Running inference on unstructured data like images or audio has always meant leaving your data warehouse. You build a separate pipeline, manage a Python service, and orchestrate data movement between systems. That pattern is starting to break. With Snowflake's release of multimodal functions in Cortex AI, you can now analyze image and video content with familiar SQL commands, directly where your data lives.
This isn't just a cosmetic change. It represents a fundamental shift in how we should think about unstructured data pipelines. The takeaway is that the complex, multi-service architectures we built for tasks like image classification or video analysis can now be radically simplified into a single SQL query. This collapses the stack, tightens security, and puts powerful AI capabilities into the hands of any analyst who knows SQL.
what just changed
Snowflake has integrated multimodal AI capabilities directly into its SQL engine through a set of features called Cortex AI Functions. Now in public preview, these functions allow you to run inference on unstructured data like images, audio, and documents that live in Snowflake stages or even external object storage like S3. This is handled without moving the data out of Snowflake's security perimeter.
The system introduces a native FILE data type, which allows SQL to directly reference and operate on this kind of data. Instead of treating an image as an inert blob, the query engine can now pass it to a model for analysis as part of a standard query. This eliminates the need for external tools and specialized skills to bridge the gap between structured and unstructured data analysis.
why it matters: collapsing the ai/data stack
The main consequence for builders is the simplification of the data stack. The typical workflow for analyzing, say, user-uploaded images in a product catalog involves several steps: an ETL process to extract the data, a separate service to call a vision API, and another process to load the resulting metadata back into the warehouse. This creates complexity, latency, and multiple points of failure.
By embedding the model call directly in SQL, you remove the need for that entire external pipeline. This offers a few concrete advantages. First, performance and cost. Snowflake claims this native integration can be over 30% faster and reduce costs by up to 60% compared to traditional, manually orchestrated AI implementations. Second, it democratizes access. Any data analyst can now perform tasks that previously required a machine learning engineer, like filtering products based on visual characteristics or aggregating customer sentiment from call recordings. Finally, it unifies governance. All data, structured and unstructured, is processed within a single, secure platform.
a practical sql example
Let's say you have a table of product listings, with structured data like product_id and price, alongside a column containing a URL to a product image stored in a Snowflake stage. You need to identify products that are missing a required safety warning label in their image.
Instead of building an external image processing service, you can now write a SQL query to do the check directly.
-- Find products missing a visible safety label from their primary image
SELECT
p.product_id,
p.product_name,
p.image_url
FROM
products p
WHERE
SNOWFLAKE.CORTEX.COMPLETE(
'pixtral-large', -- or another supported vision model
'Does this image contain a standard electrical safety warning label? Answer YES or NO.',
p.image_url
) = 'NO';
This query joins your structured product data with AI-driven analysis of your unstructured image data in a single statement. It's a simple but powerful pattern that can be extended to classification, object detection, sentiment analysis from audio, and more, all within the environment your data team already uses.
the takeaway for builders
The line between the data platform and the AI platform is dissolving. Bringing multimodal inference directly into the data warehouse is a significant step toward simplifying the productionalization of AI. For engineers and data teams, this means less time spent on plumbing and orchestration and more time spent on extracting value from data.
The most valuable data in many organizations—call recordings, support documents, product images, promotional videos—is often the hardest to analyze at scale. Tools that lower the barrier to processing this data natively are a major unlock. It's time to re-evaluate those complex Python-based data processing jobs and see if they can be replaced with a few lines of SQL.
Top comments (0)