An Outline for an Interrogative/Prompt Library to help improve output quality from Generative-AI Datasets

Abstract

This White Paper provides insight into an outline for what is termed a proposed ‘Interrogative/Prompt Library’ (IPL) designed to help optimise the quality of output when engaging with Generative-AI (Gen-AI) datasets, such as, most notably, the recently rapidly developing ChatGPT. The White Paper begins with a recap of the authors’ findings from their previous work investigating ChatGPT, as published in full during March 2023 in an earlier ARC White Paper (as referenced). Next, the importance of questioning Large Language Model (LLM) datasets so that they can be better understood is covered, before investigating 'interrogatives' (involving who, why, what, when, where, how, etc. questions) and scoping their role in analytical search processes, including fundamentals relating to how interrogative questions are structured. Following on from the above work are 'Phase 1' suggestions towards building an ‘Interrogative Library Typology', before delving more specifically into the area and activities of 'Prompt Engineering'. The White Paper then examines the development and maintenance of an Interrogative/Prompt Library, in the form of presenting a second phase. That work includes insight into the 'Interrogative Prompt Library Engine' that underpins the above work. A number of overall Conclusions and Key Takeaways are then tabled, noting especially the guidance value acquired from engaging with the activities discussed throughout this White Paper. Thereby, end-users are increasingly better armed for engaging with Gen-AI datasets helping ensure that they best reduce the risks of, amongst others, falling into 'Garbage-In, Garbage- Out' (GIGA) traps. Finally, this White Paper ends with a ‘call for action!’ for further research and development relating to what is tabled in this White Paper, paving the way for further collaboration. Appendices are also included to provide further reference detail. --- ARC White Paper (London: Analytic Research Consortium - ARC, May 2023)