Chat with PDF: how to ask questions of long reports, filings and papers (and check the answers)

This article is part of our guide to AI citation generators that point to real sources. It explains how PDF chat works in plain terms, where it fails, how to check answers, and how the main tools compare on limits, price and privacy.

Chat with PDF is a tool that lets you ask questions in plain language about a PDF and get answers drawn from its text.

How chat with PDF works: retrieval plus a language model

Almost every PDF chat tool uses a method called retrieval-augmented generation. Retrieval-augmented generation (RAG) is a method in which a system first retrieves passages related to a question from a set of documents, then gives those passages to a language model to write the answer. The idea was set out in a 2020 research paper by Patrick Lewis and colleagues.

In plain terms, the tool works like a research assistant who has an index of your document but only reads the pages the index points to:

1. Extract. The tool pulls the text out of the PDF. Scanned pages need optical character recognition (OCR) first; without it there is no text to search.
2. Split. It cuts the text into passages, often a few hundred words each, and records the page each came from.
3. Index. It turns each passage into a numerical representation of its meaning, so it can find passages on the same topic even when they use different words.
4. Retrieve. When you ask a question, it finds the handful of passages closest to your question.
5. Generate. It sends those passages and your question to a language model, which writes the answer.
6. Cite. Good tools link each part of the answer back to the passage and page it came from.

Some tools with large context windows skip retrieval for short files and give the model the whole document. That helps, but research has found that language models use information at the start and end of a long input better than information in the middle, so long documents still get uneven attention.

Where it fails

The rule that follows: treat a PDF chat answer as a pointer to a page, not as a fact.

What journalists, analysts and writers use it for

Step by step: upload, ask, check the cited page

1. Check the file. Open the PDF and try to select a sentence. If you can't, it is a scan without a text layer: run OCR (Acrobat and many free tools do this) before uploading.
2. Upload it to a tool whose privacy terms fit the document (see below).
3. Start with an orientation question. "What is this document, who wrote it, when, and how is it organized?" This shows whether the tool read it correctly.
4. Ask narrow questions. One fact per question, and ask for the page: "What was the 2025 operating loss? Give the page and quote the line."
5. Open every cited page. Read the sentence the answer relies on, plus the paragraph around it. Confirm the number, the year and who is speaking.
6. Ask what's missing. "What does the document say that contradicts this?" or "Which sections did you not use?"
7. Record what you checked. Note the claim, the page, and the exact quote in your draft or notes, so an editor can check it too.

Asking across many PDFs at once

Multi-file chat is where these tools save the most time and where errors hide best. When a tool searches twenty files, the answer may blend passages from different documents, years or parties.

• Ask for the file name and page for every claim. "For each point, give the document name and page."
• Group files by question. Put the five budget years in one folder or notebook and the court filings in another. Smaller sets retrieve better.
• Name files clearly before upload: 2025-budget-enacted.pdf, not doc(3).pdf. Tools often cite by file name.
• Compare explicitly. "Compare the transit allocation in the 2024 and 2025 budgets, citing the page in each."
• Use keyword search for completeness. Chat finds the most related passages, not all of them.

For reviewing a body of studies, see AI literature review tools, which cover search and screening as well as PDF chat.

Privacy and confidentiality: where your files go

Before uploading anything sensitive (unpublished filings, leaked documents, client reports, unproduced scripts), check:

• Training. Are uploads or chats used to train models?
• Retention. How long are files kept, and are they deleted when you delete the chat?
• Human review. Can staff read uploads, for example when you send feedback?
• Sub-processors. Is the file sent to a third-party model provider?
• Location and account type. Business and education accounts often have stricter terms than personal ones.

What two of the large vendors say, as of November 2026:

• Adobe says customer content in Acrobat's AI features "is never used to train Adobe's AI models."
• Google says that for personal accounts, Gemini Notebook data is not used for training unless you provide feedback, and that feedback may be reviewed with its full context. For Workspace accounts, uploads are not reviewed by humans and not used to train models.

For documents that must not leave your organization, the options are an enterprise plan with a data agreement, or a tool you can run in your own cloud account. For a journalist protecting a source, a document's metadata can identify the source even if the content is safe to share; strip it before uploading anywhere.

Chat with PDF tools compared

Prices and limits as of November 2026. They change often; check each pricing page before you choose.

Fair notes: Gemini Notebook is the strongest free option for large sets of sources. Acrobat is the best choice if you already pay for it and need OCR, redaction and editing in the same tool. SciSpace is built around research papers. ChatPDF is the fastest way to ask a question with no sign-up.

Good questions vs. bad questions

Add one line to any prompt: "If the document doesn't say, answer 'not in the document'."

Chatting with PDFs in WriteOrbit

WriteOrbit puts PDF chat next to the document you are writing, so the answer, the page and your draft are on one screen.

1. Upload PDFs to the Library, or import them from Zotero or Mendeley on a paid plan.
2. Open a PDF and choose Ask. Answers carry page numbers, so you can turn to the passage and check it.
3. Ask across your Library from the AI chat beside your draft by choosing Library as a source. Its answers cite the papers they draw on; open the paper and use Ask on it when you need the exact page.
4. Cite as you write. Type @ to cite a paper from the Library or from a search of 250M+ scholarly works.
5. Check the draft. AI review runs claim checks and flags retracted papers, preprints and wrong DOIs, as comments in the text.
6. Keep sensitive files in your own cloud. WriteOrbit can be deployed to your own Firebase / Google Cloud project, so documents stay in your own cloud account.

The free plan includes 10 PDFs and five chat messages a day. Check /pricing for current limits.

FAQ

How does chat with PDF work?

Most tools use retrieval-augmented generation. The tool extracts the PDF's text, splits it into passages, and indexes them. When you ask a question, it finds the passages most related to it and gives only those to a language model, which writes an answer from them. Good tools show which passage and page each claim came from.

Is there a free way to chat with a PDF?

Yes. ChatPDF, AskYourPDF, SciSpace, Gemini Notebook (formerly NotebookLM) and WriteOrbit all have free tiers. Free plans cap the number of documents per day, pages or file size, and questions. Gemini Notebook's free tier is the most generous for many files: up to 50 sources per notebook. Check each tool's pricing page, because limits change often.

Can I chat with multiple PDFs at once?

Yes, in most tools. ChatPDF uses folders, AskYourPDF has multi-document chat on paid plans, Adobe Acrobat uses PDF Spaces, and Gemini Notebook treats each notebook as a set of sources. Cross-document answers are harder to check, so ask the tool to name the file and page for every claim.

Is it safe to upload confidential PDFs to an AI chat tool?

Only if the tool's terms fit the document. Check whether uploads are used for training, how long they are kept, whether staff can review them, and where they are stored. Adobe says customer content is never used to train its AI models, and Google says personal Gemini Notebook uploads aren't used for training unless you send feedback. Court-sealed or source-protected material may need a self-hosted tool.

Why does a PDF chat tool give wrong answers?

The usual causes are retrieval misses, where the right passage was never shown to the model; scanned pages without readable text; tables and multi-column layouts that extract badly; and questions that need the whole document, like "list every mention." The model may then answer from general knowledge. Always open the cited page before you use an answer.

Where to go next

Page-cited answers are one half of trustworthy sourcing; real citations are the other. Read our guide to AI citation generators that point to real sources. If you are a reporter, see AI tools for journalists for the full newsroom workflow, from documents to publication.

Sources

1. Lewis, P. et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020): https://arxiv.org/abs/2005.11401
2. Liu, N. F. et al., "Lost in the Middle: How Language Models Use Long Contexts" (2023): https://arxiv.org/abs/2307.03172
3. ChatPDF: https://www.chatpdf.com/
4. AskYourPDF, pricing: https://askyourpdf.com/pricing
5. Adobe, Acrobat AI Assistant and plans: https://www.adobe.com/acrobat/generative-ai-pdf.html
6. Google, Gemini Notebook help, usage limits by plan: https://support.google.com/notebooklm/answer/16213268
7. Google, Gemini Notebook help, supported sources and per-source limits: https://support.google.com/notebooklm/answer/16215270
8. Google, Gemini Notebook help, privacy: https://support.google.com/notebooklm/answer/16164461
9. Google, "NotebookLM is now Gemini Notebook" (16 July 2026): https://blog.google/innovation-and-ai/products/gemini-notebook/notebooklm-gemini-notebook/