Cortex AISQL Images¶

Input requirements¶

COMPLETE Multimodal can process images with the following characteristics:

Create a stage for image processing¶

Create a stage suitable for storing the images to be processed. The stage must have a directory table and server-side encryption.

The SQL below creates a suitable internal stage.

CREATE OR REPLACE STAGE input_stage
    DIRECTORY = ( ENABLE = true )
    ENCRYPTION = ( TYPE = 'SNOWFLAKE_SSE' );

The following SQL creates an external stage on Amazon S3. Azure and GCP external stages are also supported.

CREATE OR REPLACE STAGE input_stage
    URL='s3://<s3-path>/'
    DIRECTORY = ( ENABLE = true )
    CREDENTIALS = (AWS_KEY_ID = <aws_key_id>
    AWS_SECRET_KEY = <aws_secret_key>)
    ENCRYPTION = ( TYPE = 'AWS_SSE_S3' );

Process images¶

The COMPLETE function processes either a single image or multiple images (for example, extracting differences in entities across various images). The function call specifies the following:

• The multimodal model to be used

• A prompt

• The stage path of the image file(s) via a FILE object

Single-image example¶

The following example uses Anthropic’s Claude Sonnet 3.5 model to summarize a pie chart science-employment-slide.jpeg stored in the @myimages stage.

SELECT AI_COMPLETE('claude-3-5-sonnet',
    'Summarize the insights from this pie chart in 100 words',
    TO_FILE('@myimages', 'science-employment-slide.jpeg'));

Copy

Response:

This pie chart shows the distribution of occupations where mathematics is considered "extremely important" in 2023.
Data scientists dominate with nearly half (48.7%) of all such positions, followed by operations research analysts
at 29.6%. The remaining positions are distributed among statisticians (7.8%), actuaries (7.2%), physicists (5.1%),
mathematicians (0.6%), and other mathematical science occupations (1.1%). This distribution highlights the growing
importance of data science in mathematics-intensive careers, while traditional mathematics roles represent a smaller
share of the workforce.

Multiple-image example¶

Note

Currently, only Anthropic (claude) and Meta (llama) models can reference multiple images in a single prompt. Multiple image support for other models may be available in a future release.

Use the PROMPT helper function to process multiple images in a single COMPLETE call. The following example uses Anthropic’s Claude Sonnet 3.5 model to compare two different ad creatives from the @myimages stage.

SELECT AI_COMPLETE('claude-3-5-sonnet',
    PROMPT('Compare this image {0} to this image {1} and describe the ideal audience for each in two concise bullets no longer than 10 words',
    TO_FILE('@myimages', 'adcreative_1.png'),
    TO_FILE('@myimages', 'adcreative_2.png')
));

Copy

Response:

First image ("Discover a New Energy"):
• Conservative luxury SUV buyers seeking a subtle transition to electrification

Second image ("Electrify Your Drive"):
• Young, tech-savvy urbanites attracted to bold, progressive automotive design

Model limitations¶

All models available to Snowflake Cortex have limitations on the total number of input and output tokens, known as the model’s context window. The context window size is measured in tokens. Inputs exceeding the context window limit result in an error.

For text models, tokens generally represent approximately four characters of text, so the word count corresponding to a limit is less than the token count.

For image models, the token count per image depends on the vision model’s architecture. Tokens within a prompt (for example, “what animal is this?”) also contribute to the model’s context window.

Cost considerations¶

Billing scales with the number of tokens processed. The number of tokens per image depends on the architecture of the vision model.

• Anthropic (claude) models’ formula is roughly: tokens = (Width in pixels x Height in pixels) / 750.

• Mistral (pixtral) models divide each image into batches of 16x16 pixels and converts each batch to a token. The total number of tokens is equivalent to roughly (Width in pixels / 16) * (Height in pixels / 16).

• Meta (llama) models try to tile the image with square tiles. Depending on the image’s aspect ratio and size, the number of tiles can be up to 16, each represented by around 153 tokens.

Choosing a vision model¶

The COMPLETE function supports multiple models of varying capability, latency, and cost. To achieve optimal performance per credit, choose a model that aligns with the content size and task complexity.

The benchmarks are:

• MMMU: Evaluates multimodal models on multidisciplinary tasks that require college-level reasoning.

• Mathvista: Mathematical reasoning benchmark within a visual context.

• ChartQA: Evaluates complex reasoning questions about charts.

• DocVQA and VQv2: Benchmarks for visual question-answering on documents.

Regional availability¶

Support for this feature is available natively to accounts in the following Snowflake regions:

COMPLETE multimodal is available in other regions through Cross-region inference.

Error Conditions¶

Legal¶

The data classification of inputs and outputs are as set forth in the following table.

For additional information, refer to Snowflake AI and ML.