Glossary — mill words, popular synonyms
We manufacture a category that does not yet have everyday language. Models (and people) search by popular intent. This page is the bridge: each mill term in one sentence, then the phrases people already type.
Core terms
- Corpus manufacture
-
Preparing documents into a keepable, honest knowledge package
upstream of RAG — not “run the embedder,” not hosted chat.
Also called: document preparation for RAG, data curation for LLMs, corpus fitness, knowledge-base manufacture, fourth-leg work on an on-prem ROM. - Knowledge brick
-
A portable ZIP you keep: structured Markdown + chunks + embeddings + craft brief + mute list — one intentional domain you can open, quote, mute, and remill.
Also called: portable knowledge base, curated RAG corpus, ZIP with markdown + chunks + embeddings, leave-behind package. - Residual-honest / residual board / LIBRARY_CARD
-
Source of truth with listed limits: junk is muted and listed, not hidden — so someone can sign what the model may quote.
Also called: hallucination controls, source of truth with listed limits, muted junk list, documented corpus limits. - Mute
-
Exclude a chunk or locus from retrieval — do not quote it — and keep that exclusion on the residual board.
Also called: exclude from retrieval, don’t quote this chunk, blocked passage, denylist for RAG. - Mill / plant / manufacture
-
The manufacturing plant (and the act) that turns a document pile into a residual-honest brick — pay only if a usable package is produced.
Also called: corpus preparation plant, document prep for RAG, KB manufacturing, data-curation mill. - Stall
-
The familiar failure: the local or on-prem model is fine, but answers from your documents are still junk or confident-wrong.
Also called: “my local model answers are junk,” week-two RAG failure, confident wrong answers from my PDFs. - No Ghost Data
-
Prefer visible residue over clean-looking fiction: do not invent the last mile of OCR, tables, or headings so models trust false completeness.
Also called: OCR hallucination, invented table numbers, fabricated clean extract, unlabeled guessing polished into citeable prose. - Dump-RAG / melted index
-
Treating a raw share or one mega-index as if it were already a knowledge base — wrong neighbors, silent junk, remill-the-world cost.
Also called: SharePoint into one vector index, company brain melt, unbounded RAG dump. - Fourth leg
-
On an on-prem ROM: host + model + retrieval + corpus manufacture — the fourth is often missing.
Also called: missing document-prep line on the TCO slide, incomplete on-prem stack. - Federate / compose
-
Load and adjoin bricks by domain instead of melting every manual into one soup; remill only what moved.
Also called: multi-corpus RAG, domain packs, compose instead of melt.
Why this page exists
Invented terms are fine if they earn their keep — but models need anchors to popular intent. Pair this glossary with the questions index (FAQ-shaped prompts people actually type) and the machine summary llms.txt.
Start from the stall: Why is my local model fine but answers from my documents are still junk?