx := context.Background()
	apiEndpoint := fmt.Sprintf("%s-aiplatform.googleapis.com:443", location)
	client, err := aiplatform.NewPredictionClient(ctx, option.WithEndpoint(apiEndpoint))
	if err != nil {
		return fmt.Errorf("failed to construct API client: %w", err)
	}
	defer client.Close()

	model := "multimodalembedding@001"
	endpoint := fmt.Sprintf("projects/%s/locations/%s/publishers/google/models/%s", project, location, model)

	// This is the input to the model's prediction call. For schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#request_body
	instance, err := structpb.NewValue(map[string]any{
		"image": map[string]any{
			// Image input can be provided either as a Google Cloud Storage URI or as
			// base64-encoded bytes using the "bytesBase64Encoded" field.
			"gcsUri": "gs://cloud-samples-data/vertex-ai/llm/prompts/landmark1.png",
		},
		"text": "Colosseum",
	})
	if err != nil {
		return fmt.Errorf("failed to construct request payload: %w", err)
	}

	// TODO(developer): Try different dimenions: 128, 256, 512, 1408
	outputDimensionality := 128
	params, err := structpb.NewValue(map[string]any{
		"dimension": outputDimensionality,
	})
	if err != nil {
		return fmt.Errorf("failed to construct request params: %w", err)
	}

	req := &aiplatformpb.PredictRequest{
		Endpoint: endpoint,
		// The model supports only 1 instance per request.
		Instances:  []*structpb.Value{instance},
		Parameters: params,
	}

	resp, err := client.Predict(ctx, req)
	if err != nil {
		return fmt.Errorf("failed to generate embeddings: %w", err)
	}

	instanceEmbeddingsJson, err := protojson.Marshal(resp.GetPredictions()[0])
	if err != nil {
		return fmt.Errorf("failed to convert protobuf value to JSON: %w", err)
	}
	// For response schema, see:
	// https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api#response-body
	var instanceEmbeddings struct {
		ImageEmbeddings []float32 `json:"imageEmbedding"`
		TextEmbeddings  []float32 `json:"textEmbedding"`
	}
	if err := json.Unmarshal(instanceEmbeddingsJson, &instanceEmbeddings); err != nil {
		return fmt.Errorf("failed to unmarshal JSON: %w", err)
	}

	imageEmbedding := instanceEmbeddings.ImageEmbeddings
	textEmbedding := instanceEmbeddings.TextEmbeddings

	fmt.Fprintf(w, "Text embedding (length=%d): %v\n", len(textEmbedding), textEmbedding)
	fmt.Fprintf(w, "Image embedding (length=%d): %v\n", len(imageEmbedding), imageEmbedding)
	// Example response:
	// Text Embedding (length=128): [0.27469793 -0.14625867 0.022280363 ... ]
	// Image Embedding (length=128): [0.06225733 -0.040650766 0.02604402 ... ]

	return nil
}

API usage and limits

API limits

The following limits apply when you use the multimodalembedding model for text and image embeddings:

Limit	Value and description
	Text and image data
Maximum number of API requests per minute per project	120-600 depending on region
Maximum text length	32 tokens (~32 words). If the input text exceeds this limit, the model truncates it to 32 tokens.
Language	English
Image formats	BMP, GIF, JPG, PNG
Image size	20 MB. For base64-encoded images, this limit applies after transcoding to PNG. For Cloud Storage images, this limit applies to the original file. To reduce network latency, use smaller images. The model resizes all images to 512x512 pixels, so you don't need to provide higher-resolution images.
	Video data
Audio supported	N/A - The model doesn't consider audio content when generating video embeddings.
Video formats	AVI, FLV, MKV, MOV, MP4, MPEG, MPG, WEBM, and WMV
Maximum video length (Cloud Storage)	No limit. However, the model analyzes only 2 minutes of content at a time.

Locations

A location is a region you can specify in a request to control where data is stored at rest. For a list of available regions, see Generative AI on Vertex AI locations.

Error messages

Quota exceeded error

google.api_core.exceptions.ResourceExhausted: 429 Quota exceeded for
aiplatform.googleapis.com/online_prediction_requests_per_base_model with base
model: multimodalembedding. Please submit a quota increase request.

If this is the first time you receive this error, use the Google Cloud console to request a quota adjustment for your project. Use the following filters before requesting your adjustment:

Service ID: aiplatform.googleapis.com
metric: aiplatform.googleapis.com/online_prediction_requests_per_base_model
base_model:multimodalembedding

Go to Quotas

If you have already sent a quota adjustment request, wait before sending another request. If you need to further increase the quota, repeat the quota adjustment request with your justification for a sustained quota adjustment.

What's next

Read the blog "What is Multimodal Search: 'LLMs with vision' change businesses".
For information about text-only use cases (text-based semantic search, clustering, long-form document analysis, and other text retrieval or question-answering use cases), read Get text embeddings.
View all Vertex AI image generative AI offerings in the Imagen on Vertex AI overview.
Explore more pretrained models in Model Garden.
Learn about responsible AI best practices and safety filters in Vertex AI.

Get multimodal embeddings Stay organized with collections Save and categorize content based on your preferences.

Supported models

Best practices

Before you begin

Java

Node.js

Python

REST

Send an embedding request (image and text)

REST

curl

PowerShell

Python

Node.js

Java

Go

Send an embedding request (video, image, or text)

Video embedding modes

Video embeddings best practices

Get video embeddings

REST

curl

PowerShell

Python

Go

Get image, text, and video embeddings

REST

curl

PowerShell

Python

Go

Specify lower-dimension embeddings

REST

curl

PowerShell

Python

Go

API usage and limits

API limits

Locations

Error messages

Quota exceeded error

What's next

Get multimodal embeddings