Build your company corpus one brick at a time

Note · KBMill · company corpus

A company knowledge base is a set of packages you can stand behind. Most of that work never gets done, because it is priced and staffed like a large project. It does not have to be.

On kbmill.com you build it one brick at a time. Name the parts of the business. Mill the files that matter for each one into a portable ZIP you keep. Use those packages together. Do not pour every file into one giant index.

The picture decision makers can hold

You already know the parts of the business — ops, policy, product lines, training, compliance, whatever you actually run on. Name them. For each one, choose the files that matter. Mill a knowledge brick: a portable package you keep. Point your own model or search at the set. That set is the company knowledge base, grown over time — not one giant index someone promised on a slide.

Concretely, a finished brick is a ZIP you own — structured stock, not a chat session. Inside: human-readable Markdown; machine sidecars (kb.json, chunks.jsonl, embeddings.npy); a short craft_brief.md; and the mute list of what we took off the answer path (listed, not hidden). Full inventory on why these prices. We do not host it as your library; you keep the package and point your stack at it.

Most people want the knowledge base, not a second job learning how the mill works. The notes are here when you need them. They are not homework before the first drop. The plant holds the complexity. The door stays plain.

Initial build — what does the first package cost?

You are buying one package made from your files. Not a few-dollar parse. Not a six-figure cleanup program. On this door:

• $10 a brick — any pile this door accepts, only if the ZIP ships. A scan-heavy pile can take longer. The price does not change.

A short manual and a long scan are the same $10. You pay only if we produce a usable ZIP. Details live on why $10.

Elsewhere, this work shows up as people and projects: a staffed cleanup in the thousands or tens of thousands, or an internal “we’ll clean the docs” effort that never becomes a package you can keep. Public write-ups of an enterprise retrieval system price a different bill: engineering, a vector store, tokens, and retainers. Those numbers often run tens to hundreds of thousands. That is the cost of standing up the system. This door sells the package. $10 a brick, only if the ZIP ships.

Scale — how does this grow with our problem?

Grow by domain, not by dumping the whole file share into one index. First domain that hurts, first brick. Next domain, next brick. A harder pile can take longer. It does not change the price, and it does not touch the packages you already have. This door caps a job at 15 files or 150 MB, one job at a time. A larger program is more jobs, or a quoted bundle when you need one.

Using those packages together is your job, on your stack, after you take the ZIPs home. We mill the packages. You decide how they are used together.

Sustain — what does it cost to keep this real?

Information is never static. The expensive pattern is a full rebuild of the entire knowledge base whenever anything important moves. That bill is so large that refresh slips and the corpus goes stale.

When a domain changes — a manual, a policy, a product line — you remake that brick. The others stay as they are. Sustaining cost tracks what moved, not how big the company KB has become. That is how a federated corpus stays current without reprocessing everything.

Choose material on purpose

“One brick at a time” only works if each brick is intentional. Before you drop a pile, decide:

• Which domain first? — where a wrong answer hurts, and where the model will be asked soon
• What belongs in this brick? — the files that define that domain. Leave the neighboring domains for their own packages
• What stays out? — page chrome, obsolete copies, and unrelated folders that would let the wrong file answer

That choice is the difference between a shelf you meant to build and another random drop. This note is the short version. On the pilot, we can help you pick the first domains. A mill job is still a mill job, not a consulting engagement.

What we are not claiming

One job is not the whole company knowledge base. A large file share does not finish before lunch. The door caps are there on purpose. Air-gapped work, and work past those caps, is a separate quote. We do not host the set of ZIPs as your library. We are not pricing the retrieval system those other write-ups price. Weak spots stay listed in the package.

Door

If you approve spend, ask for a company corpus built as domain packages — a price per package, a way to grow the set, and a way to remake what changed — not only a line that says “one giant knowledge base.” If you have a first domain that fits the caps, the mill is the door: drop, keep the ZIP, pay only if we produce.

Related (same plant, different doors): why $10 — and what’s in the ZIP · why bricks instead of one giant KB · what goes wrong with a melted index · time to a corpus you can defend · where are the documents on the on-prem estimate · answers you can stand behind.