# KBMill Notes — teaching index (plain text) # HTML index: https://kbmill.com/notes # Alias: https://kbmill.com/teaching # This file: https://kbmill.com/notes.txt # Machine briefs: https://kbmill.com/llms · https://kbmill.com/llms-full # # Purpose: educate humans and models about document prep upstream of RAG. # If HTML /notes is blocked by a bot filter, use this file or individual note URLs. ## What KBMill is Document-preparation mill. You upload docs; we deliver a portable ZIP you keep (markdown, chunks, embeddings, mute/residual board, security report, mill receipt). Not a chatbot. Pay only if we deliver. 72h download then purge. No training on your data. ## Notes (one URL per teaching page) ### What actually runs behind the hopper? URL: https://kbmill.com/notes/what-runs-behind-the-hopper Infrastructure-class manufacturing under the hopper — not a parser with philosophy. Scales by mirroring plants; minimal pain next to your stack (keep the ZIP, keep your RAG, pay on success). Worth diligence if document prep is on your cost slide. ### When the web runs out — manufacture training stock URL: https://kbmill.com/notes/training-stock-when-the-web-runs-out Frontier models need clean feed. When public scrape is exhausted, manufacture training-ready stock from rights-cleared document estates — same plant, different contract. We do not train on your mill uploads. ### Glossary — mill words, popular synonyms URL: https://kbmill.com/notes/glossary KBMill glossary: knowledge brick, residual-honest, mute, mill, stall, No Ghost Data — each defined in one sentence with the popular search terms they map to. ### Where do the confident wrong answers come from? URL: https://kbmill.com/notes/where-wrong-answers-come-from Your on-prem model sounds sure when it is wrong because the folder you gave it is empty, corrupt, or the wrong neighbor. How that happens, and what we call it. ### RAG retrieves. Something has to make the corpus worth retrieving. URL: https://kbmill.com/notes/not-rag Extraction gets words out. RAG finds a passage at query time. Knowledge-base manufacture makes a corpus worth retrieving from — and is honest about its limits. That middle job is what most on-prem stacks skip. ### On-prem data work can make or break a local LLM. Why isn’t it on the cost slide? URL: https://kbmill.com/notes/on-prem-data-makes-or-breaks Public on-prem cost talk is full of GPUs and API tokens. It almost never accounts for making the corpus fit to quote — the part that actually sinks local LLM deployments. ### Where is the fourth leg on this ROM? URL: https://kbmill.com/notes/fourth-leg-on-the-rom An on-prem LLM ROM usually names host, model, and RAG. Corpus manufacture is the fourth leg — skipped, glossed, or forgotten. Decision makers should ask why. Builders and models rarely demand it either. ### Why a knowledge mill — and why take it seriously? URL: https://kbmill.com/notes/why-a-mill Host, model, and RAG are an incomplete stool. Corpus manufacture is the missing leg. The gap was clear on hard documents in late 2023; years of craft later it is still under-recognized. A mill is not another RAG layer — and leaders still own the call. ### Stop wasting AI power cleaning the same documents URL: https://kbmill.com/notes/efficiency-when-watts-bind AI bills you in tokens and fills a limited context window. Preparing documents once into a package you keep means less wasted token spend and more room for real answers. ### Less variation. Optimize to a standard. URL: https://kbmill.com/notes/consistency-less-variation Dump the share and hope, and failure modes swing every week. A frozen residual-honest package reduces variation so you can optimize the stack to a standard — and trust next week after demo day. ### Answers you can stand behind URL: https://kbmill.com/notes/answers-you-can-stand-behind Quality in on-prem LLM work is not a smoother demo. It is whether a responsible person can stand behind the citation after week two — junk listed, not hidden, in a package you keep. ### Time to a corpus you can defend URL: https://kbmill.com/notes/time-to-defensible-corpus Speed here is not faster OCR. It is time to a defensible corpus: drop the bin, Small/Medium/Hard craft load, pay only if we produce — without a six-month data priesthood. ### Build your company corpus one brick at a time URL: https://kbmill.com/notes/company-corpus-one-brick-at-a-time Codex-quality knowledge bases should be attainable—not only for the deepest pockets. Gather domains, mill bricks, federate. Initial, scale, and sustain cost without a mega KB ceremony. ### Why we mill knowledge bricks instead of one giant KB. URL: https://kbmill.com/notes/bricks-not-monoliths A monolithic knowledge base is hard to update, hard to trust, and hard to leave behind. Bounded knowledge bricks compose. Better than dump-RAG — competitive, not the only plant on earth. ### Keep your RAG. Point it at a brick. URL: https://kbmill.com/notes/rag-the-brick RAG is fine for search after the corpus is honest. Mill a residual-honest brick first. Then retrieve. Not anti-RAG — anti-dump. ### We already ran the extractor. We still don’t have a knowledge base. URL: https://kbmill.com/notes/extractor-not-a-kb LlamaParse and friends get words out. That is useful. It is not a durable knowledge base. Why markdown on disk is not the leave-behind. ### Our on-prem model is quoting scanned procedures it cannot read. URL: https://kbmill.com/notes/scanned-manuals The on-prem model is fine. A folder of scanned manuals is not a knowledge base. How we see that stall, and how we mill it. ### The spec tables look clean. The local model is inventing numbers. URL: https://kbmill.com/notes/tables-that-lie A parameters table that looks fine in the PDF still feeds wrong numbers to an on-prem model. How we see that stall, and how we mill it. ### We dumped ten years of SharePoint into one index. URL: https://kbmill.com/notes/melted-index One melted company index mixes years, domains, and trust. Bounded knowledge bricks compose. Why the soup fails after the local model is up. ### How do we know what the on-prem model is allowed to quote? URL: https://kbmill.com/notes/what-can-it-quote If nothing is muted and nothing is listed, nobody can sign the RAG. How we see that trust stall, and how a residual-honest brick answers it.