Snowflake is expanding its Cortex AI platform with Cortex Search and Cortex Analyst. These LLM-based chatbots can reliably answer questions about unstructured and structured data, respectively.

Snowflake is showcasing two major enhancements to its Cortex AI platform at Snowflake Summit: Cortex Search and Cortex Analyst. Search is a module that enables companies to quickly build a chatbot that understands unstructured data, while Analyst provides answers to questions in structured data.

“Cortex Analyst translates your questions into SQL that you can also validate,” explains Christian Kleinerman, EVP Product for Snowflake. “And with Cortex Search, we want to make it easy for users to build a chatbot for text data.”

State of the art

“Both solutions are state-of-the-art,” says Baris Gultekin, Head of AI at Snowflake, during a demo of the products for ITdaily. “The chatbots provide correct answers and point out when something is missing from the data, so they are not hallucinating. If a question is not clear enough, it will not be answered until you provide additional information.”

You can configure both Cortex Search and Cortex Analyst through the new AL & ML Studio. First, we see Gultekin creating a new bot using Cortex Search. This requires just a few clicks of the mouse, in which he specifies which data will serve as the source. In this case, it is a folder with several hundred PDFs.

Answers, but not always

Cortex Search analyzes the unstructured data and extracts the data using OCR, which is then structured and indexed. The process is completed in a few minutes, and Gultekin can ask questions. “How do I install Slack on my phone?” Search provides a clear, readable answer to this question and points to the source document that contains the information. However, when Cortex Search asks to install Instagram, it says that this information is not available.

Gultekin has a preconfigured demo for Cortex Analyst ready. The chatbot has access to structured data. Gultekin asks about the Top products If it is a fictitious company, the analyst notes that this question cannot be answered without further information Great is an ambiguous concept.

SQL Specialist

A more specific question would work. The analyst translates that question into a relevant SQL query and retrieves the relevant data, all in the background. The chatbot then uses that data to formulate an answer. “GPT-4 already does SQL very well,” Gultekin said, “but thanks to our optimization, Cortex Analyst is better.” In the background, this model runs on Mistral and Llama 3.

“We made it so easy to build a chatbot that anyone can do it,” Kleinerman concludes. Snowflake illustrates this by getting an audience on the big keynote stage to work on the demo. In true demo tradition, of course, this goes wrong. We conclude that anyone can build a chatbot, but it still helps if the head of product gives some tips during the process.

Cortex Search and Cortex Analyst will be available in public preview soon.