How does Cortex AI SQL integrate LLMs with the database?

It integrates LLMs by providing new primitive operators within the SQL query language, allowing composition with existing SQL syntax like CTEs, group by, and joins. This also leverages the SQL query planner for performance optimizations.

Can Cortex AI SQL handle unstructured data and text analysis?

Yes, Cortex AI SQL offers managed functions that require only natural language instructions for tasks like sentiment analysis, summarization, and filtering text or even images. It can also enable aggregation using AI with group by clauses.

How does Snowflake optimize AI SQL query performance?

Snowflake optimizes AI SQL through techniques like minimizing LLM calls by reordering filters, using cascade strategies with proxy models for dynamic thresholding, and query rewrites that treat joins as classification tasks for efficiency.

What is Snowflake Intelligence and how does it relate to AI SQL?

Snowflake Intelligence is a chat-like interface that allows users to talk to their data. Cortex AI SQL serves as a foundational building block that powers Snowflake Intelligence, enabling it to understand and query user data.

Can Snowflake Intelligence be white-labeled for other companies?

Yes, Snowflake Intelligence can be white-labeled with custom logos, fonts, and color schemes. It's intended for business users as an application and also offers a REST API for integration into third-party front ends.

How accurate is the natural language to SQL translation in Snowflake's products?

While specific numbers for Cortex AI SQL weren't provided, the presentation mentioned that for the related Cortex Analyst product, text-to-SQL accuracy can be high. It was noted that Snowflake's internal evaluations on text-to-SQL have outperformed available models for evaluation.

Key Moments

AI Dev 25 x NYC | Benjamin Han: Snowflake: A SQL Engine to Talk to Your Data

DeepLearning.AI

Education4 min read23 min video

Dec 3, 2025|427 views|7

Save to Pod

Key Moments

TL;DR

Snowflake's Cortex AI SQL integrates LLMs into databases, enabling natural language queries and advanced data analysis.

Key Insights

Cortex AI SQL transforms traditional SQL into an AI query language by embedding LLM capabilities directly into database operations.

It offers managed functions and low-level primitives, allowing users to process text, perform sentiment analysis, summarization, and multimodal analysis using natural language instructions.

The platform enhances efficiency and ease of use by allowing composition of AI functions with standard SQL syntax like joins, WHERE clauses, and GROUP BY.

Snowflake leverages its SQL query planner for significant performance optimizations, reducing LLM call volumes and execution times.

AI SQL provides inherent governance benefits, including role-based access control, budget constraints, and data encryption, inherited from the Snowflake platform.

The technology powers higher-level agentic features like Snowflake Intelligence, a chat-like interface for data interaction, which can be white-labeled and customized.

Performance optimizations include dynamic thresholding with proxy models and query rewrites to handle large-scale AI-driven data analysis tasks efficiently.

INTRODUCTION TO SNOWFLAKE'S AI OFFERINGS

Benjamin Han introduces Snowflake's Cortex AI SQL, a project launched about a year prior and expanded from five to sixty engineers. The core goal of Cortex AI SQL is to infuse generative AI features with the efficiency, ease of use, and trust benefits inherent in the full Snowflake SQL platform. This suite includes managed models for text processing, sentiment analysis, and summarization, supporting various input methods like SQL, Python, REST, and a no-code interface, Snowflake Studio.

THE AISQL ARCHITECTURE AND PRIMITIVES

At the foundation of Snowflake's AI strategy lie scalable primitives, including document processing for extraction, parsing, and embedding, culminating in Cortex AI SQL. This layer integrates Large Language Models (LLMs) directly into the database by introducing new primitive operators into the query language. This integration allows for seamless composition with existing SQL syntax, such as Common Table Expressions (CTEs), GROUP BY, joins, and WHERE clauses, while also benefiting from performance optimizations.

FUNCTIONALITY AND USE CASES OF AISQL

Cortex AI SQL transforms SQL into an AI query language. It offers low-level LLM calls via `AI_COMPLETE` for direct integration, enabling users to select models, provide prompts, and point to data. Beyond primitives, managed functions simplify tasks through natural language instructions. Examples include `AI_FILTER` for sentiment analysis on customer transcripts, multimodal analysis of images for product identification, and `GROUP BY AI` for summarizing large text sets or identifying top complaints, all composable within standard SQL.

DEMONSTRATION: TOPIC DISCOVERY IN CUSTOMER REVIEWS

A demonstration showcased a workflow for topic discovery on customer reviews. Initially, an AI function identifies broad categories. Subsequently, this output is refined by classifying reviews into these discovered topics, revealing themes like 'value for money' and 'material quality.' When outliers emerge tagged as 'other,' further AI analysis is performed to identify underlying reasons, such as 'lack of specificity' in reviews. This iterative process, involving a 'human in the loop,' allows for refining categories and filtering out uninformative feedback for better analysis.

PERFORMANCE OPTIMIZATIONS FOR AI QUERIES

Significant performance optimizations are crucial for making AI SQL practical. Traditional SQL query plans often minimize join costs, which can lead to numerous LLM calls. Cortex AI SQL optimizes by strategically minimizing the number of LLM calls. This includes executing CPU-bound filters first to reduce the dataset before applying expensive LLM filters. Techniques like dynamic thresholding with proxy models and query rewrites, transforming joins into classification tasks for smaller candidate sets, drastically reduce execution time and improve accuracy.

GOVERNANCE AND HIGHER-LEVEL AGENTS

Leveraging the Snowflake platform, Cortex AI SQL provides built-in governance features such as role-based access control, budget constraints, and data encryption, ensuring secure and controlled AI-driven data operations. These low-level primitives serve as building blocks for higher-level agentic systems. Snowflake Intelligence, an example of such a system, is a chat-like interface allowing users to interact with their data using natural language, with AI SQL powering its ability to understand queries and retrieve relevant information.

FUTURE DIRECTIONS AND WHITE-LABELING

The architecture supports future agentic features and integrations. Notably, Snowflake Intelligence is designed to be adaptable, offering white-labeling capabilities with customizable branding. The underlying REST API allows integration into front-end applications of choice, extending its utility beyond internal Snowflake users. Efficacy testing on natural language to SQL translation shows high accuracy, with Snowflake's offerings outperforming existing models in evaluations for specific products like Cortex Analyst.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●Studies Cited

●People Referenced

Common Questions

Cortex AI SQL is a Snowflake product that transforms SQL into an AI query language. Its main benefits include efficiency, ease of use, and trust by integrating generative AI features directly into the Snowflake SQL platform.