Generate multimodal embeddings

Select a documentation version:

You can generate multimodal embeddings in AlloyDB Omni using the supported Vertex AI multimodal model, multimodalembedding@001.

You can use the Vertex AI multimodal embedding models referred to in Supported models.

This page assumes that you're familiar with AlloyDB Omni and generative AI concepts. For more information, see What are embeddings.

Before you begin

Before you use multimodal embeddings, do the following:

  1. Configure user access to Vertex AI models.
  2. Verify that the latest version of google_ml_integration is installed.

    1. To check the installed version, run the following command:

      SELECT extversion FROM pg_extension WHERE extname = 'google_ml_integration';
extversion
------------
1.5.2
(1 row)

    2. If the extension isn't installed or if the installed version is earlier than 1.5.2, update the extension.

      CREATE EXTENSION IF NOT EXISTS google_ml_integration;
ALTER EXTENSION google_ml_integration UPDATE;

      If you experience issues when you run the preceding commands, or if the extension isn't updated to version 1.5.2 after you run the preceding commands, contact Google Cloud support.

  3. To use the AlloyDB AI query engine functionality, set the google_ml_integration.enable_ai_query_engine flag to on.

    1. Use a text editor to set the following flag in the postgresql.conf configuration file for your installation of AlloyDB Omni: 
      google_ml_integration.enable_ai_query_engine = on
    2. After you save the postgresql.conf file, restart the AlloyDB Omni service for the changes to take effect. 
      sudo systemctl restart alloydbomni18

  4. Access data in Cloud Storage to generate multimodal embeddings.

Access data in Cloud Storage to generate multimodal embeddings

  • To generate multimodal embeddings, refer to content in Cloud Storage using a gs:// URI.
  • Access Cloud Storage content through your current project's Vertex AI service agent. By default, the Vertex AI service agent already has permission to access the bucket in the same project. For more information, see IAM roles and permissions index.

  • To access data in a Cloud Storage bucket in another Google Cloud project, run the following gcloud CLI command to grant the Storage Object Viewer role (roles/storage.objectViewer) to the Vertex AI service agent of your Google Cloud project.

    gcloud projects add-iam-policy-binding <ANOTHER_PROJECT_ID> \
--member="serviceAccount:service-<PROJECT_ID>@gcp-sa-aiplatform.iam.gserviceaccount.com" \
--role="roles/storage.objectViewer"

    For more information, see Set and manage IAM policies on buckets.

To generate multimodal embeddings, select one of the following schemas.

Generate multimodal embeddings

To generate text embeddings for a multimodalembedding@001 model endpoint, run the following statement:

SELECT
  ai.text_embedding(
    model_id => 'multimodalembedding@001',
    content => 'TEXT');

Replace TEXT with the text to generate the embedding for.

To generate image embeddings for a registered multimodalembedding@001 model endpoint where the image mimetype is default image/jpeg, run the following statement:

SELECT
  ai.image_embedding(
    model_id => 'multimodalembedding@001',
    image => 'IMAGE_PATH_OR_TEXT',
    mimetype => 'MIMETYPE');

Replace the following:

  • IMAGE_PATH_OR_TEXT with the Cloud Storage URI of the image, for example, gs://my-bucket/embeddings/flowers.jpeg, or the base64 string of the image.
  • MIMETYPE with the mimetype of the image, for example, image/jpeg. For the full list of supported mimetypes, see the Multimodal embeddings API.

To generate video embeddings for a registered multimodalembedding@001 model endpoint, run the following statement:

SELECT
  ai.video_embedding(
    model_id => 'multimodalembedding@001',
    video => 'VIDEO_URI');

Replace VIDEO_URI with the Cloud Storage URI of the target video, for example, gs://my-bucket/embeddings/supermarket-video.mp4, or the base64 string of the video. These are two-dimensional arrays that can be accessed as my_array[0][5] like syntax. For more information, see Arrays.

What's next