Snowflake’s Cortex Analyst helps users perform complex analyses or dashboard customizations using natural language.

Snowflake announced the public preview of Cortex Analyst, a tool that allows users to analyze datasets in its cloud data platform using natural language prompts. The company published a detailed blog post about Cortex Analyst, which is based on Meta’s Llama and Mistral models. This tool converts voice prompts into SQL code and automatically checks for errors.

User-friendliness

With Cortex Analyst, Snowflake aims to solve the problem of always having to involve data scientists for complex analysis or dashboard customizations. Instead of waiting for new dashboards or customizations, business users can now send natural language instructions to Cortex Analyst, which automatically converts them into SQL queries that can be executed in Snowflake.

Cortex Analyst is accessible through Streamlit, a no-code interface in Snowflake, and through an API that enables integration with other software. The tool uses large language models (LLMs) from Mistral AI and Meta Platforms’ Llama family, with optional support for OpenAI models on Microsoft Azure.

Bug fixes

Cortex Analyst not only converts the statements into SQL code, but also automatically checks them for errors using the Error Correction Agent. Among other things, this module prevents LLMs from generating incorrect SQL commands. The Synthesizer Agent then summarizes the corrected code into a single executable snippet.

The reliability of LLMs in SQL code generation is a known challenge. According to data from Forrester Research, which Snowflake cites, the accuracy of AI-generated SQL queries depends on the complexity of the task and ranges from 20 to 70 percent. To minimize risk, Cortex Analyst uses relatively simple data schemas and suggests alternative queries when an original task is too complex.