Long documents, and what gets lost in the middle

What "I have read your document" actually means

You attach a 200-page contract and ask a question. Something happens quickly and an answer comes back. It is worth knowing what happened, because the shape of the failure follows from it.

Two things can be going on. If the document fits inside the model's context window — the amount of text it can hold at once, now very large in the leading tools — the whole thing is in front of it. If it does not fit, or the tool is built for documents, the file has been chopped into passages, those passages have been indexed, and your question retrieves the handful that look most relevant. Only those passages reach the model. The rest of the document is not in the room.

Either way, one finding should govern how you work. Researchers testing long inputs in 2023 found a consistent U-shape: information placed at the beginning or the end of a long context is used reliably; the same information placed in the middle is used much less reliably, sometimes worse than if it had not been provided at all. The effect has softened in newer models. It has not gone away.

The consequence you must internalise

A negative answer over a long document is the weakest answer you can get.

"Does this manual mention X?" — "No" means no passage that looked relevant to your phrasing was retrieved, and nothing salient appeared at the ends. It does not mean the manual is silent. When absence matters — a missing indemnity, an unstated deadline, an obligation nobody has noticed — you cannot outsource the finding of it to a single question.

Positive answers are much safer, because you can go and look at the thing it found.

How to work over long material

Ask located questions. Not "what does this contract say about termination", but "in clauses 14 to 19, what does the contract say about termination, quoting each relevant sentence". Naming a location changes what is retrieved and gives you an address to check.

Go section by section. Thirty questions over ten sections beats one question over three hundred pages, and it takes less time than it sounds because each answer is short and checkable.

Demand quotations, always. "Quote the exact sentence and give the clause or page number. If the document does not address this, say 'not addressed' and nothing else." An answer with no quote is not a weak answer — treat it as no answer.

Use the tool to navigate, not to conclude. The highest-value question over a long document is usually where, not what. "Which sections discuss data retention?" gives you four numbers, and you read four sections instead of three hundred pages. Wrong answers are visible in seconds because you open the section and it is about something else.

Ask the same question two ways. "What are the notice requirements?" and "what must a party do before terminating?" retrieve different passages. When both point to the same clause, your confidence is earned rather than assumed.

Notebook-style tools

A family of tools now exists specifically for this: you assemble a set of sources, and the assistant answers only from them, with inline citations you can click back to the source passage. Google's NotebookLM is the best known and has a free tier; several document platforms now ship the same shape.

What they get right is the discipline this whole module is about — grounding answers in your material and showing you where each claim came from. That is a genuinely better default than a chat window, and for a body of material you return to repeatedly, it is worth the setup.

What they do not fix: retrieval still misses things, so a negative answer is still weak; and a citation proves a passage exists, not that the passage says what the summary claims. Click it. Read the sentence. Most of the errors that survive in citation-carrying tools are the ones where somebody trusted the presence of a footnote.

Two practical constraints

Scanned documents are images. If the PDF was produced by a scanner, the text layer may be poor or absent, and a page of numbers read by OCR is a specific hazard — a smudged 3 becomes an 8 silently. Check that you can select the text. Free OCR that works: Tesseract, and the OCR built into most PDF readers.

Large files also cost time and, on paid tiers, money. The whole document goes in every time you ask. Splitting a 300-page manual into its chapters is faster, cheaper and more accurate than uploading it whole, and takes two minutes.

The habit

For anything that matters, the sequence is always the same: narrow, quote, open, read. Narrow to a section. Demand a quote. Open the source. Read the sentence yourself.

That is four steps and about ninety seconds, and it converts a tool that is sometimes right into a tool that reliably saves you an afternoon.