Artificial intelligence (AI) experts from Amazon and the PGA Tour have developed a prototype virtual assistant that uses large language models (LLMs) and the Retrieval-Augmented-Generation (RAG) framework to provide real-time, interactive data and information to fans. This post summarises a project that involved the AWS Generative AI Innovation Center, AWS Professional Services and the PGA Tour, using Amazon Bedrock, a fully managed service offering foundation models for AI application development.
The objective was to create a Q&A chatbot that could interpret and respond to natural language queries about the PGA Tour’s vast data history, for improved fan engagement. The AWS team developed a text-to-SQL converter that used Amazon Athena to extract relevant data as SQL code, which was then interpreted by an LLM that could generate the necessary response.
Amazon Kendra provided an efficient and convenient means to index text data for LLM use; this information was then used for context and reinforcement during the generation of answers. By following this approach, AWS and the PGA Tour were able to create a more robust and versatile assistant that was able to answer questions and provide data on any event, player, or detail in PGA Tour history. The application was developed to be scalable and suitable for real-time use, capable of generating content and data appropriate for the context of live events on the tour.
The cloud-based virtual assistant prototypes were carefully tested, with the Anthropic’s Claude v2 and Claude Instant models in Amazon Bedrock proving the most effective for the pipeline and use case. The key to effective use of LLMs was effective prompting: various prompts were developed and tested for different tasks, with multiple fallback options and potential error correction inputs prepared for various issues. The virtual assistant was developed to use both structured and unstructured data pipelines, with the Anthropic Claude 2 model providing the best overall results.
Building the prototype involved a Lambda function that would guide prompt selection and communication between Amazon Bedrock, Kendra and Athena. Scott Gutterman, SVP of Broadcast and Digital Properties at the PGA Tour, stated that the PGA Tour aims to use this advanced AI technology to enhance the fan experience and explore new ways to engage with the wealth of data available from the event’s rich history.
This project demonstrated that AI technology can be used to create next-generation virtual assistants capable of advanced fan engagement and unlocking efficiencies. By making structured databases and content more accessible, this technology has potential applications across other industries. For companies considering similar applications, AWS offers the Agents for Amazon Bedrock and Knowledge Bases for Amazon Bedrock as possibilities, two fully managed solutions for intelligent automation and data search abilities.