Generate and manage vector embeddings

This page describes how to generate and store vector embeddings based on a model.

For more information, see Build generative AI applications using Cloud SQL.

Cloud SQL lets you use an embedding model hosted by Vertex AI to translate a text string into an embedding, which is the model's representation of the given text's semantic meaning as a numeric vector.

Cloud SQL implements embeddings as arrays of real values. You can use generated embeddings as inputs for pgvector extension functions.

Before you begin

Some requirements differ depending on whether you want to use Cloud SQL to generate embeddings, or whether you only need to work with embeddings that are stored in your database from another source.

Regional restrictions

To generate Vertex AI embeddings with Cloud SQL, your instance must reside in a region where generative AI foundational models are supported. The Vertex AI models that Cloud SQL can use for embeddings, text-embedding and textembedding-gecko, are located in that region.

Required database extensions

To work with embeddings, install the google_ml_integration extension on your Cloud SQL instance. For Vertex AI models, install version 1.2 or later. For third-party or custom models, install version 1.4.2 or later.

Optionally, if you want to store these embeddings, and use vector functions and operators with the embeddings, then you also need the pgvector extension.

Cloud SQL has both of these extensions. You can install them on any database in your instance. For more information, see Configure PostgreSQL extensions.

Set up model access

Before you can generate embeddings from a Cloud SQL instance, you must configure Cloud SQL to work with a text embedding model.

To work with the cloud-based text-embedding or textembedding-gecko model, you need to integrate Cloud SQL with Vertex AI.

Grant database users access to generate embeddings

Grant permission for database users to use the embedding function to run predictions:

  1. Connect a psql client to the primary instance, as described in Connect using a psql client.

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

    \c DB_NAME

GRANT EXECUTE ON FUNCTION embedding TO USER_NAME;

    Replace the following:

    • DB_NAME: the name of the database for which you're granting permissions

    • USER_NAME: the name of the user for whom you're granting permissions

Generate embeddings

Cloud SQL provides a function that lets you translate text into a vector embedding. You can then store that embedding in your database as vector data, and optionally use pgvector functions to run queries on it.

Generate an embedding

To generate an embedding using Cloud SQL, use the embedding function that the google_ml_integration extension provides:

SELECT embedding( 'MODEL_IDVERSION_TAG', 'TEXT');

Make the following replacements:

  • MODEL_ID: the ID of the model to query.

    If you're using the Vertex AI Model Garden, then specify text-embedding-004 or
    text-multilingual-embedding-002. These are the cloud-based models that Cloud SQL can use for text embeddings. For more information, see Text embeddings.

  • VERSION_TAG (optional): the version tag of the model to query. For versions of textembedding-gecko prior to text-embedding-004 or text-multilingual-embedding-002, prepend the tag with @.

    If you're using one of the textembedding-gecko models with Vertex AI, then specify one of the version tags listed in Model versions.

  • TEXT: the text to translate into a vector embedding.

The following example uses the text-embedding-004 model to generate an embedding based on a provided literal string:

SELECT embedding( 'text-embedding-004', 'Cloud SQL is a managed, cloud-hosted SQL database service.');

Store a generated embedding

The return value of the embedding() function is an array of real values. To store this value in a table, add a real[] column:

ALTER TABLE TABLE ADD COLUMN EMBEDDING_COLUMN real[DIMENSIONS];

Make the following replacements:

  • TABLE: the table name

  • EMBEDDING_COLUMN: the name of the new embedding column

  • DIMENSIONS: the number of dimensions that the model supports.

    If you're using one of the text-embedding or textembedding-gecko models with Vertex AI, then specify 768.

Optionally, if you have installed the pgvector extension, then you can store embeddings as vector values:

ALTER TABLE TABLE ADD COLUMN EMBEDDING_COLUMN vector(DIMENSIONS);

After you create a column to store embeddings, you can populate it based on the values already stored in another column in the same table:

UPDATE TABLE SET EMBEDDING_COLUMN = embedding('MODEL_IDVERSION_TAG', SOURCE_TEXT_COLUMN);

Make the following replacements:

  • TABLE: the table name.

  • EMBEDDING_COLUMN: the name of the embedding column.

  • MODEL_ID: the ID of the model to query.

    If you're using the Vertex AI Model Garden, then specify text-embedding-004 or
    text-multilingual-embedding-002. These are the cloud-based models that Cloud SQL can use for text embeddings. For more information, see Text embeddings.

  • VERSION_TAG (Optional): the version tag of the model to query. For versions of textembedding-gecko prior to text-embedding-004 or text-multilingual-embedding-002,Prepend the tag with@`.

    If you're using one of the textembedding-gecko models with Vertex AI, then specify one of the version tags listed in Model versions.

  • SOURCE_TEXT_COLUMN: the name of the column that's storing the text. You translate this text into embeddings.

The previous command works for both real[] and vector embedding columns. If your embedding column is of the vector type, then Cloud SQL casts the return value of embedding() from a real array to a vector value implicitly.

The following example uses the text-embedding-004 model to populate the messages.message_vector column with embeddings based on the content of the messages.message column:

UPDATE messages SET message_vector = embedding( 'text-embedding-004', message);

Query and index embeddings using pgvector

The pgvector PostgreSQL extension lets you use vector-specific operators and functions when you store, index, and query text embeddings in your database. Cloud SQL has its own optimizations for working with pgvector, letting you create indexes that can speed up queries that involve embeddings.

Create a nearest-neighbor index

pgvector supports approximate nearest-neighbor (ANN) searches through indexing.

To create an HNSW index, use the hnsw function, as shown in the following example:

CREATE INDEX ON TABLE
  USING hnsw (EMBEDDING_COLUMN DISTANCE_FUNCTION)
  WITH (m = M, ef_construction = EF_CONSTRUCTION);

Make the following replacements:

  • TABLE: the table to which you're adding the index.

  • EMBEDDING_COLUMN: a column that stores vector data.

  • DISTANCE_FUNCTION: the distance function to use with this index. Choose one of the following:

    • L2 distance: vector_l2_ops

    • Inner product: vector_ip_ops

    • Cosine distance: vector_cosine_ops

  • M (optional): the maximum number of connections with neighboring data points in a graph. We recommend a range of 5 to 48. The default is 16.

  • EF_CONSTRUCTION (optional): the size of the list that holds the closest candidates during the graph traversal when building the index. Higher values lead the algorithm to consider more candidates, allowing a better index to be created. The default size is 64.

To create this index on an embedding column that uses the real[] data type instead of vector, cast the column into the vector data type:

CREATE INDEX ON TABLE
  USING hnsw ((CAST(EMBEDDING_COLUMN AS vector(DIMENSIONS))) DISTANCE_FUNCTION)
  WITH (m = M, ef_construction = EF_CONSTRUCTION);

Replace DIMENSIONS with the dimensional width of the embedding column.

The next section demonstrates an example of this kind of index.

Make a nearest-neighbor query with given text

After you store and index embeddings in your database, the full range of pgvector query functionality is available to you.

To find the nearest semantic neighbors to a piece of text, use the embedding() function to translate the text into a vector. In the same query, apply this vector to the pgvector nearest-neighbor operator, <->, to find the database rows with the most semantically similar embeddings.

Because embedding() returns a real array, you must cast the embedding() call to vector to use these values with pgvector operators.

  SELECT RESULT_COLUMNS FROM TABLE
    ORDER BY EMBEDDING_COLUMN
    <-> embedding('MODEL_IDVERSION_TAG', 'TEXT')::vector
    LIMIT ROW_COUNT

Make the following replacements:

  • RESULT_COLUMNS: the columns to display from semantically similar rows.

  • TABLE: the table containing the embedding to which you compare the text.

  • EMBEDDING_COLUMN: the column containing the stored embeddings.

  • MODEL_ID: the ID of the model to query.

    If you're using the Vertex AI Model Garden, then specify text-embedding-004 or
    text-embedding-multilingual-002. These are the cloud-based models that Cloud SQL can use for text embeddings. For more information, see Text embeddings.

  • VERSION_TAG (Optional): the version tag of the model to query. Prepend the tag with @.

    If you're using one of the textembedding-gecko models with Vertex AI, then specify one of the version tags listed in Model versions.

  • TEXT: the text that you want so that you can find the nearest stored semantic neighbors.

  • ROW_COUNT: the number of rows to return. If you want only the single best match, then specify 1 as the value for this parameter.

To run this query with a stored embedding column that uses the real[] data type instead of vector, cast the column into the vector data type:

  SELECT RESULT_COLUMNS FROM TABLE
    ORDER BY EMBEDDING_COLUMN::vector
    <-> embedding('MODEL_IDVERSION_TAG', 'TEXT')::vector
    LIMIT ROW_COUNT

Use model version tags to avoid errors

Cloud SQL strongly recommends that you always use a stable version of your chosen embeddings model. For most models, this means setting a version tag explicitly.

Calling the embedding() function without specifying the version tag of the model is valid syntactically, but it can be prone to error.

If you omit the version tag when using a model in the Vertex AI Model Garden, then Vertex AI uses the latest version of the model. This might not be the latest stable version. For more information about available Vertex AI model versions, see Model versions.

A given Vertex AI model version always returns the same embedding() response to a given text input. If you don't specify the model version in your embedding calls, then it's possible that a newly published model version could abruptly change the returned vector for a given input. This can cause errors or other unexpected behavior in your applications.

What's next