This page documents AlloyDB Omni version 18.1.0 using the containers deployment option. Choose a different deployment option.

Perform vector search in AlloyDB Omni

Select a documentation version:

This tutorial describes how to set up and perform a vector search in AlloyDB Omni using the Google Cloud console. Examples are included to show vector search capabilities, and they're intended for demonstration purposes only.

For information about how to use filtered vector search to refine your similarity searches, see Filtered vector search in AlloyDB Omni.

To learn how to perform a vector search with Vertex AI embeddings, see Getting started with Vector Embeddings with AlloyDB Omni AI.

Objectives

Create an AlloyDB Omni cluster and primary instance.
Connect to your database and install required extensions.
Create a product and product inventory table.
Insert data to the product and product inventory tables and perform a basic vector search.
Create a ScaNN index on the products table.
Perform a basic vector search.
Perform a complex vector search with a filter and a join.

Costs

In this document, you use the following billable components of Google Cloud:

Agent Platform API

To generate a cost estimate based on your projected usage, use the pricing calculator.

New Google Cloud users might be eligible for a free trial.

When you finish the tasks that are described in this document, you can avoid continued billing by deleting the resources that you created. For more information, see Clean up.

Prerequisites

Complete the following prerequisites before performing a vector search:

Install AlloyDB AI in AlloyDB Omni
Run and connect to AlloyDB Omni as the postgres user.

Install the vector, alloydb_scann, and the google_ml_integration extensions.

CREATE EXTENSION IF NOT EXISTS vector;
CREATE EXTENSION IF NOT EXISTS alloydb_scann;
CREATE EXTENSION IF NOT EXISTS google_ml_integration CASCADE;

Insert product and product inventory data and perform a basic vector search

Run the following statement to create a product table that does the following:
- Stores basic product information.
- Includes an embedding vector column that computes and stores an embedding vector for a product description of each product.
Note: If you have more than 100k rows in a table, we don't recommend using the embedding() function to generate stored embeddings on existing data in a table. It is best suited for inline embedding generation.
```
  CREATE TABLE product (
    id INT PRIMARY KEY,
    name VARCHAR(255) NOT NULL,
    description TEXT,
    category VARCHAR(255),
    color VARCHAR(255),
    embedding vector(768) GENERATED ALWAYS AS (embedding('text-embedding-005', description)) STORED
  );
```
If needed, you can view the logs and troubleshoot errors.
Run the following query to create a product_inventory table that stores information about available inventory and corresponding prices. The product_inventory and product tables are used in this tutorial to run complex vector search queries.
```
CREATE TABLE product_inventory (
  id INT PRIMARY KEY,
  product_id INT REFERENCES product(id),
  inventory INT,
  price DECIMAL(10,2)
);
```

Run the following query to insert product data into the product table:

INSERT INTO product (id, name, description,category, color) VALUES
(1, 'Stuffed Elephant', 'Soft plush elephant with floppy ears.', 'Plush Toys', 'Gray'),
(2, 'Remote Control Airplane', 'Easy-to-fly remote control airplane.', 'Vehicles', 'Red'),
(3, 'Wooden Train Set', 'Classic wooden train set with tracks and trains.', 'Vehicles', 'Multicolor'),
(4, 'Kids Tool Set', 'Toy tool set with realistic tools.', 'Pretend Play', 'Multicolor'),
(5, 'Play Food Set', 'Set of realistic play food items.', 'Pretend Play', 'Multicolor'),
(6, 'Magnetic Tiles', 'Set of colorful magnetic tiles for building.', 'Construction Toys', 'Multicolor'),
(7, 'Kids Microscope', 'Microscope for kids with different magnification levels.', 'Educational Toys', 'White'),
(8, 'Telescope for Kids', 'Telescope designed for kids to explore the night sky.', 'Educational Toys', 'Blue'),
(9, 'Coding Robot', 'Robot that teaches kids basic coding concepts.', 'Educational Toys', 'White'),
(10, 'Kids Camera', 'Durable camera for kids to take pictures and videos.', 'Electronics', 'Pink'),
(11, 'Walkie Talkies', 'Set of walkie talkies for kids to communicate.', 'Electronics', 'Blue'),
(12, 'Karaoke Machine', 'Karaoke machine with built-in microphone and speaker.', 'Electronics', 'Black'),
(13, 'Kids Drum Set', 'Drum set designed for kids with adjustable height.', 'Musical Instruments', 'Blue'),
(14, 'Kids Guitar', 'Acoustic guitar for kids with nylon strings.', 'Musical Instruments', 'Brown'),
(15, 'Kids Keyboard', 'Electronic keyboard with different instrument sounds.', 'Musical Instruments', 'Black'),
(16, 'Art Easel', 'Double-sided art easel with chalkboard and whiteboard.', 'Arts & Crafts', 'White'),
(17, 'Finger Paints', 'Set of non-toxic finger paints for kids.', 'Arts & Crafts', 'Multicolor'),
(18, 'Modeling Clay', 'Set of colorful modeling clay.', 'Arts & Crafts', 'Multicolor'),
(19, 'Watercolor Paint Set', 'Watercolor paint set with brushes and palette.', 'Arts & Crafts', 'Multicolor'),
(20, 'Beading Kit', 'Kit for making bracelets and necklaces with beads.', 'Arts & Crafts', 'Multicolor'),
(21, '3D Puzzle', '3D puzzle of a famous landmark.', 'Puzzles', 'Multicolor'),
(22, 'Race Car Track Set', 'Race car track set with cars and accessories.', 'Vehicles', 'Multicolor'),
(23, 'RC Monster Truck', 'Remote control monster truck with oversized tires.', 'Vehicles', 'Green'),
(24, 'Train Track Expansion Set', 'Expansion set for wooden train tracks.', 'Vehicles', 'Multicolor');

Optional: Run the following query to verify that the data is inserted in the product table:
```
SELECT * FROM product;
```

Run the following query to insert inventory data into the product_inventory table:

INSERT INTO product_inventory (id, product_id, inventory, price) VALUES
(1, 1, 9, 13.09),
(2, 2, 40, 79.82),
(3, 3, 34, 52.49),
(4, 4, 9, 12.03),
(5, 5, 36, 71.29),
(6, 6, 10, 51.49),
(7, 7, 7, 37.35),
(8, 8, 6, 10.87),
(9, 9, 7, 42.47),
(10, 10, 3, 24.35),
(11, 11, 4, 10.20),
(12, 12, 47, 74.57),
(13, 13, 5, 28.54),
(14, 14, 11, 25.58),
(15, 15, 21, 69.84),
(16, 16, 6, 47.73),
(17, 17, 26, 81.00),
(18, 18, 11, 91.60),
(19, 19, 8, 78.53),
(20, 20, 43, 84.33),
(21, 21, 46, 90.01),
(22, 22, 6, 49.82),
(23, 23, 37, 50.20),
(24, 24, 27, 99.27);

Run the following vector search query that tries to find products that are similar to the word music. This means that even though the word music isn't explicitly mentioned in the product description, the result shows products that are relevant to the query:
```
SELECT * FROM product
ORDER BY embedding <=> embedding('text-embedding-005', 'music')::vector
LIMIT 3;
```
The result of the query is as follows:

Performing a basic vector search without creating an index uses exact nearest neighbor search (KNN), which provides efficient recall. At scale, using KNN might impact performance. For a better query performance, we recommend that you use the ScaNN index for approximate nearest neighbor (ANN) search, which provides high recall with low latencies.

Without creating an index, AlloyDB Omni defaults to using exact nearest-neighbor search (KNN).

To learn more about using ScaNN at scale, see Getting started with Vector Embeddings with AlloyDB AI.

Create a manually-tuned ScaNN index on products table

Run the following query to create a product_index ScaNN index on the product table:

CREATE INDEX product_index ON product
USING scann (embedding cosine)
WITH (mode='MANUAL', num_leaves=4);

For more information on creating a ScaNN index, see Create a ScaNN index.

Perform a vector search

Run the following vector search query that tries to find products that are similar to the natural language query music. Even though the word music isn't included in the product description, the result shows products that are relevant to the query:

SET LOCAL scann.num_leaves_to_search = 2;

SELECT * FROM product
ORDER BY embedding <=> embedding('text-embedding-005', 'music')::vector
  LIMIT 3;

The query results are as follows:

The scann.num_leaves_to_search query parameter controls the number of leaf nodes that are searched during a similarity search. The num_leaves and scann.num_leaves_to_search parameter values help to achieve a balance of performance and recall.

Perform a vector search that uses a filter and a join

You can run filtered vector search queries efficiently even when you use the ScaNN index. Run the following complex vector search query, which returns relevant results that satisfy the query conditions, even with filters:

SET LOCAL scann.num_leaves_to_search = 2;

SELECT * FROM product p
JOIN product_inventory pi ON p.id = pi.product_id
WHERE pi.price < 80.00
ORDER BY embedding <=> embedding('text-embedding-005', 'music')::vector
LIMIT 3;

Accelerate your filtered vector search

You can use the columnar engine content store to improve the performance of vector similarity searches, specifically K-Nearest Neighbor (KNN) searches, when combined with highly selective predicate filtering —for example, using LIKE— in databases. In this section, you use the vector extension and the AlloyDB Omni google_columnar_engine extension. For more information on how the columnar engine works, see Columnar engine overview.

Performance improvements come from the columnar engine's built-in efficiency in scanning large datasets and applying filters —such as LIKE predicates— coupled with its ability, using vector support, to pre-filter rows. This functionality reduces the number of data subsets required for subsequent KNN vector distance calculations, and it helps to optimize complex analytical queries involving standard filtering and vector search.

The columnar store offers two options to manage its content:

Automatically manage the column store content: new AlloyDB Omni instances use auto-columnarization by default. Alternatively, you can manually run the auto columnarization functionality.
Manage column store content manually: if you need to manually manage the columns in the column store for your workload, you can disable auto columnarization.

To compare the execution time of a KNN vector search filtered by a LIKE predicate before and after you enable the columnar engine, follow these steps:

Enable the vector extension to support vector data types and operations. Run the following statements to create an example table (items) with an ID, a text description, and a 512-dimension vector embedding column.
```
CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE items (
    id SERIAL PRIMARY KEY,
    description TEXT,
    embedding VECTOR(512)
);
```

Populate the data by running the following statements to insert 1 million rows into the example items table.

-- Simplified example of inserting matching (~0.1%) and non-matching data
INSERT INTO items (description, embedding)
SELECT
    CASE WHEN g % 1000 = 0 THEN 'product_' || md5(random()::text) || '_common' -- ~0.1% match
    ELSE 'generic_item_' || g || '_' || md5(random()::text)    -- ~99.9% don't match
    END,
    (SELECT array_agg(random()) FROM generate_series(1, 512))::vector
FROM generate_series(1, 999999) g;

Measure the baseline performance of the vector similarity search without the columnar engine.

SELECT id, description, embedding <-> '[...]' AS distance
FROM items
WHERE description LIKE '%product_%_common%'
ORDER BY embedding <-> '[...]'
LIMIT 100;

Enable columnar engine and vector support.
1. Enable the google_columnar_engine.enabled and google_columnar_engine.enable_vector_support database flags.
```
ALTER SYSTEM SET google_columnar_engine.enabled = 'on';
ALTER SYSTEM SET google_columnar_engine.enable_vector_support = 'on';
```
2. Restart AlloyDB Omni
  Docker
  
  docker restart CONTAINER_NAME
  Docker
  
  docker restart CONTAINER_NAME
  Podman
  
  podman restart CONTAINER_NAME
  Podman
  
  podman restart CONTAINER_NAME
  Replace CONTAINER_NAME with the name of the container you created in Install AlloyDB Omni in containers.

Add the items table to the columnar engine:

SELECT google_columnar_engine_add('items');

Measure the performance of the vector similarity search using the columnar engine. You re-run the query that you previously ran to measure baseline performance.

SELECT id, description, embedding <-> '[...]' AS distance
FROM items
WHERE description LIKE '%product_%_common%'
ORDER BY embedding <-> '[...]'
LIMIT 100;

To check whether the query ran with the columnar engine, run the following command:

explain (analyze) SELECT id, description, embedding <-> '[...]' AS distance
FROM items
WHERE description LIKE '%product_%_common%'
ORDER BY embedding <-> '[...]'
LIMIT 100;

Clean up

To uninstall AlloyDB Omni, see Manage and monitor AlloyDB Omni.

What's next

Learn how to build generative AI applications using AlloyDB AI.
Create a ScaNN index.