Chat with a PDF

Ask for the wording, and the page it sits on

Put the file in and ask the question you would put to a colleague who had just finished reading it. "What does it say about ending the agreement early?" "Does this paper report an effect size anywhere?" "Which page covers the payment schedule?"

Then ask for the sentence rather than a description of the sentence. A paraphrase of a clause is something you have to take on faith. A quoted line with a page number against it is something you can go and look at, and looking takes about as long as reading the paraphrase did. Text comes out of the file one page at a time, so the page number is real: turn to it and the words are either there or they are not. Make that check on the first answer of any document that matters, and you will know quickly how much rope to give the rest.

When your PDF is a photograph of a document

Some PDFs hold text. Others hold a picture of text, which is what comes out of a scanner, a photocopier with a send button, or a phone held over a page. Everything about what happens next depends on which one you have, and the test takes two seconds: open the file and try to drag your cursor across a word. If the word highlights, the text is in there. If you get a rectangle over the whole page instead, you are holding an image.

Plenty of scanners do run recognition as they scan, the setting sitting in the menu under OCR, and tuck the words invisibly behind the picture, which is why some scans do highlight after all. When yours does not, the pages get read as images instead, and the words come back. What you should not do is trust the digits. Recognition is strong on clean body text and weakest exactly where the stakes are: a figure in a small-print table, a handwritten amendment, a stamp over a date, anything that was faxed at some point in its life. Ask which page a number came from and go and look at that number.

Getting a table off a page

A table in a spreadsheet is data. A table in a PDF is lines and numbers painted at positions on a page, with nothing anywhere recording that this figure belongs in that column. Pulling one out means rebuilding the rows from where things sit.

Tables with ruled lines around the cells come out cleanly, because the lines say where the columns are. Tables held together by nothing but whitespace are the hard case, and the tell is a column that arrives one row out of step with its neighbours. Two things help. Say which page it is on, and ask for it back as a spreadsheet file rather than as text in the reply, because a file opens next to the original and a wrong cell is then visible rather than plausible.

The same goes for turning a PDF into a Word file. What comes back is the text rebuilt as a new document, not your original with a different extension. Headings, paragraphs and tables survive. The exact layout does not.

Where a long document starts to strain

There is no page count at which the file gets turned away, and nothing to compress before you upload. What changes with length is how the reading happens. The text comes out of the file and gets read in windows of about two thousand lines, which on a normally set contract is something like forty pages, rather than all in one go. Point a question at a section and it goes to that section. Ask for everything at once across three hundred pages and it works through the file in passes, which is slower and blurrier than asking about section seven.

Layout is the other thing that arrives with length. A two-column academic paper comes out in the order the text was placed on the page rather than the order a person reads it, so a footnote or a caption can land in the middle of a sentence. That is a nuisance in a whole-document summary and a non-issue when you asked about one clause, which is one more reason to ask narrow and follow up.