The crawler will not open your manual, and the manual resists being split

Open the manual and try to select a paragraph and copy it. If you get the words, there is a text layer and the file can be read. If you get nothing, or you get a picture, the manual is scanned and uploading it will produce a source that contributes essentially nothing while appearing to have been added successfully.

Do this per section rather than once, because mixed documents are common. A manual can be perfectly readable text for two hundred pages and then contain a chapter that was pasted in from a scan, or an appendix of screenshots with all the specifications inside the pictures. That chapter is invisible and nothing will say so.

Where a section is images, it has to be converted to text before upload. That is real work with real cost, and it is worth doing only for the sections people ask about. A scanned appendix nobody references is not worth converting; a scanned specifications table that generates half your questions is.

Prepare a version of the manual for upload that is not the version you print. Take out the running headers, the footers, the page numbers, the confidentiality notice and the document reference that repeats on every page. None of them mean anything in a passage and all of them appear in every passage.

Reflow anything in two or more columns into one. This is the change that fixes the most severe corruption, because a two-column page read straight across produces sentences that are grammatically plausible and factually invented, alternating halves of two different paragraphs. That is much harder to notice than an obviously broken passage.

Deal with the floating material at the same time. Callout boxes, margin notes and sidebars land at unpredictable points in the extracted stream. Either move each one inline, immediately after the paragraph it belongs to, or pull it out into its own short section with a heading that says what it is about.

Find every table that crosses a page boundary. On the continuation page the header is gone, and what remains is rows of values with no columns to name them. Either resize the table so it fits on one page, repeat the header row at the top of each continuation, or convert the whole table into one sentence per row with the column names written into the sentence.

The prose version is the most robust, and for tables people ask specific questions about it is worth the space. A row that reads the compact model has a capacity of that many litres and weighs that much is answerable on its own; the same row inside a grid, one page after its header, is not.

Apply the same test to any table whose meaning lives outside its cells: a merged heading covering three columns, a shaded row indicating something, an asterisk pointing to a footnote at the bottom of the page. Each of these carries meaning that does not come across as text, and each has to be written into the cells or into the surrounding sentences instead.

A four-hundred page manual uploaded as one file is one source whose passages all come from an enormous document about everything. Splitting it into per-topic files, one for installation, one for maintenance, one for specifications, one for safety, gives each set of passages a document that is about a single subject and makes it far easier to see later which topics are actually being asked about.

The cost is real and worth stating plainly. It is manual work each time the manual is revised, because a new edition means splitting again rather than replacing one file. It uses more of your page allowance, which is set by plan: fifty on the free tier, five hundred on Starter, five thousand on Growth and fifty thousand on Agency. And you have to keep track of which split files correspond to which edition, or you will end up with two editions of the same chapter uploaded at once, which produces exactly the contradiction you would expect.

One thing you do not have to worry about is accidental duplication. Two files whose extracted text comes out identical are stored once, so a printer-friendly variant of a chapter you have already uploaded adds nothing rather than doubling it. That protects you from the most common accident, but not from two genuinely different editions, which differ in their text and so both count.

The distinctive failure of a PDF manual is silence where there should be an answer. The manual is definitive, it covers the question in detail, and the assistant declines because nothing in the extracted text resembles the question closely enough. The material contains page headers, page numbers, and paragraphs of interleaved column text that read as gibberish to a matcher, and the actual specification was a picture.

The louder version is a specification quoted from the wrong row. A table split across a page break yields a passage of bare values, the model is handed the numbers without the column names, and the customer is told a capacity or a torque figure that belongs to a different model. In a manual that is not an embarrassment, it is a safety problem, and it is the one case in this whole family where a wrong answer can hurt somebody.

Both failures share a cause and a cost: the upload appeared to work. Nothing is missing from the list of sources, nothing is flagged as unreadable, and the owner reasonably assumes the manual is covered. The gap is only discovered by asking the assistant a question you already know the answer to, which is the single most useful thing to do after uploading any manual.