PDF AI Chat does not read your whole document for every question. It cuts the text into short excerpts, picks up to eight that share words with your question, and sends only those to the model. Once you know that, you know when to trust an answer and when to open the page.
This guide walks through what happens between your question and the answer in PDFCheck's AI chat, with numbers from a three-page sample I processed locally. Most of it applies to any retrieval-based PDF chat tool.
Pages become numbered excerpts
After upload, the text of each page is extracted and cut into excerpts of up to 220 words. Consecutive excerpts on the same page overlap by 40 words, so a sentence near a boundary lands in both. Every excerpt gets an ID made of its page and position, such as P3-C1, and those IDs are what answers cite.
A three-page agreement of 627 words produced four excerpts:
| Excerpt ID | Words | Covers |
|---|---|---|
P1-C1 | 220 | Page 1, words 1 to 220 |
P1-C2 | 100 | Page 1, words 181 to 280, including the 40-word overlap |
P2-C1 | 191 | All of page 2 |
P3-C1 | 156 | All of page 3 |
If no page yields any text at all, as with a fully scanned document, OCR is attempted instead: the first 8 pages are rendered with Ghostscript at 160 dpi and sent to OpenAI for transcription. OCR runs only when the whole document has no text. A scanned page inside an otherwise digital PDF is simply not read.
Which excerpts travel with your question
Excerpts are chosen by word matching, not by meaning. Words of three or more letters in your question are compared with each excerpt, after a short list of filler words such as "the", "what" and "does" is dropped. Each occurrence of a word scores 4 points and its presence adds 1. Up to eight excerpts are sent: any you selected yourself go first, and the best-scoring ones fill the remaining places.
When fewer than two excerpts match anything, the top of the ranking is used anyway. For a question with no matching words, that means the start of the document. Here is how three prompts scored on the sample:
| Prompt | Matching excerpt and score | What happened |
|---|---|---|
| When does the warranty expire? | P3-C1, 9 | "warranty" appears twice; "expire" scores nothing because the page says the warranty "lasts" 12 months |
| How late can payment be made? | P2-C1, 18 | Page 2 contains "payment" three times and "late" once |
| Summarize the document | None, all 0 | No question word occurs in the text, so excerpts are taken in document order |
The warranty row is the one I keep coming back to. The right page was found, but only because "warranty" happened to be on it. Had I asked "how long is the guarantee?", nothing would have matched and the model would have seen page 1 instead. So use the document's own words. Section titles, defined terms and the exact noun on the page beat a paraphrase every time.
What a page citation proves
The model is told to answer only from the supplied excerpts and to cite them by ID. Each citation is checked against the excerpts that were actually sent; references to anything else are dropped, and the page number you see comes from the excerpt record, not from the model. A citation chip therefore reliably points to a real passage on that page.
The quoted wording next to the chip is different. The model writes that quote, up to 240 characters, and it is not compared with the excerpt. So the routine is short but not optional:
- Click the chip and read the source passage, not only the quote.
- Read the paragraph around it. Exceptions tend to live in the next sentence.
- Compare names, figures and dates character by character.
- If the answer says something is "not mentioned", search the PDF yourself before accepting that. The model can only report that the excerpts it received lack the information.
To tighten the next question, select a passage in the extracted text and add it as context. Selected excerpts are always sent first.
Answers that need you to open the page
- Summaries of long documents. With no matching words, eight excerpts of at most 220 words from the opening pages are all the model sees. Ask section by section instead, or split the PDF so each chat covers one part.
- Totals, counts and comparisons across many pages. They rarely fit in eight excerpts.
- Tables and multi-column layouts. Extraction flattens them into one run of words, so rows and columns can get mixed. The text extractor shows you the same flattened text the chat works from.
- Scanned pages beyond the first eight, or scanned pages in a document that also has digital text.
- "Is this accurate?" questions. The instructions limit accuracy checks to consistency within the excerpts. Nothing is checked against outside sources.
- Legal, medical and financial conclusions. Use the chat to find the clause, then read it and ask a qualified person.
For an exact word or number, plain search is still the better tool. Acrobat's Find and Advanced Search read every page, and so does Ctrl+F in Chrome's viewer. Chat is for questions in your own words that pull from more than one passage. For anything important, use both.
What the chat cannot tell you about the file
The chat reads text, not file properties. It cannot say how the PDF was made, whether an invoice is genuine, or whether a signature can be trusted. The metadata check on the home page shows creation dates and software names, which are clues rather than proof, and the signature checker lists signers and whether bytes were added after signing without judging whether the signer is trustworthy.
Plan, provider and what happens to your PDF
AI Chat is included in the AI plan only: 1,000 messages per calendar month and uploads up to 100 MB. The Free plan and the current Pro plan do not include it; older paid subscriptions keep the chat access they were sold with. A message counts against the allowance only when an answer comes back, so a provider failure costs nothing. Prices are on the pricing page.
Each question is sent with the selected excerpts and up to the last six completed exchanges. The configuration uses OpenAI by default, with gpt-4o-mini as the default model setting and DeepSeek as the fallback when the first provider fails. Those providers' own retention terms apply to what they receive, and OpenAI's own PDF guide describes the same pattern of extracted text and page images reaching the model. The privacy policy says the same in section 2.3. Check your organization's rules before uploading anything confidential.
The original PDF, the extracted excerpts, your messages, the answers and any notes stay with your account until you delete the chat or the account. Deleting a chat cannot recall content already sent to a provider. The Download Markdown button exports the extracted text page by page under its excerpt IDs, plus notes and comments, which is the honest picture of what the chat could see. If you only need clean text for another AI tool, read PDF to Markdown for AI workflows.
Messages you may see
- "This PDF is still being prepared for chat." Extraction or OCR has not finished.
- "This PDF did not contain extractable text, and OCR could not read it yet." Neither text extraction nor OCR produced words.
- "The AI provider could not answer right now. Please try again." Both providers failed; the attempt does not use a message.
- "You have used your AI message allowance for this month." It resets with the next calendar month.
- "AI Chat requires the AI plan." The account has no AI Chat access.
The whole method, chunk, match, cite, is described in OpenAI's retrieval guide as well. The practical lesson is the same everywhere: a retrieval chat is a very fast way to find the page. Reading the page is still your job.