I. Introduction
Turning to search engines to address daily information needs has become a common behavior. In response to a given query, search engines typically employ the established information retrieval (IR) pipeline, specifically the index-retrieve-rank strategies [1], [2], to generate a ranked list of documents. Over the past several decades, the inverted index [1] has been foundational to term-based or sparse retrieval methods. With the advent of pre-trained language models (PLMs) [3], [4], [5], [6], sophisticated representation learning approaches [7], [8], [9], [10], [11] have been employed. These techniques are adept at capturing the intricate semantics of both queries and documents, producing superior representation vectors. Such advancements have notably enhanced the search quality within the index-retrieve-rank framework.