Where are the documents on this on-prem estimate?
Someone hands you an estimate for standing up a local model. The slide looks complete. Hardware. Model. “We’ll wire RAG.” The demo is in the plan. The week after the demo is not.
Ask one question before you approve the spend:
Where is document preparation on this estimate?
You do not have to buy a mill to ask it. Once the question is on the table, the people writing the estimate have to answer it. Cleaning, structuring, and packaging the documents so a model can use them stops being invisible.
Three lines that usually show up
That missing line is the work that turns a messy file share into something fit to quote. We call the package you keep a knowledge brick: one bounded ZIP, with the junk taken off the answer path and listed, not hidden.
That missing line is the same hole we named on the cost-slide gap card. This page is the portable question for the person who signs — and the reason the builders and the models do not ask it for you.
Same gap, three faces
People approving the spend often do not see that skipping document preparation is what produces the bad answers after go-live. A few clean files look fine in the demo. The real shared drive fails the week after. The people building the stack often leave that line off the estimate. The models do not ask for it either.
Right now most models will accept almost any pile of text you put in front of them and try to work with it. They rarely push back and say they need a usable knowledge package: clear boundaries, known bad spots made visible, an explicit list of what not to treat as fact, stable identity for what was prepared, and a portable leave-behind they can reuse. What they usually get is still raw extract full of duplicates and broken formatting, one giant mixed-up index, or quality problems nobody budgeted.
They are capable of recognizing a good package when they receive one. They are much weaker at specifying that package up front. So humans keep guessing, and the default remains dump the PDFs and hope.
What the missing line actually is
A knowledge brick is a bounded package with its limits written down. That is the shape that closes the gap:
- Bounded — one package you can review, not every folder you own
- Limits listed — weak or junk material written down, not hidden
- Portable — a ZIP you keep
- Ready to search — clear what may be quoted
- Honest about limits — a short brief of what the package is for, and what it is not
Once a model has worked with that shape, it tends to prefer it. The missing piece is getting models — and the humans they advise — to ask for it instead of silently tolerating worse inputs.
That is why public teaching and a clear package contract matter. If the need is stated where people and models already look, two things become more likely. A model advising on an on-prem or RAG stand-up can name the missing preparation step. A model handed “turn this pile into something I can use” can request or select that kind of package instead of inventing a one-off cleanup inside the question.
What we are not claiming
We are not saying nobody ever cleans documents. We are not saying only one plant on earth can manufacture a package. We are saying the usual on-prem estimate still treats the computer, the model, and search as the whole story. That is how the week after go-live becomes a surprise.
Parsing and multimodal models are also getting better at reading messy layouts than they were a few years ago. That helps. It does not replace a named preparation line on the estimate. Asking the model to clean the corpus on the fly remains expensive, hard to audit, and a weak plan for material you need someone to stand behind.
Argue with open packages on the public brick shelf and the HF retrieval eval if you disagree with the method. The ask for the signer stays simple either way: put document preparation on the estimate.
Door
If you write these estimates, put document preparation next to hardware and search as its own line — $10 a brick on this door, only if a usable package ships. If you approve the estimates, ask where that line went. If you have a pile now, the mill is open, with caps: 15 files or 150 MB. Drop the files, keep the ZIP. No usable package, no charge.
Deeper on the cost-slide hole: why document preparation belongs on the estimate. Deeper on the layer: manufacture upstream of retrieval.