Chat with PDF: how to ask the document instead of re-reading everything
You have a 137-page PDF, a tender notice, a contract, or study material. You need one specific piece of information: the deadline, the exception, the item that changes everything. You hit Ctrl+F, type a word, and get 42 hits of the same expression scattered across the document. None of them answers your question, because your question wasn’t a word — it was an idea.
Chat with PDF exists for that gap. Instead of hunting for the exact term you think the author used, you ask in your own words and the answer is built from the document’s text. The difference looks small in theory and is enormous in practice, especially when the file is long and you need it more than once.
What Ctrl+F doesn’t solve
Searching for an exact word works when you already know what you’re looking for. It fails in three very common situations:
- The answer is spread out. The deadline is on page 12, the condition that suspends it is on page 48, and the exception is on page 96. None of those pages answers on its own.
- You remember the concept, not the term. The document says “termination without cause,” you search for “cancellation.” Zero results.
- The question is comparative. “Which of the two methodologies requires a smaller sample?” doesn’t exist as a string anywhere in the file.
Re-reading the whole document works, but it costs the same time every time. And when the PDF has 100 pages of dense text, you end up reading out of obligation instead of consulting out of need.
What a chat with PDF is (and what it isn’t)
A chat with PDF is a conversation whose answers are anchored in that file’s content. It isn’t a generic chatbot answering from memory with the PDF as decoration. The difference shows up in practice: one answers “appeal deadlines are usually 10 days”; the other answers “the deadline in this document is 5 business days, in section 4.2.”
That difference depends less on the language model and more on how the PDF was read. That’s what a study on PDF structure recognition applied to RAG shows: it compared a simple PDF extractor with a parser that preserves tables, headings and section hierarchy — the better parser answered better on 47% of the questions, tied on 38%, and lost on only 15%. In other words, when the extracted text loses tables, headers and structure, the answer gets worse — no matter how good the model is.
That’s why a reliable chat with PDF starts before the chat: it starts at extraction.
How to do it in Sintesy, from file to answer
1. Upload the PDF
In the dashboard, use New → Upload file and select the PDF. The limit per file is 64 MB — above that the system itself suggests splitting the document into parts. Scanned PDFs (the ones that are an image, with no text layer) go through OCR, so it’s worth sending in that scan you got by email and could never copy from.
2. Wait for the Synthesis to be ready
Sintesy doesn’t hand back just a run of text. It delivers the document’s Synthesis: a summary, the full text extracted in the transcription tab, and a mind map with the main sections. This matters because you get two reading speeds: the summary to know what the document is about, and the full text to track down the detail.
3. Open the chat and ask with context
In the chat panel, the field reads “Ask about your Sintesy…”. That’s where the conversation happens. Three habits change the result:
- Name what you want. “List the supplier’s obligations with deadline and penalty” works better than “summarize the obligations.”
- Ask for the output format you’re going to use. A table, a list with owner and date, or two paragraphs.
- Accept “I couldn’t find it.” If the answer isn’t in the document, that’s what you want to hear. A made-up answer in an official document is worse than no answer.
4. Check the cited passage
Every answer comes with the sources it used inside the document. Reading the passage takes five seconds and settles the doubt: you check whether the interpretation matches the original text before you take the information into a decision.
Questions that work (and the ones that stall)
| Weak question | Why it stalls | Version that works |
|---|---|---|
| ”Summarize the document” | You get the same summary as the Synthesis, not an answer | ”What are the 5 obligations with a defined deadline, with the date of each one?" |
| "Does it talk about deadlines?” | Yes/no isn’t a useful answer | ”What is the deadline for each step, and what happens if it’s missed?" |
| "Explain chapter 4” | It asks for reading, not consulting | ”In chapter 4, what changes compared to what chapter 2 set out?" |
| "Is this right?” | The document has no way to validate your thesis | ”What does the document say, literally, about this point? Cite the section.” |
A sequence that works well on long documents: start with the map (sections), then ask about the specific section, and only then ask for a comparison between two parts.
Cross several documents in the same question
This is where the chat stops being a luxury and becomes a method. You can add more than one source as context in the same question — two PDFs, the PDF and the meeting where it was discussed, three lectures from the same course. With the sources attached, the question crosses the files:
- “Does the signed contract keep the clause we discussed in Tuesday’s meeting?”
- “What does the notice require, and where does the study material cover it?”
- “Do the three authors agree on point 2? Point out the disagreement.”
Without that, you’d ask the question three times, in three tabs, and still have to piece the answers together in your head.
The assistant also acts, not just answers
Sintesy’s chat isn’t just a field for questions about one document. It operates on your library: it searches across your syntheses and notes (“where did we talk about delivery deadlines last quarter?”), creates a note from a YouTube link or a web page, creates and organizes folders, builds checklist items from what was agreed, and can reprocess a note when the audio or the file came out bad.
Destructive actions (deleting a note, folder or checklist) don’t happen on their own: the assistant returns a confirmation code and only runs it after your “confirm.” In work documents, that guard rail is worth more than speed.
Who gets the most out of it
- Anyone studying for an exam or certification. A 200-page PDF becomes a summary, a mind map, and a place where asking “what’s tested on tax jurisdiction?” is faster than flipping through.
- Anyone reading contracts, tender notices and regulations. In those documents the answer exists, is written down, and has to be cited. Asking is faster than digging and less risky than trusting memory.
- Anyone turning raw material into reference material. Meeting minutes, technical reports, scientific papers: the value isn’t in the file, it’s in being able to come back to it.
And there’s a side effect worth noting: asking retains more than re-reading. In the classic Roediger and Karpicke experiment on learning through testing, a week after studying the same text, those who only re-read remembered about 42% of the ideas; those who tested themselves once, 56%; those who tested themselves more than once, 61%. Five minutes later, re-reading looked better — which is exactly why it fools you. Asking the document is a way to test yourself without waiting for the exam.
Limits and precautions
- Size: the free plan accepts files up to 20 MB; on Pro the limit per file goes to 2 GB, with PDFs capped at 64 MB. On the free plan, daily usage is 1 hour of audio and up to 10 syntheses per day.
- OCR isn’t perfect: in a scanned document, double-check proper names, amounts and dates before citing.
- Complex tables deserve a visual check: merged cells and numbers in wide columns are the weakest point of any PDF extraction.
- The critical decision is still yours: chat with PDF speeds up consultation, it doesn’t replace reading the whole thing when the subject is legal, medical or financial.
The shortest path
Re-reading a 100-page PDF costs an hour and you forget half of it. Asking costs two minutes and leaves a trail: the answer, the passage and the page. If you already have the file in hand, the work is uploading it once and changing the habit afterward. Anyone who wants to start with a simple task can find the step-by-step in how to study a PDF with AI without re-reading everything from scratch and, for the PDF that needs to become a visual structure, it’s worth seeing turning a PDF into a mind map. And if the content you consult comes from video or a webinar, the flow is the same: recorded webinar into text and a summary.
