Ask all of it at once.
Answered before you blink.

Corpus Desk

Where were we invoiced above the rate we signed?

Who is working on our sites with expired insurance?

What renews itself before we get the chance to cancel?

Which customers are still buying at last year's prices?

01 / INPUT

Bring the mess.

Do not prep. Do not sort. Do not clean up. Just point Sheaf at the raw intake and let the engine handle the reality of operational documents.

Scans

Phone Photos

Merged Packets

Application Dumps

Wrong-order Pages

Junk Pages

Dense Tables

Signatures

02 / PROCESS

Every file, one question.

The reading is already done, so the asking is free. Ask a question in plain words and Sheaf answers it across everything you have ever put through it — then names the pages it read to get there.

Index

Extraction is the index.

Nothing extra runs. The fields pulled when a document was filed are the same fields a question searches, so a file becomes answerable the moment it is approved.

Ask

Plain words, no query builder.

Type it the way you would say it to a colleague. There is no syntax to learn, no filter panel to assemble, and nothing to set up before the first answer comes back.

Cross

Answers no single file holds.

What is missing across four thousand files, where two documents contradict each other, what your own average actually is. None of it is written down anywhere until the files are read together.

Cite

Every figure names its pages.

An answer arrives with the documents behind it and the pages they sat on. Open one and you are looking at the scan itself, not a summary of it.

Technical detail How corpus questions work, and why an answer has to name its pages Read the spec →

03 / OUTPUT

Structured payload.

The file doesn't need to arrive perfect, but the data leaves that way. Ready for your existing infrastructure.

A question comes back as rows, not prose. The same shape every time, ready to drop straight into a spreadsheet, a report, or whatever already runs your week.

Every row carries the documents it was drawn from, so a number in a board pack can be walked back to a page in a scan without leaving the record.

{
  "question": "invoiced above signed rate",
  "files_read": 3914,
  "elapsed_ms": 61,
  "rows": [
    {
      "subject": "Meridian Freight",
      "delta_usd": 4120,
      "cites": ["RC-4471 p1", "INV-88213 p2"]
    }
  ],
  "total_usd": 16765
}

How corpus questions work

The expensive part of answering a question about a document is reading it. Sheaf has already done that. What is left is a lookup — which is why it costs nothing and finishes before you have let go of the key.

The reading happened once.
Everything after it is a lookup.

The work was done at intake.

Every document Sheaf files leaves behind three things: the fields it was read for, the pages those fields sat on, and the fields your schema asked for that the document turned out not to contain. Those become rows the moment the document is approved — not overnight, not on a schedule, and not by a separate indexing job somebody has to remember to run. By the time you have a question, the answer is already sitting in a table.

Three questions a pile of documents cannot answer.

What is missing

Search can only find what is there. Sheaf records what a document was supposed to contain and didn’t, so “which of these four thousand files has no current certificate of insurance” is a question with an answer rather than a project.

Where two documents disagree

The signed rate confirmation says one number and the invoice says another. Neither document looks wrong on its own, which is exactly why nobody catches it. Sheaf compares the fields that are meant to agree and lists the ones that don’t.

What ordinary looks like

Your own files are the benchmark. What your average actually is, which agreements drift from your standard, which month broke the pattern. None of it is written down anywhere until the files are read together.

What people put to it.

A freight brokerage

“Where were we invoiced above the rate we signed?” Six discrepancies out of 3,914 files, $16,765, in 61 milliseconds — each one pointing at the rate confirmation and the invoice that disagreed.

A marketing agency

“Which clients’ effective CPM moved more than ten percent this quarter, and was it the rate card or the delivery?” Answered across three years of platform invoices and every insertion order behind them.

A property brokerage

“Which listing agreements expire in the next forty-five days?” and “what repair credit did we concede, against our own average?” One is a calendar, the other is a coaching conversation.

An auction house

“Across four thousand lots, whose estimates beat hammer most consistently?” A question about specialists, answered out of consignment agreements and results that were never in the same place.

Any employer

“What is median tenure by hiring manager?” Offer letters and exit paperwork, read once on the way in, and never opened again to answer it.

It stays fast because it stops reading.

Answering by re-reading documents costs time and money in proportion to how many there are, which is why nobody offers it across a whole archive — at ten thousand files it is not expensive, it is impossible. Reading once and answering from the rows costs the same whether you hold four hundred files or forty thousand. The searching is instant and effectively free. The only thing left to pay for is the thinking, and there is very little of it left to do.

Every row keeps its documents and its page numbers, which is what makes a fast answer safe to act on.

Ask all of it at once.Answered before you blink.