Snowflake Cortex AI Functions in SQL: AI_COMPLETE, AI_FILTER, AI_AGG and More

Last updated: October 2026

Introduction

Snowflake announced Cortex AISQL in June 2025 as a way to use language models inside ordinary SQL: filtering rows by meaning, classifying text and images, and summarising whole groups of records. The functions have since become generally available, the documentation now calls them Cortex AI Functions, and several of the models used in early examples, including the screenshots in the first version of this article, have been retired. This version covers the functions as they stand in October 2026, with examples you can run today.

From preview to general availability

The core functions moved out of preview in three steps:

The family also includes AI_EXTRACT, AI_SENTIMENT, AI_TRANSLATE, AI_SUMMARIZE (still in preview), AI_PARSE_DOCUMENT, AI_REDACT and AI_MULTI_EMBED, plus helper functions such as TO_FILE, PROMPT and AI_COUNT_TOKENS. Our article on translation, summarisation, sentiment and classification covers the text-analysis functions in more detail.

AI_COMPLETE and choosing a model

AI_COMPLETE sends a prompt to a model you name and returns its response. It accepts positional or named arguments: the model and prompt are required, and model parameters, a response format and a show_details flag are optional. In the model parameters, temperature and top_p range from 0 to 1 and default to 0, and max_tokens defaults to 4096. For a single image or file, the form is AI_COMPLETE(model, predicate, file [, model_parameters]), with three required arguments.

SELECT AI_COMPLETE(
    model => 'llama3.3-70b',
    prompt => 'Summarise the main complaint in this review in one sentence: ' || review_text,
    model_parameters => {'temperature': 0, 'max_tokens': 120}
) AS complaint
FROM product_reviews
LIMIT 20;

-- Structured output: the response follows a JSON schema
SELECT AI_COMPLETE(
    model => 'llama3.3-70b',
    prompt => 'Extract the product and the complaint from this review: ' || review_text,
    response_format => {
        'type': 'json',
        'schema': {
            'type': 'object',
            'properties': {
                'product':   {'type': 'string'},
                'complaint': {'type': 'string'}
            },
            'required': ['product', 'complaint']
        }
    }
) AS extracted
FROM product_reviews
LIMIT 20;

Model choice is where older examples break. Snowflake retires models on a schedule, and a query that names a retired model fails. Since the first version of this article, llama2-70b-chat was deprecated in the 2025_05 behavior change bundle, Claude 3.5 Sonnet was deprecated on 31 March 2026, and mistral-7b and llama4-maverick became legacy models on 12 August 2026, which means accounts that had not used them before that date cannot call them. The models listed for AI_COMPLETE now include current Claude, Llama, Mistral and DeepSeek models, and availability depends on your region. We recommend keeping the model name in one configuration table or variable, so that moving to a newer model is a one-line change.

AI_CLASSIFY

AI_CLASSIFY needs at least two distinct categories, and Snowflake notes that more than about twenty can reduce accuracy. An optional configuration object accepts a task description of up to 50 words, 'output_mode': 'multi' for multi-label results, and few-shot examples. The result is an object with a labels array.

SELECT ticket_id,
       AI_CLASSIFY(
           ticket_body,
           ['technical issue', 'billing', 'account access', 'general inquiry'],
           {'task_description': 'Classify the main reason the customer contacted support'}
       ):labels[0]::STRING AS ticket_type
FROM support_tickets;

AI_FILTER

AI_FILTER returns TRUE or FALSE for a condition written in natural language, which makes it usable directly in a WHERE clause or a join condition. It has three forms: a single input string, a predicate plus a file, and a PROMPT() template that combines several columns.

-- 1. A single input string
SELECT AI_FILTER('Is Canada in North America?');

-- 2. A predicate and a file: find cat pictures on a stage
WITH pictures AS (
    SELECT TO_FILE(file_url) AS img FROM DIRECTORY(@file_stage)
)
SELECT FL_GET_RELATIVE_PATH(img) AS file_path
FROM pictures
WHERE AI_FILTER('Is this a picture of a cat?', img);

-- 3. A PROMPT() template over columns
SELECT review_id, review_text
FROM product_reviews
WHERE AI_FILTER(PROMPT('This review mentions a delivery problem: {0}', review_text));

AI_AGG and AI_SUMMARIZE_AGG

The aggregate functions work across many rows at once. AI_AGG takes a text expression and an instruction describing the aggregation, and AI_SUMMARIZE_AGG summarises a group without an instruction. Unlike AI_COMPLETE and AI_SUMMARIZE, both handle groups of text larger than the model's context window, which makes them suited to questions such as what customers say about each product.

SELECT restaurant_id,
       AI_AGG(review, 'Summarize the restaurant reviews for potential consumers') AS summary
FROM reviews
GROUP BY 1;

SELECT product_id, AI_SUMMARIZE_AGG(review_text) AS summary
FROM product_reviews
GROUP BY product_id;

When calling AI_AGG with named arguments, note that the second argument is called instruction.

AI_SIMILARITY

AI_SIMILARITY returns a score from -1 to 1 for two texts or two images; it does not compare a text with an image. An optional configuration object can name the embedding model to use. It is useful for matching near-duplicate descriptions or finding records that say the same thing in different words.

SELECT AI_SIMILARITY(
    'Stainless steel chef knife, 8 inch',
    '20 cm chef''s knife in stainless steel'
) AS similarity;

Keeping costs under control

AI functions are billed by the tokens they process, and a function applied to every row of a large table processes a great many tokens. Three habits keep costs predictable. Develop against a sample with LIMIT, and estimate the input size with AI_COUNT_TOKENS before a full run. Materialise results into a table instead of calling the function again in every dashboard query. And note that resource monitors cover warehouses only; spending on AI services needs a Snowflake budget to be tracked and capped.

Key takeaways

• Snowflake's AI functions are generally available, and the documentation now calls them Cortex AI Functions.
• AI_COMPLETE needs a model that is still available in your region; retired models such as llama2-70b-chat and Claude 3.5 Sonnet fail.
• AI_CLASSIFY needs at least two categories and returns a labels array; AI_FILTER fits naturally in WHERE clauses.
• AI_AGG and AI_SUMMARIZE_AGG handle groups of text larger than a model's context window.
• Develop on samples, materialise results, and track AI spending with budgets.

Frequently Asked Questions