Our on-prem model is quoting scanned procedures it cannot read.
You stood up a local model so the files would stay inside the building. Someone pointed it at the maintenance share. Demo day was fine. Week two it cites a torque spec, a warning, or a step that is not on the page — or it stares at a picture of a page and guesses. The model is usually fine. A folder of scanned manuals is not a knowledge base.
How we see the problem
Most of the “corpus” in a shop, a field team, or a training locker is not text. It is a picture of text: a TM, an FM, a procedure book that went through a copier in 1998, a PDF that is only images. The on-prem stack did what it was sold to do. It chunked the file. It embedded the chunks. It searches them. It cannot read a page that was never turned into honest words, and it will not tell you which pages are still pictures. That is ghost data with a shop-manual face.
- PDF container · “400 pages indexed”
- Image layers · figures · 1998 copier dirt
Looks like a manual corpus. Often isn’t text.
- Semantic void → invent a typical torque
- Or OCR mash → quote as if it were doctrine
Ghost data / hallucination fuel — still sure.
What you already tried
OCR then RAG the dump
You ran an extractor. You got some words out. You threw the markdown at RAG. The model still quotes doubled letters, a caption with no figure, or a footer that ranked like a heading. Retrieval searched the mess. Extraction got words out. Neither left a package a person can trust on the floor.
How we mill it
We mill a knowledge brick: a portable ZIP you keep. Pages that can be read are structured. Known junk is taken off the answer path and listed (LIBRARY_CARD and muted chunk flags — the residual board) — not hidden, not rewritten into a fake-clean manual. You point your existing local model at that package. You do not replace the stack. You do not start a chatbot. You pay only if we produce.
This is not “we OCR’d it.” OCR is one station — a sound tool upstream. The product is a residual-honest brick a person or a model can actually use — including the holes. Sound next to “OCR then RAG the dump”; not a claim that no other plant could mill scans.
An analogous brick (not your manual)
We do not have your TM. We do have scanned procedure-class stock from the same kind of pile. Open one. The residual is the lesson — not a highlight reel.
FM 7-85 — mid-size field manual
FM 7-85 Ranger Unit Operations — figures, multi-section doctrine, the kind of PDF a plant library holds. We mill it text-first. A figure dump once made a huge ZIP nobody wants; that is a trap, not a product. Readable package plus muted stretches listed on the card.
Galileo Dialogo — hostile OCR boundary
Galileo’s Dialogo — long scan, OCR that breaks words, plates, furniture. We still ship it. We do not pretend it is a critical edition. If your worst manual looks more like that than like a clean digital TM, that is useful to know before you drop the pile.
Same mill:
Public brick shelf
· OCR-hard retrieval stress on
HF
(nasa_skylab — our scan-hard sample on the eval set;
re-run cosine). Your shop manual is a different job. The method
is the same.
What this costs, when it is your pile
A clean short manual is often Small. A scan-heavy procedure book is Medium or Hard — manufacture difficulty, not a vanity page count. Class can move if you take files out of the hopper. Why $10.
Drop the pile on the mill. You keep the ZIP. We do not host your library. After the download window we purge (72 hours). If we cannot produce an honest brick, there is no charge. We are not a chatbot, not a vector database, and not a rewrite service. Last word on whether it is safe to use in your shop is still your trusted people.