Somewhere in your files is the answer to the question a customer asked this morning. It is in a manufacturer bulletin from last spring, or section 4.2 of the warranty terms, or a spec sheet attached to an email three jobs ago. The person who took the call did not go find it. They answered from memory, because finding it would have taken twenty minutes nobody had.
The filing-cabinet version of how this works
A general-purpose AI has read the public internet. It has never read your warranty terms, your price book, or the bulletin that changed your fastener pattern last spring. Retrieval is the plumbing that fixes that. When a question comes in, the system does not answer from general knowledge — it goes to the filing cabinet first, pulls the three paragraphs across your documents that actually bear on the question, and drafts the answer from those paragraphs. Then it shows you which pages it used.
Take a real question: does the 30-year shingle warranty cover wind damage on a low-slope tear-off? Without retrieval, a model guesses from general roofing knowledge — confident, plausible, possibly wrong. With retrieval, the answer is composed from your warranty's wind-exclusion section and the manufacturer's minimum-slope spec, and both sources appear under the answer with a file name and a page number.
Why every answer carries a file and a page
The citation is not decoration. It converts an answer you have to trust into an answer you can check in ten seconds. "Covered, subject to the 130-mph limit — 2026 Warranty Terms, page 41, section 4.2." The person asking can open the page, read the paragraph, and confirm before repeating it to a customer. On an ordinary day that check takes seconds. On the day a claim is disputed, the citation is the whole point: you do not want the answer, you want the page.
There is a second, slower effect. After a few weeks of answers that name their sources, staff learn where things live — which manual covers what, which document governs which question. The system ends up teaching the filing structure it searches.
One rule follows from this, and it is worth insisting on: no source retrieved, no answer given. When the documents do not contain the answer, the honest output is "I could not find this in the documents" — which tells you either that the answer was never written down or that the document never made it into the system. Both are things you want to know.
How a document gets cut apart, and why you should care
Documents are stored in pieces a few hundred words long, because the system retrieves passages, not entire 400-page manuals. Where those cuts fall is where these systems quietly break.
Cut a warranty clause in the wrong place and you can invert its meaning. The catastrophic case is short enough to memorize: a stored fragment that begins "...this exclusion does not apply," retrieved on its own, severed from the paragraph above it that says which exclusion, under what conditions. An answer built from that fragment states the exact opposite of what the document says — delivered in a confident tone, with a citation attached that makes it look verified.
The prevention is unglamorous. Cut on section boundaries instead of every few hundred words. Keep tables in one piece. Keep each clause attached to the heading that scopes it. Stamp every fragment with its origin — document name, section, effective date — so that a retrieved piece carries its own context wherever it goes. None of this is exotic engineering. It is the difference between a system that embarrasses you in front of a customer and one that does not, and it is why "we indexed your PDFs" is not a sufficient description of anyone's work.
The 2019 price sheet problem
Retrieval finds the passage most similar to the question — not the most current one. Ask what you charge for a 40-gallon gas water heater install, and if the 2019 price sheet and the 2026 price sheet are both in the system, the old one is every bit as similar to the question as the new one. Sometimes more similar, if its wording happens to match yours. The system will quote seven-year-old prices with total confidence and a perfectly accurate citation to an obsolete document.
The fix is metadata and discipline. Every document carries an effective date. Superseded documents are marked superseded. The retrieval step filters to current documents unless someone explicitly asks for history. And — the part that involves no software at all — when your supplier reprices in March, loading the new sheet and retiring the old one has to be someone's actual job. A document system without that habit decays into a very fast way to get old answers.
What goes in, and what shape it arrives in
Native digital PDFs extract nearly perfectly. A scan of a coffee-stained bulletin photographed in a truck cab may extract into garbage. Price tables are the most fragile of all: a table pulled apart into an undifferentiated wall of numbers will produce confidently wrong prices, which is worse than no answer.
So part of the setup is triage, not technology. Which documents are clean digital originals. Which are scans that need character recognition and a human check. Which are worth retyping. A short stack of clean, current, correct documents beats a large pile of everything — and the sorting is work the business does once, with help, at the start.
What this does not fix
It reduces wrong answers. It does not eliminate them. Retrieval controls what the model reads; the model still composes the reply, and composition can go wrong. Anything high-stakes — a warranty position, a price, a spec — gets checked against the cited page before it goes out the door. The citation exists precisely so that check is fast.
It cannot answer what was never written down. If the pricing for odd jobs lives in the owner's head, no retrieval system will surface it. And when two documents contradict each other, the system's answers will be inconsistent — which is a documents problem the system exposes rather than causes, and usually the most valuable thing the first month reveals.
For medical, dental, and legal practices, one more line matters: this architecture keeps your records on infrastructure you control, which is the right starting point for a compliance conversation — not a substitute for one. Involve your own counsel or compliance officer before patient or client records go anywhere near it.
Where to start
List the ten documents people in your shop actually ask about — the warranty terms, the current price book, the two manufacturer specs that come up every week. For each, ask three questions: is it current, is it correct, and is there a clean digital copy. That inventory takes an afternoon, it is where most of the value of any retrieval project begins, and it is worth doing even if you never build the system.