PlanRAG: A Plan-then-Retrieval Augmented Generation for Generative Large Language Models as Decision Makers

In general, the decision-making task requires performing the following three steps: (1) making a plan for which kind of analysis is needed for decision; (2) retrieving necessary data using queries; (3) making a decision (i.e., answering) based on the data (Troisi et al., 2020; Sala et al., 2022). To make the Steps (2) and (3) easier, a lot of decision support systems have been developed and utilized during the past few decades (Gupta et al., 2002; Eom and Kim, 2006; Power, 2007; Hedgebeth, 2007; Power, 2008; Kasie et al., 2017). However, humans still have been in charge of the most hard part, Step (1). The goal of this study is to investigate the possibility of replacing the human role with a Large Language Model(LLM) such that it performs not only Steps (2) and (3) but also Step (1), that is, all the Steps end-to-end.