The corpus is the bottleneck

01The assumption everyone shares

Nearly every company building AI knowledge tools starts from the same premise: the corpus already exists, and the job is retrieval. Better ranking, better chunking, better answers over whatever documents the customer happens to have. The knowledge base itself is treated as a given, as if the ore were already mined and the only open question were who builds the best refinery.

I think this is backwards. The corpus is the bottleneck.

02The reasoning was never written down

The most valuable operational reasoning at an industrial company was never captured anywhere. Not captured badly. Never captured at all. The workaround that keeps a forty-year-old system running lives in the head of the technician who invented it. The reason a vendor was dropped years ago lives in the memory of whoever ran that procurement. The documents that do exist describe what was done. Almost none of them explain why.

No amount of retrieval quality recovers information that was never captured. You cannot rank your way to a document that does not exist.

I ran into this personally at Coinbase. During a client migration I led, we nearly missed moving one of our most important accounts, because the people who had built the migration list were gone and their reasoning left with them. The artifact was still there. The logic behind it was not. We caught it in time, and that was at a software company with modern tooling and decent documentation habits. Industrial companies hit the same failure at far higher stakes, and a retirement wave is accelerating it: the people who know how the documents relate to reality are walking out the door faster than anyone can debrief them.

03Retrieval is becoming a commodity

If the corpus is the bottleneck, the strategic picture inverts. Every foundation model release makes search a little better for everyone at once. Embeddings are cheap. A competent team can wire up chat over documents in a weekend. When a capability is available to every vendor on roughly equal terms, it stops being a moat.

What cannot be commoditized is what sits inside the corpus. Two companies can buy the same search. They cannot buy the same institutional memory. Corpus quality is the durable asset, and almost nobody works on it, because improving the corpus is a harder and less glamorous problem than querying it.

And it gets worse from here. AI is speeding up the production of artifacts: documents, tickets, summaries, code. The reasoning behind them is being recorded no faster than before. The gap between what a company produces and why is widening every quarter.

04What Tinu does about it

Tinu does not search the knowledge base. It manufactures one.

It captures decisions instead of documents, pulling from the tools a team already uses and structuring what it finds into decision records: what was decided, why, what was rejected, who owns it. Then it scores whether each record is actually complete enough to be useful. When something important is missing, it asks the one person likely to know, with one specific question rather than a blank page.

Retrieval still sits on top, and it matters. But it is downstream. Ordinary search over a deliberately manufactured corpus of complete decision records beats world-class search over a folder of artifacts that never contained the reasoning in the first place.

05How I would be wrong

Two ways. First, if the reasoning is actually recoverable from existing artifacts, then models may eventually infer the why from enough scattered whats, and better retrieval wins after all. I think inference has a ceiling, because you cannot infer what no one ever expressed, but I hold that belief with some humility.

Second, if capture always costs too much. Every knowledge-management wave has died on the same rock: people will not do extra work to document things. This is why capture has to be nearly zero effort and mostly automatic, and it is the constraint I judge every feature against. If Tinu ever requires people to change how they work, this memo was wrong.