From SQL to Gen-AI: A First Look at Snowflake Cortex AISQL

Snowflake's new Cortex AISQL functions let you run large language model tasks such as classification, extraction, translation, and even image Q&A directly in SQL. Here is what it means for data teams, how it works in practice, and where we at Viewnear see the biggest opportunities.

Eduardo Javier Ramos

CEO

June 21, 2025 · 4 min read

Connect on LinkedIn →

From SQL to Gen-AI: A First Look at Snowflake Cortex AISQL

Key takeaways

Cortex AISQL embeds state-of-the-art LLMs directly inside the Snowflake engine, so there is no extra AI infrastructure to stand up and your data never leaves the platform.
Analysts can prototype LLM workflows with nothing more than a SELECT statement, classifying sentiment, extracting fields, and answering questions about images in a single query.
Pairing PARSE_DOCUMENT with AI_COMPLETE handles PDFs and mobile photos in one pipeline, which is ideal for mortgage and insurance use cases.
Existing roles, masking policies, and row-level security still apply, and credits scale with input tokens and the chosen model, so prototype small and track usage.
The hosting question is largely solved; the real decision is now which business problem to tackle first.

Why This Matters

At Snowflake Summit 2025 the company unveiled Cortex AISQL, a family of built-in functions that embed state-of-the-art LLMs directly inside the Snowflake engine. For teams that already store governed data in Snowflake, this brings three immediate benefits:

No extra infrastructure: Everything executes inside the platform, so you avoid spinning up a separate AI service.
Unified security: Data never leaves Snowflake, and existing roles, masking policies, and row access controls still apply.
Rapid iteration: Analysts can prototype LLM workflows using nothing more than a SELECT statement.

What's In The Toolbox?

Core Text And Document Functions

AI_COMPLETE: free-form completions, summarization, Q&A
AI_CLASSIFY: zero-shot and few-shot text classification
AI_FILTER: policy or sentiment filtering
AI_SIMILARITY: vector similarity search over text
PARSE_DOCUMENT: OCR and structural parsing for PDFs, images, and Word docs

Image-Aware Features

The same AI_COMPLETE interface can caption images, extract entities, and answer questions about visuals such as charts or diagrams.

A Quick, Practical Demo

Imagine you store product-review screenshots in a stage named @reviews_stage. You want to classify each review as positive, neutral, or negative and pull out the star rating shown in the image.

-- Extract review text, classify sentiment, and pull the star rating
WITH IMGS AS (
  SELECT
    RELATIVE_PATH                              AS FILE_NAME,
    TO_FILE('@REVIEWS_STAGE', RELATIVE_PATH)   AS IMG_FILE
  FROM DIRECTORY(@REVIEWS_STAGE)
),
TEXT AS (
  SELECT
    FILE_NAME,
    -- 1) Extract the visible text from each screenshot
    AI_COMPLETE(
      'claude-3-5-sonnet',
      'Extract the visible product-review text. Respond with raw text only.',
      IMG_FILE
    )::STRING AS REVIEW_TEXT
  FROM IMGS
)
SELECT
  FILE_NAME,
  REVIEW_TEXT,
  -- 2) Classify overall sentiment
  AI_CLASSIFY(
    REVIEW_TEXT,
    ARRAY_CONSTRUCT('positive','neutral','negative')
  )['LABEL']::STRING AS SENTIMENT,
  -- 3) Parse the star rating (1-5) mentioned in the text
  AI_COMPLETE(
    'claude-3-5-sonnet',
    'Return only the star rating (1-5) mentioned in this review.',
    REVIEW_TEXT
  )::STRING AS RATING
FROM TEXT;

All three steps run inside Snowflake with no batch exports or external endpoints.

Governance And Cost Notes

Privileges: Grant the CORTEX_USER role plus warehouse usage. Fine-grained masking and row-level security continue to apply.
Credits: Each function consumes credits based on input tokens and the chosen model. Prototype with limited data first, then track usage through credit-consumption views to avoid surprises.

Early Lessons From The Viewnear Team

Barrier to entry is low. Analysts with some coding experience built sentiment dashboards in a morning.
Prompt clarity matters. A clear system prompt plus a few shot examples cut wrong answers by about forty percent.
Hybrid workflows shine. Pairing PARSE_DOCUMENT with AI_COMPLETE lets us handle PDF contracts and mobile photos in the same pipeline, ideal for mortgage and insurance clients.
Warehouse sizing affects latency. LLM calls run in a single thread, so moving from a small to a medium warehouse often halves response time.

Where It's Heading Next

Realtime agents: Combine AISQL with Snowpark Container Services to create always-on AI copilots that can invoke external APIs when needed.
Governed vector search: The Native App Framework now supports restricted-caller rights, making it safer to package and share custom AI models.
Industry blueprints: Expect pre-built templates such as invoice extraction, support-ticket triage, and clinical note summarization to arrive via partners, including Viewnear.

Closing Thoughts

Cortex AISQL blurs the line between data warehousing and generative AI. The question is no longer "Where do we host the model?" but "Which business problem do we tackle first?" If you are ready to experiment, the barrier is now a single SQL query.

The Viewnear team is already weaving AISQL into upcoming client sprints. If you would like to see a live demo tailored to your data, get in touch. Happy querying!

Written by

Eduardo Javier Ramos

CEO

Connect on LinkedIn →