Invoke online predictions from AlloyDB databases

This page shows how to invoke online predictions from an AlloyDB for PostgreSQL database.

AlloyDB provides you the ability to get online predictions in your SQL code by calling the google_ml.predict_row() function. For more information about using machine learning (ML) models with AlloyDB, see Build generative AI applications.

Before you begin

Before you can invoke online predictions from an AlloyDB database, you must prepare your database and select an appropriate ML model.

Prepare your database

  1. Set up integration between your database and Vertex AI.

  2. Grant permission for database users to execute the google_ml.predict_row() function to run predictions:

    1. Connect a psql client to the cluster's primary instance, as described in Connect a psql client to an instance.

    2. At the psql command prompt, connect to the database and grant permissions:

      \c DB_NAME

GRANT EXECUTE ON FUNCTION google_ml.predict_row TO USER_NAME;

      Replace the following:

      • DB_NAME: the name of the database on which the permissions should be granted

      • USER_NAME: the name of the user for whom the permissions should be granted

Select an ML model

When you call the google_ml.predict_row() function, you must specify the location of an ML model. The model that you specify can be one of these:

  • A model that's running in the Vertex AI Model Garden.

    The google_ml.predict_row() function supports invoking predictions only on tabular or custom models.

  • A Vertex AI model with an active endpoint that you have Identity and Access Management (IAM) permission to access.

    AlloyDB doesn't support private endpoints for getting online predictions.

Invoke online predictions

Use the google_ml.predict_row() SQL function to invoke online predictions against your data.

The format of the function's initial argument depends on whether the ML model that you want to use is in the Vertex AI Model Garden or is an endpoint running in a Google Cloud project.

Use a model in the Vertex AI Model Garden

To invoke an online prediction using an ML model that's running in the Vertex AI Model Garden, use the following syntax for the google_ml.predict_row() SQL function:

SELECT google_ml.predict_row('projects/PROJECT_ID/locations/REGION_ID/publishers/google/models/MODEL_ID', '{ CONTENTS }');

Replace the following:

  • PROJECT_ID: the ID of your Google Cloud project

  • REGION_ID: the ID of the Google Cloud region that the model is located in—for example, us-central1 for gemini-pro

  • MODEL_ID: the ID of the ML model to use—for example, gemini-pro

  • CONTENTS: the inputs to the prediction call, in JSON format

If the ML model is stored in the same project and region as your AlloyDB cluster, then you can abbreviate this function's first argument:

SELECT google_ml.predict_row('publishers/google/models/MODEL_ID', '{ CONTENTS }');

For information about the model's JSON response messages, see Generative AI foundational model reference.

For examples, see Example invocations.

Use a Vertex AI model endpoint

To invoke an online prediction using a Vertex AI model endpoint, use the following syntax for the google_ml.predict_row() SQL function:

SELECT google_ml.predict_row('projects/PROJECT_ID/locations/REGION_ID/endpoints/ENDPOINT_ID', '{ CONTENTS }');

Replace the following:

  • PROJECT_ID: the ID of the Google Cloud project that the model is located in

  • REGION_ID: the ID of the Google Cloud region the model is located in—for example, us-central1

  • ENDPOINT_ID: the ID of the model endpoint

  • CONTENTS: the inputs to the prediction call, in JSON format

If the endpoint is located in the same project and region as your AlloyDB cluster, then you can abbreviate this function's first argument:

SELECT google_ml.predict_row('endpoints/ENDPOINT_ID', '{ CONTENTS }');

For information about the model's JSON response messages, see PredictResponse.

Example invocations

The following example uses gemini-pro, available in the Model Garden, to generate text based on a short prompt that is provided as a literal argument to google_ml.predict_row():

SELECT google_ml.predict_row('gemini-1.0-pro:generateContent', '{ "contents": [{ "role": "user", "parts": [{ "text": "What is AlloyDB?" }] }] }');

The response is a JSON object. For more information about the format of the object, see Response body.

The next example modifies the previous one in the following ways:

  • The example uses the contents of the current database's messages.message column as input.

  • The example demonstrates the use of the json_build_object() function as an aid to formatting the function parameters.


select google_ml.predict_row('projects/PROJECT_ID/locations/us-central1/publishers/google/models/gemini-1.0-pro:generateContent', json_build_object('contents', json_build_object('text', message))) from messages;

The returned JSON object now contains one entry in its predictions array for every row in the messages table.

Because the response is a JSON object, you can pull specific fields from it using the PostgreSQL arrow operator:

select google_ml.predict_row('projects/PROJECT_ID/locations/us-central1/publishers/google/models/gemini-1.0-pro:generateContent', json_build_object('contents', json_build_object('text', message)))->'predictions'->0->'content' FROM messages;

For more example arguments to google_ml.predict_row(), see Quickstart using the Vertex AI API.