Follow the Entities: A Corpus Map for Agentic Search

Vince Okay, that is the annoying part right there. The agent already has full-corpus access, but it still has to rediscover the same document links every time, which is just token burn with better branding.

Ava Yeah. A flat corpus is basically a pile of files with no topology. If the approval is in one doc, the requirements are in another, and the status lives somewhere else, the model is doing graph discovery from scratch on every query.

Vince And that’s the product angle I actually like. CorpusMap is saying, stop making the model be a little archaeologist every single time. Build the map once, let it reuse the structure, and ship something that feels faster because it is faster.

Ava Right, and the interesting move is that they don’t invent a new corpus abstraction at query time. They resolve recurring entity mentions offline, then turn each entity into an Entity Page that points to every document referring to it. So the graph is shared across queries instead of re-learned.

Vince That part matters a lot. If I’m reading a project tracker and it mentions a codename, I don’t want the agent to shrug and go search the whole world again. I want it to follow the codename to the email, the notes, the spec, whatever actually closes the loop.

Ava Exactly. The paper’s mechanism is pretty plain, which I mean as a compliment. Identify mentions, resolve the same entity across documents, attribute facts back to sources, and keep links to every doc that talks about that entity. The original documents stay in place, so this is a navigation layer, not a replacement corpus.

Vince Mm-hm.

Ava Then at inference time the agent can traverse entity-to-document and document-to-entity links from whatever it already found. That means the query doesn’t have to explicitly name the hidden connection. If a document mentions a project, the agent can walk from that project page to the other docs that mention it and collect the missing evidence.

Vince That’s the bit that makes it feel shippable instead of cute. It’s offline, it’s reusable, and it maps to stuff teams already have: projects, people, incidents, products. Honestly, this is the opposite of one of those platform launches nobody can picture using.

Ava Also, the evaluation is not just one happy-path demo. They test seven models across three benchmarks: EnterpriseRAG-Bench, WixQA, and HERB. CorpusMap beats raw-corpus agentic search and also outperforms four other navigation layers, including LLM Wiki and Corpus2Skill.

Vince No way.

Ava And the gain is not subtle. They report about six point four to eleven point seven quality points up, while using thirty-four to fifty-seven percent fewer input tokens on average. So you’re getting better evidence discovery and less repeated search work at the same time.

Vince That’s the thing I’d actually bet on. If you’re running enterprise search over a messy corpus, the cost isn’t just retrieval, it’s re-discovery. Every extra hop is another chance for the agent to miss the complementary doc that makes the answer click.

Ava Sure, but I do want to keep one eyebrow up. Entity resolution is doing a lot of work here. If your corpus has ambiguous names, aliases, bad metadata, or a ton of near-duplicates, the map can get noisy fast. Then the nice graph turns into a confident wrong turn.

Vince Yeah, but that’s also why this feels more practical than a lot of agent-search papers. They’re not asking the model to become smarter in some mystical way. They’re giving it better structure around the thing it already does badly, which is remembering where the evidence lives.

Ava Exactly. And because the links are built offline, you can imagine real production constraints around it: index refreshes, entity merge rules, maybe manual review on high-value corpora. That’s boring, but boring is what makes the thing survive contact with users.

Vince Okay, that’s such an Exploring Next take. But you’re right. This is the kind of layer that could sit under an enterprise assistant and quietly make the whole system feel less stupid.

Ava And I like that they’re not claiming some grand new reasoning breakthrough. It’s a navigation move. The paper’s basically saying the corpus already contains a graph, so stop pretending a flat folder is the right interface for agentic search.

Vince That’s the line, honestly. Not a smarter model, just a less ridiculous path through the documents.

Ava Yeah. If I were building this, I’d care most about the entity resolution quality and the refresh pipeline. The architecture is sound, but the operational win lives or dies on whether the map stays trustworthy as the corpus changes.

Vince Still, this is one of the cleaner shippable ideas we’ve seen in a while. Not flashy, but it gives the agent receipts, and that tends to matter once someone actually has to use it.

Ava Mm-hm.

Vince All right, Ava, I’m calling it a good Wednesday if we spent it arguing that files need a map. Honestly, that feels on-brand for us.