llm_embedding

The llm_embedding function generates vector embeddings that represent the semantic meaning of text from specified table columns.

Modalities

Text

Supported. Pass one or more text columns in context_columns.

SELECT llm_embedding(
  {'model_name': 'text-embedding-3-small'},
  {'context_columns': [{'data': product_name}, {'data': product_description}]}
) AS product_embedding
FROM VALUES
  ('Wireless Headphones', 'Premium noise-cancelling headphones with 30-hour battery life'),
  ('Gaming Laptop', 'High-performance laptop with RTX graphics and 16GB RAM')
AS t(product_name, product_description);

Image

Not supported directly. Describe the image with llm_complete (type: 'image'), then embed the description.

WITH image_descriptions AS (
  SELECT
    image_id,
    llm_complete(
      {'model_name': 'gpt-4o'},
      {
        'prompt': 'Describe this image in detail.',
        'context_columns': [{'data': image_url, 'type': 'image'}]
      }
    ) AS generated_description
  FROM image_catalog
)
SELECT
  image_id,
  llm_embedding(
    {'model_name': 'text-embedding-3-small'},
    {'context_columns': [{'data': generated_description}]}
  ) AS image_embedding
FROM image_descriptions;

Voice

Not supported on raw audio. Transcribe with llm_complete (type: 'audio'), then embed the transcript.

WITH transcripts AS (
  SELECT
    audio_id,
    llm_complete(
      {'model_name': 'gpt-4o'},
      {
        'prompt': 'Transcribe the following audio file.',
        'context_columns': [{
          'data': file_path,
          'type': 'audio',
          'transcription_model': 'whisper-1'
        }]
      }
    ) AS transcript
  FROM VALUES
    (1, '/data/audio/note_01.m4a'),
    (2, '/data/audio/note_02.m4a')
  AS t(audio_id, file_path)
)
SELECT
  audio_id,
  llm_embedding(
    {'model_name': 'text-embedding-3-small'},
    {'context_columns': [{'data': transcript}]}
  ) AS embedding
FROM transcripts;

1. Simple Usage (without data)

1.1 Basic Embedding Generation

SELECT llm_embedding(
    {'model_name': 'text-embedding-3-small', 'secret_name': 'embedding_secret'},
    {'context_columns': [{'data': product_name}, {'data': product_description}]}
) AS product_embedding
FROM VALUES
    ('Wireless Headphones', 'Premium noise-cancelling headphones with 30-hour battery life'),
    ('Gaming Laptop', 'High-performance laptop with RTX graphics and 16GB RAM'),
    ('Smart Watch', 'Fitness tracker with heart rate monitor and GPS')
AS t(product_name, product_description);

Description: This example generates vector embeddings for each product, combining the product_name and product_description columns using the text-embedding-3-small model. The output is a semantic vector that represents the content of the product’s name and description.

1.2 Similarity Search

WITH sample_products AS (
    SELECT *
    FROM (VALUES
        (1, 'Wireless Headphones', 'Premium noise-cancelling headphones with 30-hour battery life'),
        (2, 'Bluetooth Earbuds', 'Compact wireless earbuds with charging case'),
        (3, 'Gaming Laptop', 'High-performance laptop with RTX graphics and 16GB RAM'),
        (4, 'Office Laptop', 'Lightweight laptop perfect for business and productivity')
    ) AS t(product_id, product_name, product_description)
),
product_embeddings AS (
    SELECT
        product_id,
        product_name,
        llm_embedding(
            {'model_name': 'text-embedding-3-small'},
            {'context_columns': [{'data': product_name}, {'data': product_description}]}
        ) AS product_embedding
    FROM sample_products
)
SELECT
    a.product_name,
    b.product_name,
    array_cosine_similarity(a.product_embedding::DOUBLE[1536], b.product_embedding::DOUBLE[1536]) AS similarity
FROM product_embeddings a
JOIN product_embeddings b
ON a.product_id != b.product_id
WHERE similarity > 0.8;

Description: This example demonstrates how to use the vector embeddings for similarity search. It calculates the cosine similarity between embeddings of different products to find similar items based on their semantic meaning. Only product pairs with a similarity greater than 0.8 are included.

2. Input Parameters

The llm_embedding function accepts two primary inputs: model configuration and column mappings.

2.1 Model Configuration

Parameter: model_name and secret_name

2.1.1 Model Selection

Description: Specifies the model used for text generation.
Example:
{ 'model_name': 'gpt-4o' }

2.1.2 Model Selection with Secret

Description: Specifies the model along with the secret name to be used for authentication when accessing the model.

Example:

{ 'model_name': 'gpt-4o', 'secret_name': 'your_secret_name' }

2.2 Context Columns Configuration

Parameter: context_columns array
Description: Specifies the text columns from the table to be passed to the model for embedding generation. Each column can have two properties:
- data: The SQL column data (required)
- name: Custom name for the column (optional)
Note: Embeddings are text-only. For image or voice content, derive text first (see Modalities above).

Example:

{ 'context_columns': [{'data': product_name}, {'data': product_description}] }

3. Output

The function returns a JSON array containing floating-point numbers that represent the semantic vector of the input text. Example Output:
For a product with the description “Wireless headphones with noise cancellation”, the output might look like this:

[0.342, -0.564, 0.123, ..., 0.789]

This array of floating-point numbers encodes the semantic meaning of the product description in high-dimensional space.

Home

Getting Started

Resource Management

Scalar Functions

Aggregate Functions

Hybrid Search

Operations

Contributing

llm_embedding

Modalities

Text

Image

Voice

1. Simple Usage (without data)

1.1 Basic Embedding Generation

1.2 Similarity Search

2. Input Parameters

2.1 Model Configuration

2.1.1 Model Selection

2.1.2 Model Selection with Secret

2.2 Context Columns Configuration

3. Output

​Modalities

​Text

​Image

​Voice

​1. Simple Usage (without data)

​1.1 Basic Embedding Generation

​1.2 Similarity Search

​2. Input Parameters

​2.1 Model Configuration

​2.1.1 Model Selection

​2.1.2 Model Selection with Secret

​2.2 Context Columns Configuration

​3. Output

Modalities

Text

Image

Voice

1. Simple Usage (without data)

1.1 Basic Embedding Generation

1.2 Similarity Search

2. Input Parameters

2.1 Model Configuration

2.1.1 Model Selection

2.1.2 Model Selection with Secret

2.2 Context Columns Configuration

3. Output