Why a knowledge mill?

Note · KBMill · the missing line

Most on-prem plans name three things: the computer, the model, and “we’ll wire RAG.” That list is incomplete. Something has to make the documents fit to quote. That work is what this mill does. Leave it off the estimate and the failure shows up after the demo, when the real files are in front of the model.

KBMill is that door. It is not another search layer. The gap was already obvious on hard documents in late 2023. The packages on the shelf are the proof.

Where the journey began

Late 2023 and early 2024, the pressure was hard documents — the kind that do not forgive a demo on clean samples. RAG, as it was being delivered, had clear failure modes:

Sludge in the extract, then confident wrong answers
A raw folder treated as if it were already a knowledge base
A model that looked fine choked by a polluted corpus

That was not a missing button on a chatbot. It was a structural miss in how the stack was being sold and stood up. First code for what became this plant came from that recognition — from the gap on the desk, not from a marketing brief that needed an AI feature.

What has not changed

Here is the surprising part. Through all the iteration since — roughly two and a half years of grinding on the failure modes — the structural gap has not closed. Host, model, and retrieval still dominate the story. Manufacture is still under-named. Week-two junk, polluted indexes, and cost slides quiet on documents are still all around us.

People are only just beginning to recognize it. That is not because latecomers are stupid. The map was missing. The old gospel felt complete. Evidence can sit in plain sight when nobody has settled words for the missing leg.

How we saw it

We are not claiming a private genius. We stumbled onto the signals in a way that made the problems clearer: hard corpus, time on the desk, and enough isolation from the demo loop that week-two failure could not be papered over with a smooth walkthrough.

If someone else had been looking that way — isolation, hard material, refusal to shrug the sludge — they would likely have noticed the same thing. The gap was in the documents and in the incomplete stool, not in a secret insight. What we did next was stay with what we saw and build a door.

Why a mill, not another RAG product

Three jobs get conflated. Keeping them separate is the whole shift:

Retrieval finds a passage at query time. It does not by itself turn a messy estate into something a responsible person can stand behind.
Extraction gets words out of files. It is not yet a leave-behind with listed limits.
Manufacture is the middle job: bounded packages, known junk muted and listed, something you keep and point your existing stack at.

Keep your model. Keep your search if you have it. Those do not prepare the documents. The question to ask is simple: where are the documents on this estimate? The layer underneath is here: search retrieves; something has to make the files worth retrieving.

What you can check

The plant has been built against hard documents since late 2023. What you can check is on the shelf:

Weak spots listed, not hidden
Pay only if we produce a usable package
A public proof shelf you can open and a retrieval eval you can re-run

You can run one job on your own files before you fund a staffed cleanup. $10, only if the ZIP ships. You keep the package.

Door

If the estimate names the computer, the model, and search, and says nothing about the documents, start at the mill or the questions we answer. Open a package on the shelf and see.

The pilot is intentional and bounded: drop your files, keep the package if we produce — no honest brick, no charge.