Keep your RAG. Point it at a brick.
You are not required to throw away retrieval to fix on-prem answers. You are required to stop treating search over a dump as the whole data plan. Mill a residual-honest knowledge brick first. Then point your existing RAG — or your model’s tools — at that package. Keep the retriever. Change what it is allowed to see.
How we see the problem
This page is the companion to RAG retrieves — something has to make the corpus worth retrieving. That card names three jobs people conflate (extract, retrieve, manufacture). This card is only about the handoff: manufacture first, then retrieve. Careful engineers who hear a bare “not RAG” brace for a holy war. We are not in that war. Retrieval is sound after the corpus is fit to search. The failure mode is dump-RAG — chunk the share, embed, hope. That path feeds ghost data and hallucination fuel straight into the prompt.
- Raw share → vector index → prompt
- Noise, old years, scan debris in
Retriever did its job. The folder was the plan.
- Files → plant → knowledge brick
- Your RAG / tools → that ZIP
Keep the stack. Change what it may see.
What you already tried
When answers fail, teams often reach for better retrieval — not a better corpus:
Hybrid search, rerankers, bigger embeds
Useful on clean stock. They do not invent a mute list.
GraphRAG, chunk-boundary knobs
Same story. If the index is a melt or a scan dump, better retrieval retrieves the wrong thing more efficiently.
Fair comparison
Dump-RAG
Optimize ingest and search over whatever landed. Fast demos; weak trust.
Chat the PDF / giant context
Different job. Temporary session. Not a durable KB.
RAG over a milled brick
Same retrieval machinery, honest leave-behind underneath — including a mute list dump-RAG never invents. Simpler at the door than building a full internal plant: drop the bin, keep your stack.
How we mill it
KBMill manufactures the brick: portable ZIP; junk muted and
listed on the LIBRARY_CARD and in muted chunk flags (that
is the residual board);
craft_brief.md and
SECURITY_REPORT.md in the ZIP.
Standard sidecars (kb.json,
chunks.jsonl, embeddings.npy) so
your loader can ingest without a proprietary runtime.
Point your model and your RAG at it. We do
not host your chat. You
pay only if we produce.
Proof of shaped packages worth retrieving — cosine top-hits on milled stock, not vibes: Public brick shelf · HF retrieval eval (re-run it).
Simple at the door
You can try one pile without a staffed project. Drop files in the bin on the mill. Complexity stays in the automated plant. $10 a brick are manufacture difficulty.