Follow the Entities: A Corpus Map for Agentic Search

This paper addresses the inefficiency of LLM agents iteratively searching flat document collections for scattered evidence. The authors introduce CorpusMap, a navigation layer that organizes a corpus around recurring entities identifiable directly from the documents. CorpusMap represents each recurring entity as an Entity Page aggregating related information and linking to every referencing document, forming an entity-document graph that agents can traverse to gather disconnected evidence. Using 7 different models across 3 benchmark datasets, the results demonstrate that CorpusMap improves evidence discovery and answer quality while consuming fewer tokens on average compared to raw-corpus search.

CorpusMap organizes the corpus around recurring entities to link documents across different sources.

The navigation layer is constructed offline, sharing links across queries instead of rediscovering them at inference time.

It improves both evidence discovery and answer quality while using fewer tokens on average.