It reads the files you already have,
and answers only from them

There is no language model inside it. It finds the passage in your own documents that best matches the question, quotes up to three lines of it word for word, and tells you which document and which lines. If nothing in your records scores well enough, it declines. Everything else on this site follows from that.

Four things are true of the product, and every page on this site agrees with all four. If a page contradicts one, the page is wrong.

1 · It answers only from the records it was given

The answer is a verbatim line out of a document you already had. Nothing is composed. The engine's answer path is extractive: at most three lines, at most two citations, and every citation carries a document id and a line locator. If you have never seen the sentence before, it did not come from this product.

2 · It abstains when the records do not carry the answer

When the best passage it can find does not score high enough, or does not cover enough of the words in your question, it returns one fixed sentence: “Your records don't contain enough to answer that with confidence.” That refusal is the product. The citation on every other answer is the proof that the refusal means something.

3 · It runs on your own machine

It listens on the machine it is installed on and nowhere else. The store is a SQLite file on local disk. The model that reads your text runs on the device. It will not start a component that would send your files to somebody else — that is refused at startup, not warned about.

4 · It does not train on anything

Retrieval and configuration, never weights. Your records are indexed so they can be searched; they are never fed into a model, ours or anyone's. That is a property of how it is built, and it is also what keeps one person's corpus separable from every other.

Answering gets the attention. Conversion has to happen before anything can be answered, and it is where most of the engineering has gone.

Half one · the filing cabinet, made readable

A drawer of paper is not searchable. Conversion turns scans, spreadsheets, .docx and .pdf files into plain readable Markdown, each output carrying a note of where it came from and how it was read. Every file resolves to a status, and nothing is silently skipped. In the proving run, 56,687 files went in and 53,776 readable documents came out, 98% clean, with 2 set aside for a person to look at (counted, from the engineering record). A person who never asks a single question still gets a readable archive out of this half.

Trust is the product. The conversion layer earns it by never losing a file and never being stopped by one.

Half two · an answer you can check

Ask a plain question. The engine looks it up three ways at once — by exact words, by weighted words, and by meaning — fuses the three, and takes the best passage. It quotes up to three lines of that passage and names the document and the lines. Open the file; the words are there. Arithmetic — a total, a rate times a length — is never done by a model. Ordinary code does it, and the estimator reproduces the founder's own historic job anchor of 376 m at $834.95 a metre exactly.

It is for anyone whose answers are already written down somewhere in their own files: a contractor with ten years of tickets and job costs, a farm with a shoebox of receipts, an engineering practice with its calculations, a household with a decade of scans.

It is not for you if:

• You want something that writes. It quotes. It will not draft a letter, summarise a report or explain a contract. It shows you the three lines that best answer your question, and stops.
• You want an answer to every question. It is built to decline, and on a thin corpus it will decline often. A refusal is not a bug to report; it is the product working.
• Your records are mostly photographs of paper. Scanned and photographed documents are not searchable yet. In the proving corpus, one company's scan queue stands at 0 of 883 candidates read (counted). Digital files are what it answers from today.
• You need the software to act — to move money, file a return, send an email or certify a figure. It proposes and shows where it got it. A person accepts or holds. Nothing enters the record without you.
• You need a stated accuracy. There is not one. The last honest measurement is on the record, and it was not good enough to advertise.