ThoughtSpot Inc. is climbing aboard the large language model bandwagon with today’s announcement that it is integrating the popular GPT-3 LLM into its business intelligence platform.

GPT-3 is the model that underlies the enormously popular ChatGPT chatbot from OpenAI LLC. ThoughtSpot’s platform allows users to query business data from multiple sources using search engine-like terms and to visualize results as charts, graphs and maps. The company, which has raised more than $660 million, claims to have four of the five largest U.S. companies as customers and more than one-third of the Fortune 100.

The new offering will be packaged as ThoughtSpot Sage, combining GPT-3 with the company’s patented search technology. Support for additional LLMs is planned in the future.

The major benefits of GPT-3 integration are to query data and get results using natural language, said Chief Development Officer Sumeet Arora. ThoughtSpot Sage can also be used to assist in data modeling and the company’s support team has been equipped with LLM-based assistance.

The addition of LLM support is also intended to ease the burden on data analysts who tend to “get caught up building dashboards and taking small requests,” Arora said. “That is the drudgery we are aiming to remove. We allow analysts to create the guard rails for how search analytics work.”

The end of clicks?

Arora said natural language processing will make drag-and-drop query construction obsolete. “The mouse-and-click interface is dead,” he said. “We have harmonized human behavior with a human experience in the search bar.”

What natural language lacks, however, is precision. Asking a computer what is the company’s most successful product, for example, could invite an answer according to sales, customer reviews, profitability or other criteria. ThoughtSpot gets around this by using search tokens, which classify queries by such criteria as column, operator, value and keyword.

“We started by creating a relational interface with tokens that dramatically reduces the effort to get answers, but you have to be analytics-fluent,” Arora said. “We’ve now added a layer with search tokens that guarantees accuracy but is natural language-like.”

Key to that is that queries are translated into ThoughtSpot’s native Sage Grammar query language and presented to the user for approval. “We will tell you how we translated your query and allow you to edit it,” he said.

Noting that GPT-trained models have sometimes been called confident liars, Arora said ThoughtSpot plans to add color-coded confidence ratings to its answers, a feature ChatGPT lacks. The LLM model for each customer is automatically trained over time in the terms and queries that are specific to that organization. Customers can specify which examples to use to train the system and training data is never shared across customers, Arora said.

GPT-3 will be incorporated into the company’s cloud service over the next few weeks at no extra charge.

Image: geralt/Pixabay