Lowe’s fine-tunes OpenAI’s models to improve ecommerce data quality

Lowe’s is a Fortune 50 home improvement company that processes approximately 16 million customer transactions a week in the United States. Lowe’s has become a leading omnichannel retailer, using best-in-class technology to elevate customer and associate experiences. The data and AI team at Lowe’s saw an opportunity to further improve the data quality in its website search by fine-tuning an OpenAI model.

Accurately presenting product information to customers throughout their shopping journey is a challenge faced by large retailers, aiming to ensure that what shoppers order will be exactly represented upon delivery.

To elevate the shopping experience for Lowe’s customers on Lowes.com, Lowe’s Data and AI team wanted to improve the accuracy of product descriptions. Inaccurate descriptions make it difficult for Lowe’s to plan and organize inventory, while creating friction in the online shopping experience. For instance, if a customer searches for a ‘copper bathroom faucet,’ they might see numerous chrome bathroom faucets instead. Similarly, a search for an ‘18-inch dishwasher’ could yield results for 24-inch dishwashers due to inaccurate product catalog titles or descriptions. As a result, fewer customers may complete purchases on Lowes.com.

The challenge of accurate product tagging is familiar to online retailers and even AI teams. AI engineers have been using techniques ranging from basic token matching with rule-based regular expressions to advanced natural language processing (NLP) methods to address these challenges. Humans then verify the flagged errors, but human error rates can be high. Human intervention also burdens associates with additional workload that doesn’t solve the problem of incorrectly flagged errors.

In January 2023, engineers and AI specialists at Lowe’s decided to use OpenAI’s GPT-3.5 model to help address data quality discrepancies. Initial tests showed a positive impact on result accuracy, and it looked like this was just the solution they needed.

“Excitement in the team was palpable when we saw results from fine-tuning GPT 3.5 on our product data, and we knew we had a winner on our hands!” said Nishant Gupta, senior director, data, analytics and computational intelligence (DACI).

In a newly defined field of prompt engineering, Lowe’s tried different combinations of prompts to see if accuracy of this process could be improved. Within a few months, they saw remarkable improvements in specific product categories and scaled the process further. When the team put the AI-based product classification into production on Lowes.com, they saw an immediate reduction in the workload of associates responsible for vetting these errors.

As great as this solution was, accuracy still needed to be improved, and the size of prompts was increasing at an unprecedented rate. This is when the Lowe’s team started using OpenAI’s fine-tuning API to fine-tune GPT 3.5 to better understand the Lowe’s dataset, which led to a 20% increase in accuracy. Now, with the product tagging complete, Lowe’s plans to fine-tune newer versions of the model based on errors flagged by team members.

Lowe’s continues to improve its accuracy and minimize errors, increasing productivity in the error vetting process. In some categories, Lowe’s rate of detecting errors has increased approximately 60%—leading to far more accurate product descriptions in search results.

With OpenAI’s help, Lowe’s has made its product information more accurate and consistent, reducing friction across the customer experience and driving ecommerce growth.