Language models and library material search system - case study on NDL search -

Abstract

Log analysis of NDL search showed that there are two types of users: those who search by simple keywords and those who search by a combination of search conditions, and the ratio of these two types of users is about 50-50. For the former type of users, ambiguous search using a large-scale language model (LLM) might be effective. RAG (Retrieval-Augmented Generation) using LLM directly is costly and has some problems, but it was confirmed that vector search using a part of the components of RAG has practical potential. For the latter type of user, a function that guides the user in how to search is considered more effective than ambiguous search. We will introduce article recommendation functionalities of NDL search.