KAIST and Microsoft cut a grep agent's input tokens 57% by mapping named entities first

Give an agent grep, find and cat over a flat directory and it starts each query with no pointer from one file to another ^[4]. A dev.to write-up of "Follow the Entities: A Corpus Map for Agentic Search" (arXiv:2609.37226) ^[1] says the model has to rebuild the relationships between documents from scratch, at inference time, on every query ^[4]. The write-up's own example is an agent that reached turn six before it learned that one project's approval slip, specifications and status report sat in three different folders. Getting there cost about 200,000 tokens ^[14].

Per-query counts understate the gap, because the cheaper configuration was also right more often. Divide input tokens by correctness on EnterpriseRAG-Bench and raw grep costs about 332,500 tokens per correct answer ^[1]. CorpusMap costs about 119,400 ^[2], roughly 64% less ^[3]. The free-form wiki spent fewer tokens per query than raw grep but about 339,300 per correct answer ^[4].

The folder-index baseline roughly doubled per-query input on EnterpriseRAG-Bench ^[5]. Reading twice as much to score lower takes some effort. The write-up's explanation is that folder trees follow team charts or file formats, so the agent reads redundant directory summaries before it reaches evidence ^[15]. The wiki failed differently. Without strict grounding against the raw files, free-form wikis suffer hallucinations and missing cross-references, according to the write-up ^[16].

The layer that won is the most constrained one. I think that is the right design. Pages exist only for entities that recur across independent documents: people, projects, systems, vendors, code modules ^[8]. Each backlink is verified against a source file ^[9]. The agent keeps its terminal tools and still opens the raw documents to check what a page claims ^[10]. The authors argue that fixed retrieve-then-generate pipelines such as GraphRAG and HippoRAG cannot inspect downstream sources once the initial graph walk misses an edge ^[7].

Whether the 57% transfers depends on the workload ^[11]. The numbers come from two benchmarks, EnterpriseRAG-Bench and WixQA [c2, c3], with GPT-5.5 as the model on the EnterpriseRAG-Bench run ^[11]. A similar saving elsewhere needs questions that turn on entities spread across several files. A corpus of standalone documents gives the extractor little to link. The map also has to stay current as files change, since a stale backlink sends the agent to the wrong document. And the build has to pay for itself. The write-up does not give the token cost of the offline extraction pass ^[9]. At 118,400 input tokens saved per EnterpriseRAG-Bench query ^[6], the break-even query count is that extraction bill divided by 118,400.