How to chat with a PDF and check every answer against the page

When you ask a question, a document chat tool does three things. It reads the text of every page (and runs OCR on scanned pages), it finds the passages most likely to contain the answer, and it writes an answer from those passages. The quality of the result depends on all three steps, not only on the writing.

The part that matters most for business use is the last link in the chain: can you see which passage each sentence came from? An answer that says "either party may terminate with 60 days' notice" is only useful if you can open page 14 and read the clause yourself. That is why every answer in ChatDocument carries a citation with the document name and page, and clicking it opens the PDF at the highlighted passage.

A few minutes of preparation saves most of the frustration people have with document chat.

• Use the final version. If you have a draft and a signed copy, upload the signed one, or upload both and name them clearly ("MSA signed 2026", "MSA draft v3"). The file name appears in every citation, so a clear name makes answers easier to read.
• Check scanned pages. Scanned PDFs are read with OCR. Faint stamps, handwriting in margins and tables photographed at an angle read less reliably than typed text. If a key page is a poor scan, rescan it at a higher resolution.
• Keep related files together. An amendment changes the main agreement. Put both in the same space so a question about the notice period can find the amended clause too.
• Split giant bundles. A single 900 page PDF with a contract, twelve exhibits and email printouts works, but separate files with clear names give citations that are easier to follow.

The way you ask decides how precise the answer is. Some habits that work across contracts, policies and reports:

1. 1Ask one thing at a time. "What is the notice period for termination for convenience?" gets a cleaner answer than "Tell me about termination, renewal and payment."
2. 2Use the document's own words when you know them. If the agreement says "Customer" and "Provider", ask "When can the Provider suspend the service?" rather than "When can the vendor cut us off?"
3. 3Name the scope. "In the employee handbook, how many days of parental leave are offered?" helps when the space holds several documents.
4. 4Ask for the form you need. Choose a short answer, an answer with quotes, or a bullet summary. Quotes are the best choice when the exact wording matters, such as indemnity or a liability cap.
5. 5Follow up in the same thread. After "What is the liability cap?" ask "Does it exclude data breaches?" The thread keeps the context, and each follow-up gets its own citations.

Questions worth asking of common documents

• Contracts. Who are the parties, when does the term start and end, does it renew automatically, what is the notice period, what is the liability cap, what is excluded from it, which law governs.
• Policies and handbooks. Who is eligible, what is the approval process, what are the limits, what happens on exceptions, which version is current.
• Financial reports. What drove the change in revenue, what are the stated risks, what guidance was given, which figures were restated.

Verification is the habit that separates a useful tool from a risky one. For each answer you plan to rely on:

1. 1Click the citation. The document opens at the cited page with the passage highlighted.
2. 2Read the full sentence and the one around it. A clause often has a qualifier in the next sentence: "except as set out in Section 9.3".
3. 3Follow cross references. If the passage points to another section, ask about that section next, or open it from the viewer.
4. 4Check numbers character by character. Dates, amounts and day counts are where a misread costs most.
5. 5Note the page in your own work. When you write the summary for a colleague, keep "p. 14" next to the statement. It saves the next reader the same search.

ChatDocument does part of this check for you before you see anything: each cited passage is checked in code to exist on the cited page and to contain the quoted text. A sentence whose citation fails that check is not shown. You still read the clause, but you never chase a page that does not say what the answer claims. Our article on why AI answers need citations goes deeper on this routine.

Sometimes the document simply does not cover the question. A lease may say nothing about subletting; a handbook may not mention remote work abroad. A trustworthy tool says so plainly. ChatDocument answers "This document does not contain an answer to that question" and shows the closest passages it checked, with their pages, so you can confirm the gap yourself.

That refusal is a feature, not a failure. An answer filled in from general knowledge ("most leases allow subletting with consent") can be wrong for your document and sounds just as confident as a real one. When you get a refusal, try rephrasing once with the document's own terms. If the second attempt also comes back empty, the gap is probably real, and that is useful information for a negotiation or a policy update.

Most teams start with a single file and quickly need more. A few patterns that follow naturally:

• Several versions of one contract. Use compare documents to see what changed between them instead of asking the same question twice.
• Many contracts of the same kind. Use extract to table to pull the same fields from each one into a spreadsheet with every cell linked to its page.
• A shared library. Put policies, past proposals or product documentation in a team space, so colleagues ask the same source and get the same cited answers.

A short checklist to keep next to your desk

• One clear file name per version.
• One question per message, in the document's own terms.
• Quotes when exact wording matters.
• Click every citation you plan to rely on.
• Treat "not in the document" as an answer.