Perform intelligent SQL queries using AI functions

Preview

This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms. Pre-GA features are available "as is" and might have limited support. For more information, see the launch stage descriptions.

AlloyDB AI AI functions integrate Large Language Models (LLMs) like Gemini directly with your AlloyDB for PostgreSQL data to perform intelligent data operations. This feature includes built-in functions for the following:

• Filtering (ai.if)
• Semantic ranking (ai.rank)
• Text generation (ai.generate)

These AI functions use the AlloyDB AI query engine to help your application process data using artificial intelligence models at various scales, from single-row responses to large-scale database operations. You can use the ai.if, ai.rank, and ai.generate operators to combine natural language with SQL queries. AlloyDB AI reserves and creates the ai schema.

Note: Using AI functions with your database can result in charges for processing and model token usage.

There are three categories of AI functions which differ based on how they handle input data and memory allocation: scalar, array-based, and cursor functions. Choosing the right AI function depends on the scale of your data and your performance requirements. Use the following table to learn about these functions and their use cases:

Category	Description	Recommended use case
Scalar	Designed for basic, one-to-one processing. Takes a single input and returns a single output.	Use when you want a basic function that provides acceptable performance for queries that make a small number—less than 50—of scalar function calls.
Array-based	Processes data as an array of rows in a single function call. Takes an array as input and returns array as output.	Use for small-to-medium datasets where the entire array of eligible rows can fit within memory limits. This provides high throughput for set-based operations.
Cursor	Takes a cursor as input and returns a cursor as output.	Use for processing large numbers of rows–for example, 10,000 rows.

Before you begin

• Verify that the google_ml_integration extension is installed and that you're using version 1.5.2 or later. To obtain further performance benefits with array-based functions, use version 1.5.4 or later.
• Verify that the google_ml_integration.enable_ai_query_engine flag is set to on.
• Integrate with Vertex AI.
• Use a Gemini model that's supported in your region. Gemini 2.0 Flash is the AI query engine's default model. If you don't specify a Gemini model in the query, the AI query engine default model is chosen for the queries.

Integrate with Vertex AI and install the extension

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.

   SQL

   1. Enable the AI query engine for the current session.
      

                    SET google_ml_integration.enable_ai_query_engine = on;
                    

   2. Enable features for a specific database across sessions.
      

                    ALTER DATABASE DATABASE_NAME SET google_ml_integration.enable_ai_query_engine = 'on';
                    

   3. Enable the AI query engine for a specific user across sessions and databases.
      

                    ALTER ROLE postgres SET google_ml_integration.enable_ai_query_engine = 'on';
                  

   Console

   To modify the value of the google_ml_integration.enable_ai_query_engine flag, follow the steps in Configure an instance's database flags.

   gcloud

   To use the gcloud CLI, you can install and initialize the Google Cloud CLI, or you can use Cloud Shell.

   You can modify the value of the google_ml_integration.enable_ai_query_engine flag. For more information, see Configure an instance's database flags.

                gcloud alloydb instances update INSTANCE_ID \
                  --database-flags google_ml_integration.enable_ai_query_engine=on \
                  --region=REGION_ID \
                  --cluster=CLUSTER_ID \
                  --project=PROJECT_ID
               

Use a Gemini model that's supported in your region

If your AlloyDB for PostgreSQL cluster is in a region where gemini-2.0-flash isn't supported, you can use one of the other available Gemini models in your region using the model_id parameter.

Note: Verify model availability in your region. For more information, see Gemini 2.0 Flash.

Alternatively, you can register a Gemini model endpoint and provide that model ID to the AI operators. For more information, see Register and call remote AI models using model endpoint management.

The following example shows how to register another Gemini endpoint. In this example, this second Gemini endpoint is the global endpoint for gemini-2.0-flash. You can use this registered model with AI operators by passing model_id =>gemini-2.5-flash-global` as an additional argument.

CALL
  google_ml.create_model(
    model_id => 'gemini-2.5-flash-global',
    model_type => 'llm',
    model_provider => 'google',
    model_qualified_name => 'gemini-2.5-flash',
    model_request_url =>  'https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/google/models/gemini-2.5-flash:generateContent',
    model_auth_type => 'alloydb_service_agent_iam'
);

Use Gemini 3.0 models

Some Gemini models, such as gemini-3.0-pro-preview, are available only through the global endpoint. You must register such models as follows:

CALL
  google_ml.create_model(
    model_id => 'gemini-3-preview-model',
    model_request_url => 'https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/google/models/gemini-3-pro-preview:generateContent',
    model_qualified_name => 'gemini-3-pro-preview',
    model_provider => 'google',
    model_type => 'llm',
    model_auth_type => 'alloydb_service_agent_iam'
);

Replace PROJECT_ID with the ID of the project where the Vertex AI model is available. Note that the AlloyDB service account must have the Vertex AI User Role in that project.

After registering the model, you can use it in AI functions as follows:

SELECT ai.generate(prompt => 'What is AlloyDB?', model_id => 'gemini-3-preview-model');

Use filters in your queries

AlloyDB AI offers AI-powered SQL functions that let you use natural language processing and LLMs directly within your database queries, including the ai.if and ai.rank operators.

Note: if you don't specify a batch size in your queries, the default batch size is 30.

Perform filtering using scalar functions

To evaluate whether a condition stated in natural language is met, use the ai.if/google_ml.if operator. The function returns boolean true or false value, and returns false if the output isn't clearly detected.

- Function signature
FUNCTION ai.if(prompt TEXT, model_id VARCHAR(100) DEFAULT NULL) RETURNS bool

The following example shows the use of the ai.if operator as a filter to find restaurants with more than 500 positive reviews located in cities with a population greater than 100,000. The example uses restaurant_reviews and contains data like reviews and city location. The ai.if operator helps you understand the review sentiment and combine the locations from the database with Gemini general knowledge about the population in these locations.

SELECT r.name, r.location_city
FROM restaurant_reviews r
WHERE
  AI.IF(r.location_city || ' has a population OF more than 100,000 AND the following is a positive review; Review: ' || r.review)
GROUP BY r.name, r.location_city
HAVING COUNT(*) > 500;

The following shows the same example using the model that you registered in Use a Gemini model that's supported in your region.

SELECT r.name, r.location_city
FROM restaurant_reviews r
WHERE
  AI.IF(r.location_city || ' has a population of more than 100,000 AND the following is a positive review; Review: ' || r.review, model_id => 'gemini-2.5-flash')
GROUP BY r.name, r.location_city
HAVING COUNT(*) > 500;

Perform a join on a query that uses the if operator

To perform a join operation, use the ai.if/google_ml.if operator with join. The following sample query finds the number of reviews that mention each menu item from the restaurant menu.

    SELECT item_name, COUNT(*)
    FROM menu_items JOIN user_reviews
      ON ai.if(
        prompt => 'Does the following user review talk about the menu item mentioned ? review: ' || user_reviews.review_text || ' menu item: ' || item_name)
    GROUP BY item_name;

Perform filtering using array-based functions

The following example identifies restaurants with more than 10 positive reviews by using an AI model (gemini-2.0-flash) to analyze review sentiment and filter the results. ARRAY_AGG is used to transform individual rows of data into structured arrays so they can be processed by the AI model in bulk rather than one row at a time.

WITH initial_arrays AS (WITH initial_arrays AS (
  SELECT
    ARRAY_AGG(r.id ORDER BY r.id) AS review_ids,
    -- Assuming ai.if takes an array of prompts and returns a boolean array
    ai.if(
      prompts => ARRAY_AGG('Is the review positive: ' || r.review ORDER BY r.id)
      model_id => 'gemini-2.5-flash',
      batch_size => 20
    ) AS truth_values
  FROM restaurant_reviews r
),
reviews AS (
SELECT
  initial_arrays.review_ids[i] AS review_id,
  initial_arrays.truth_values[i] AS truth_value
FROM
  initial_arrays,
  generate_series(1, array_length(initial_arrays.review_ids, 1)) AS i
)
SELECT rest_review.city, rest_review.name
FROM restaurant_reviews rest_review JOIN reviews review ON rest_review.id=review.review_id
WHERE review.truth_value = 't'
GROUP BY rest_review.city, rest_review.name
HAVING COUNT(*) > 10;

Perform filtering using cursors

The following example shows how to filter a large set of restaurant reviews by streaming them through a cursor.

Note: if you don't specify a batch size, the default batch size is 30.

CREATE TABLE filtered_results(input text, result bool);

DO $$
DECLARE
    prompt_cursor REFCURSOR;
    result_cursor REFCURSOR;
    rec RECORD;
BEGIN
    -- 1. Open a cursor for the input data
    OPEN prompt_cursor FOR
        SELECT r.location_city || ' has a population of > 100,000 and is a positive review; Review: ' || r.review
        FROM restaurant_reviews r;

    -- 2. Call the array-based function using the input cursor
    result_cursor := ai.if(
        'Is the given statement true? ',
        prompt_cursor,
        model_id => 'gemini-2.5-flash'
    );

    -- 3. Fetch results from the output cursor and store them
    LOOP
        FETCH result_cursor INTO rec;
        EXIT WHEN NOT FOUND;
        INSERT INTO filtered_results VALUES(rec.input, rec.output);
    END LOOP;

    CLOSE result_cursor;
END $$;

Text generation and summarization

AlloyDB AI offers text generation operators like scalar ai.generate and array-based and cursor-based ai.generate.

Perform text generation using scalar functions

The ai.generate function produces text by combining provided data with the user's prompt.

-- Function Signature
FUNCTION ai.generate(prompt TEXT, model_id VARCHAR(100) DEFAULT NULL) RETURNS TEXT

For example, you can use the following query to generate a concise summary of each user review.

SELECT
  ai.generate(
    prompt => 'Summarize the review in 20 words or less. Review: ' || review) AS review_summary
FROM user_reviews;

Perform text generation using array-based functions

The following query uses UNNEST and ai.generate to summarize multiple reviews efficiently.

SELECT
  UNNEST(
    ai.generate(
      prompts => ARRAY_AGG('Summarize the review in 20 words or less. Review: ' || review),
      model_id => 'gemini-2.5-flash',
    )
  ) AS review_summary
FROM user_reviews;

Perform text generation using cursors

To generate summaries or translations for millions of rows without hitting memory bottlenecks, use batch generation with cursors.

CREATE TABLE summary_results(summary text);

DO $$
DECLARE
    prompt_cursor REFCURSOR;
    summary_cursor REFCURSOR;
    rec RECORD;
BEGIN
    OPEN prompt_cursor FOR SELECT review_text FROM user_reviews ORDER BY id;

    summary_cursor := ai.generate(
        'Summarize the review in 20 words or less. Review:',
        prompt_cursor,
    );

    LOOP
        FETCH summary_cursor INTO rec;
        EXIT WHEN NOT FOUND;
        INSERT INTO summary_results VALUES(rec.output);
    END LOOP;

    CLOSE summary_cursor;
END $$;

Score your query results

If you need to sort query results using custom, natural language instructions, use the ai.rank operator.

Perform scoring using scalar functions

The following function lets you provide a prompt which describes the ranking criteria and returns a score for each item.

-- Function signature
FUNCTION ai.rank(prompt TEXT, model_id VARCHAR(100) DEFAULT NULL) RETURNS real

For example, the following query gets the top 20 most positive restaurant reviews, using scores from an LLM.

SELECT review AS top20
FROM user_reviews
ORDER BY ai.rank(
  'Score the following review according to these rules:
  (1) Score OF 8 to 10 IF the review says the food IS excellent.
  (2) 4 to 7 IF the review says the food is ok.
  (3) 1 to 3 IF the review says the food is not good. Here is the review:' || review) DESC
LIMIT 20;

Perform scoring using array-based functions

This ai.rank function lets you score and rank query results based on custom natural language instructions.

FUNCTION ai.rank(prompts TEXT[], model_id VARCHAR(100) DEFAULT NULL) RETURNS real[]

The following query uses UNNEST and ai.rank to score multiple reviews efficiently.

SELECT
  UNNEST(
    ai.rank(
      ARRAY_AGG('Score the following review according to these rules:
  (1) Score OF 8 to 10 IF the review says the food IS excellent.
  (2) 4 to 7 IF the review says the food is ok.
  (3) 1 to 3 IF the review says the food is not good. Here is the review:' || review),
    )
  ) as review_scores
FROM user_reviews;

Perform scoring using cursors

This function is used for scoring large datasets without hitting memory limits.

FUNCTION ai.rank(context TEXT, input_cursor REFCURSOR, model_id VARCHAR(100) DEFAULT NULL) RETURNS REFCURSOR

The following example shows you how to score large volumes of text according to specific natural language criteria.

CREATE TABLE scored_results(input text, score real);

DO $$
DECLARE
    prompt_cursor REFCURSOR;
    score_cursor REFCURSOR;
    rec RECORD;
BEGIN
    OPEN prompt_cursor FOR SELECT review FROM user_reviews ORDER BY id;

    score_cursor := ai.rank(
        'Score the following review: (1) 8-10 if excellent, (2) 4-7 if ok, (3) 1-3 if not good. Review:',
        prompt_cursor,
    );

    LOOP
        FETCH score_cursor INTO rec;
        EXIT WHEN NOT FOUND;
        INSERT INTO scored_results VALUES(rec.input, rec.output);
    END LOOP;

    CLOSE score_cursor;
END $$;

What's next

• Register a model endpoint with model endpoint management.

• Rank and score search results for RAG.

• Semantically filter SQL queries and rank results.

• Perform time-series forecasting.