The Fifth Elephant 2023 Winter

In this talk, we will outline the journey of building policyGPT, a system designed to automate responses to health insurance support queries. The system can guide users with educational context such as how-to guides and provide instant, precise answers to a wide range of coverage questions interpreting policy terms and conditions. Given this context’s crucial high accuracy requirements, our main focus was ensuring that the model provides the user with the most reliable and useful information.

The talk will explore the various steps we took to ensure high accuracy and a great user experience, from preparing high-quality input data to indexing strategy, temperature control, memory management and prompt engineering to enhance discreet reasoning.

We will use many examples from policyGPT’s early prototypes to production-ready and handy tips on evaluation framework.

Speaker

Karthik B.
Engineering, Plum
linkedin profile