Follow the Entities: A Corpus Map for Agentic Search

Follow the Entities. A Corpus Map for Agentic Search. By Soyeong Jeong and three others, from KAIST and Microsoft. Published on the twenty-ninth of September. No code repository is linked to it.

Answering questions and completing tasks over large document collections often means connecting evidence spread across several documents. A project's approval is recorded in one, its requirements in another, its latest status in a third. No single document holds the answer. Recent LLM agents approach this by iteratively searching the full corpus, rather than reading only a fixed set of top-ranked documents.

The trouble is how the corpus is presented to them. Exposed only as a flat collection of files, a relevant document gives no indication of how it relates to the others. The agent has to rediscover those relationships for every query. In doing so it often misses complementary evidence, and at the same time it consumes substantial additional tokens. It is paying more and finding less. Every search begins as if no earlier search had ever happened, even though the same connections between documents hold for query after query.

CorpusMap is a navigation layer that organizes the corpus around its recurring entities. These entities are identifiable from the documents themselves, and a single one can link a document to many others, across sources. The project in the example is such an entity: it is what the approval, the requirements and the status report have in common. Find the entity, and the documents that mention it come with it.

Each recurring entity is represented as an Entity Page. The page aggregates information about the entity and links to every document that refers to it. Taken together, the pages form a graph between entities and documents, which the agent can traverse to gather evidence that would otherwise stay disconnected. The map is constructed offline, by resolving mentions of the same entity across documents. That timing matters. Because the links are built once, ahead of time, they are shared across queries rather than rediscovered repeatedly at inference time. The work of noticing that two documents concern the same thing is done one time, when the map is made, instead of again and again while an agent is trying to answer. At query time, the agent can land on one relevant document, step to the entity pages it mentions, and from there reach the other documents that complete the picture.

The evaluation uses seven different models and three benchmark datasets. Against agentic search over the raw corpus, CorpusMap improves both evidence discovery and answer quality, while using fewer tokens on average. It also outperforms four alternative navigation layers, which suggests that the choice of entities, specifically, is doing useful work. Better answers and lower cost arrive together, rather than one being traded for the other.

A flat folder of files makes every query start from nothing. The entities were already there, threaded through the documents. CorpusMap pulls that thread out once, in advance, so the agent can follow it instead of finding it again. The authors' conclusion is that entities serve as effective anchors for navigating large document collections.