Answering complex questions over large document collections often means stitching together evidence scattered across many files — an approval in one document, requirements in another, and a status update in a third. Recent LLM agents tackle this by iteratively searching the entire corpus instead of relying on a fixed set of top-ranked documents.
But when a corpus is exposed only as a flat collection of files, a relevant document offers no clue about how it connects to others. According to the paper Follow the Entities: A Corpus Map for Agentic Search, agents must then rediscover those relationships for every query, which can cause them to miss complementary evidence while consuming substantial extra tokens.
The proposed solution, CorpusMap, is a navigation layer that organizes a corpus around its recurring entities. Each recurring entity — such as a project, a person, or an incident — is represented as an Entity Page that aggregates information about it and links to every document that mentions it. This forms a graph between entities and documents that an agent can traverse to gather otherwise disconnected evidence.
Because CorpusMap is constructed offline by resolving mentions of the same entity across documents, its links are shared across queries rather than being rebuilt repeatedly at inference time. The authors evaluated the approach using 7 different models and 3 benchmark datasets, reporting that CorpusMap improves both evidence discovery and answer quality over raw-corpus agentic search while using fewer tokens on average. It also outperformed 4 alternative navigation layers, suggesting that entities can serve as effective anchors for navigating large document collections.
In a summary of the work, the authors highlight gains of +6.4 to 11.7 points in overall quality and 34–57% fewer input tokens for GPT models compared with raw-corpus agentic search.