RAG retrieves. Something has to make the corpus worth retrieving.
Your local model is fine. The corpus is not. Teams often blame “RAG” or buy another retriever. Retrieval usually did its job: it searched what you gave it. The skipped job is knowledge-base manufacture — making a corpus worth retrieving from, and being honest about its limits. We are upstream of RAG, not opposed to it.
Three jobs people conflate
The market conflates three different jobs into one word. Separating them is the teaching move:
1 · Extraction
“Get the words out of the file.” Tools people know: LlamaParse, Docling, house scripts. Yield: raw text / markdown. Sound for that job. Not a knowledge base by itself.
2 · Retrieval
RAG in the strict sense — “Find the relevant passage at query time and put it in context.” Tools people know: vector search, hybrid RAG, rerankers. Yield: a passage in the prompt. Sound after the corpus is fit to search.
3 · Manufacture — almost none
“Make a durable corpus worth retrieving from — mute what must not be quoted, list the holes, leave a package you can keep.”
Tools people know for this box: almost none. That empty cell is the argument. It is why answers stay junk after the box is up.
You already own the first two boxes. We are pointing at the third. Brand words for that box on our door: mill, brick, plant. Learn the category in plain language first; the brand is flavor on top.
A skeptic will say manufacture is “just good ingestion” that RAG guides already cover. Cleaning and chunking are real — and not enough. The wedge is the owned leave-behind: a portable ZIP with junk muted and listed on the LIBRARY_CARD and in chunk flags (that listed mute story is the residual board), not a hosted index you rent and hope. Prep folded into a pipeline is still a service. A brick is an artifact you keep.
What “dump RAG” gets wrong
People stretch “RAG” to mean the whole pipeline: ingest, chunk, embed, search. Under that umbrella, dumping the SharePoint into an index is “doing RAG.” Strictly, retrieval only searches. The failure mode we refuse is dump-RAG — treat the raw share as if it were already a knowledge base. That path feeds ghost data and hallucination fuel into the prompt with confidence.
So we do not teach “not RAG.” We teach not dump-RAG: do not skip manufacture. Companion page: Keep your RAG. Point it at a brick.
Why an artifact, not another index
Manufacture yields a knowledge brick you own — not another vector DB you subscribe to. You can hand the ZIP to any model you already run, read its documented limits, mute one locus without boiling a company index, and keep the packet when the vendor relationship ends. That is what folder-plus-dump-RAG does not give you.
Argue with the package
Same mill: Public brick shelf · HF retrieval eval. Open a ZIP. Re-run the cosine hits. If you disagree, argue with the package — not with our adjectives. We are not the only plant that could mill; we claim the method is sound and competitive.
Simple at the door
Complexity stays in the automated plant. You drop files in the bin on the mill. No staffed project required to try one pile. $10 a brick are manufacture difficulty. Why bricks, not one melt: so you can compose and swap them — human-sized blast radius, not one unbounded soup.