Reinforcement Learning based automated curation of Prompt Stores
Prompts today are the primary mode of interaction with large language models (LLMs). Prompts need to be tuned according to the user need, providing the right context and guidance to the LLM — to maximize the chances of getting the ‘right’ response.
It has led to the rise of prompt engineering [1] as a professional discipline, where prompt engineers systematically perform trials, recording their findings, to arrive at the ‘right’ prompt to elicit the ‘best’ response. The list of such successful prompts are then organized in the form of a library such that they can be efficiently reused — referred to as a prompt store.
Unfortunately, curating and maintaining a high quality prompt store remains challenging. The overarching goal of a prompt store is to be able to retrieve the optimal prompt for a given task, without having to repeat the whole experimentation process. However, this retrieval is easier said than done primarily due to the overlapping nature of prompts.
Problem Statement
Let us try and understand the issue of overlapping prompts with the help of a couple of prompts from the field of content writing (one of the areas with highest Gen AI adoption today):