AWS powers the PGA TOUR with generative AI

AWS has been the Official Cloud partner of the PGA TOUR since 2021. The PGA TOUR uses AWS machine learning (ML), storage, compute, analytics, database, and media services to quickly process and distribute video footage from each golf tournament, while giving fans new ways to engage with TOUR content. The PGA TOUR also captures hundreds of hours of raw footage at every tournament, and leverages AWS to rapidly transform content into exciting personalized, digital experiences. And now, with the rise of generative AI, we are delighted to announce our latest service for the TOUR.

Powered by AWS, generative AI and large language models (LLMs) enabled the development of a prototype virtual assistant for the PGA TOUR. This assistant, built using AWS services like Amazon Bedrock, Amazon Athena, and Amazon Kendra, allows fans to ask questions and retrieve information about PGA TOUR events, players, statistics, and more in a conversational format. The solution combines a text-to-SQL approach for structured data and a retrieval-augmented generation (RAG) approach for unstructured text data.

“Working with Amazon Bedrock and Anthropic’s CLAUDE 2.0 model, we build a multi-modal chatbot prototype that accesses 30 years of media guides, PGA TOUR competition data between 2005 and 2023, and our entire video archive. This is just the very beginning of where we are headed.”

- Scott Gutterman, SVP Digital Operations of the PGA TOUR

For structured data, the team processed and labeled the tabular competition data, adding auxiliary columns to help the LLM understand golf-specific concepts. They used Amazon Athena to run SQL queries generated from the user's natural language input. For unstructured text data like media guides, they chunked the documents and indexed them using Amazon Kendra for efficient retrieval. The relevant information from both data sources was then provided to the LLM to generate the final response.

The team evaluated various LLMs from Amazon Bedrock, ultimately selecting Anthropic's Claude v2 for its performance. They spent considerable effort on prompt engineering, developing fallback prompts and error correction mechanisms. The resulting prototype virtual assistant can field a wide range of questions, blending structured and unstructured data to provide informative and engaging responses. The PGA TOUR sees this as the beginning of their generative AI journey, enabling new fan experiences and internal efficiencies.

You can read the full blog of the PGAT’s generative AI virtual assistant here??

https://aws.amazon.com/blogs/machine-learning/the-journey-of-pga-tours-generative-ai-virtual-assistant-from-concept-to-development-to-prototype/