Revisiting Text Ranking in Deep Research

The article presents a comprehensive evaluation of text ranking methods in deep research, focusing on the effectiveness of retrieval units, pipeline configurations, and query characteristics. Experiments conducted on the BrowseComp-Plus dataset involved two open-source agents, five retrievers, and three re-rankers, revealing that passage-level units are more efficient in constrained contexts, and that a proposed query-to-question (Q2Q) method enhances performance by mitigating query mismatches. This research is significant for practitioners as it clarifies the impact of various configurations on retrieval effectiveness, guiding the design of more efficient LLM-based search agents.