Paste or Attach

Tessa has the supplier contract on her desktop as a PDF, and Waymark's refund policy sitting halfway down an email thread. The chat box offers her two doors. She can select the text and paste it into her message, or she can attach the file and let the product handle it. Both doors open into the same room — tokens in the window — but the corridor between them is different, and walking it once explains most of what goes wrong when people start feeding documents to a model.

Picture two ways of getting a letter to someone who is listening on the phone. You can read it aloud yourself, choosing what to include and hearing every word as it goes. Or you can post it to a clerk who will read it aloud on your behalf. The listener's experience is identical either way — spoken words, nothing else. The difference is that in the second case the clerk's reading has quietly become part of your message. Attaching a file hires the clerk.

Paste, or attach the fileyour selection, or the whole document

→

Extraction, attach onlylayouts interleave · scans come out empty

→

Tokens in the windowcharacters only, no pages

→

Spot-check with a quotequote the first sentence of page 12

Whichever Door, It Becomes Tokens

Start with the part that never changes. The model does not receive files. It receives a stream of tokens (Chapter 1), and that is the only kind of thing it can receive. There is no PDF inside the window, no spreadsheet, no folder — only what the product turned into tokens before the model was involved. Whatever route your material takes, it arrives stripped down to what that step produced.

That one fact does a lot of explaining. It is why a carefully laid-out proposal loses its columns on the way in. It is why what the model can say about a chart depends entirely on what survived the trip. And it is why the two doors are worth comparing at all: they differ only in who turns your document into text — you, with a selection and a paste, or the product, with a step of its own.

What Attaching Actually Does

When you attach a file, the product first runs it through extraction: ordinary software that pulls the readable text out of the file's format and hands that text on to be tokenized. On ordinary documents it does the job well. A clean text-based PDF, a word-processor file, a simple spreadsheet — all of them come out roughly as you would expect.

Three kinds of document give extraction trouble. Complicated layouts, with columns and sidebars and footnotes, can come out interleaved, so sentences from opposite ends of the page arrive spliced together. Tables can lose their alignment and turn into a smear of numbers with no visible columns. And a scanned page is, to a computer, a photograph of a document rather than a document: unless something converts the picture back into characters first, there is nothing to extract, and that page arrives empty. Products differ here, increasingly in your favour — some run the picture through character recognition, some hand the model an image of the page alongside the extracted text — but nothing on the screen tells you which one you got.

The word to hold on to is quietly. Extraction happens before the model sees anything, it usually raises no error, and the model then works cheerfully with whatever it was handed. A page that arrived blank is not a page the model knows is missing. It will answer from the rest, in the same confident voice it uses for everything else — which is Chapter 3's lesson arriving from a new direction.

Choosing a Door

Paste when you want control. A pasted passage is exactly the text you selected, in the order you selected it, and you watched it go in. For a policy paragraph, a guest email, a section of a contract — anything you can reasonably select — pasting is the honest option, and it has the side benefit of forcing you to decide what is actually relevant.

Attach when the material is long, or when there are several files, or when selecting the text by hand would take twenty minutes. Then spend ten seconds on a spot-check: ask the model to quote a sentence you already know is in there. "Quote the first sentence of the deposits section." If the quote comes back matching your copy, that part of the document made it into the window. If the answer is a vague paraphrase, or a sentence that is not in your document at all, something between the file and the model went wrong — and you found out before building an afternoon's work on it.

That habit is verification (Chapter 3) pointed in a new direction. Everything so far has been about checking what comes out. This checks what went in, and it is cheaper, because you already know the right answer.

Attaching Does Not Make the Window Bigger

One last trap, and it is the one that leads into the rest of the chapter. An upload that succeeds tells you the file reached the product. It tells you nothing about whether the file's text fits in the context window, and a 120-page contract does not. Products differ in what they do about the overflow — some take only the first part, some refuse, some do something cleverer — and they do not always say which.

So a document can be attached, accepted, and still only half present. The spot-check catches it: quote something from page 12, then quote something from page 110. When the first works and the second does not, you are not looking at a mysterious failure. You are looking at a document that is bigger than the window, which is the subject of the next page and the central collision of this chapter.