CorpusMap builds an offline entity graph for agent search
TL;DR
- CorpusMap builds an offline graph linking every document to the entities it mentions, so an agent reuses the same map across queries.
- Across seven models and three benchmark datasets, the abstract reports better evidence discovery and answer quality using fewer tokens on average.
- The method also outperforms four alternative navigation layers, the authors argue, positioning entities as anchors for large document collections.
An agent working across a large document collection typically has to rediscover how those documents relate to one another on every query. The paper Follow the Entities: A Corpus Map for Agentic Search, by Soyeong Jeong, Sujay Kumar Jauhar, Sung Ju Hwang, and Andrew Joohun Nam, argues this rediscovery is wasteful and proposes moving the work to ingest time.
They call the resulting structure CorpusMap. It "represents each recurring entity as an Entity Page that aggregates information about it and links to every document that refers to it, forming a graph between entities and documents that the agent can traverse to gather otherwise disconnected evidence." Because the graph is assembled offline by resolving mentions of the same entity across documents, its links "are shared across queries rather than rediscovered repeatedly at inference time."
The evaluation runs across seven different models and three benchmark datasets. On those, the abstract reports that CorpusMap "improves both evidence discovery and answer quality over raw-corpus agentic search while using fewer tokens on average," and further outperforms four alternative navigation layers.
Neither the seven models nor the three benchmark datasets are named in the abstract, and no per-model or per-dataset numbers appear there either. It joins seven other RAG papers we've tracked in the last ninety days.
Originally reported by huggingface.co
Read the original article →Original headline: HF paper 'Follow the Entities' builds an offline corpus map to speed agentic search