Skip to content

Tools

Chat with PDF

By Ofrai Editorial (AI Model Curators)

Ofrai Text Studio accepts PDF and DOCX uploads. Ask questions against the extracted text with any live chat model — this is document Q&A, not a separate ChatPDF subscription.

Open studio

Chat and video use credits from your Ofrai balance.

How to use chat with PDF

Open Text Studio, attach a PDF or Word file, pick Claude for careful reading or DeepSeek and GPT for faster passes, then ask your question in the same thread where the file lives. The attachment stays available across the conversation, so follow-up questions do not require re-uploading. Stay within the upload limits shown in the studio, and upload one document at a time when you are comparing sources — two files in one thread invites blended answers that cite neither.

A good first question is descriptive rather than evaluative: "summarise the scope and the parties named on page one" tells you immediately whether the extraction worked, before you spend ten questions on a document the model never actually read.

What it is not

It is not a full legal-discovery platform, and it is not OCR for scanned image-only pages when extraction fails. It does not browse the web, so it cannot fill gaps in the document with outside knowledge unless you deliberately paste that context. It does not remember previous files once you start a new thread. And it does not guarantee that a number in the answer matches the number on the page — extraction can drop a superscript, a decimal comma, or a footnote marker without telling you.

None of that makes it useless; it makes it a reading and question tool rather than a system of record. The right mental model is a very fast research assistant who has read your file and is working from memory while you read along.

When to switch to AI coding

After the document question is answered, return to AI coding mode in the same Chat studio and paste the summary into your IDE or ticket. The thread and credit balance stay put.

What kinds of PDF work best

Native text PDFs give the cleanest answers because the extractor can read the characters directly. That means anything exported from Word, InDesign, Google Docs or a modern publishing tool. Layered files, where the visible text sits over a scanned image, usually extract fine. Problems start with PDFs built from screenshots of slides, exported from old CAD or plotting tools, or re-saved by a scanner app that flattened every page to a picture.

Password-protected files fail at upload rather than at question time, which is the good kind of failure. If you are unsure which category you have, open the file, select a paragraph of body text, and see whether the selection is real text or a rectangular highlight over an image.

Formats that upload cleanly to chat with PDF

PDF is the reliable default, DOCX works well because the text layer is unambiguous, and plain text or markdown is the most accurate of all. Spreadsheet exports are readable but lose the meaning of merged cells and column headers, so a question about totals needs the header row restated in your prompt. Presentations convert with one slide per page and usually survive, but chart labels often come out as separate fragments.

Email threads saved as PDF carry quoted history that inflates the page count and confuses the model about who said what. For the awkward formats, the pragmatic move is to export the relevant section rather than the whole file: a twenty-page extract with a clear heading beats a two-hundred-page upload where the answer is buried on page 160.

Answers that cite the right page

Ask for citations explicitly or you will get confident prose. "Answer in three sentences and quote the sentence you relied on, with its page number" forces the model to show its work, and now you can check the source in two seconds instead of re-reading the whole document. Ask for a quote before a paraphrase on anything contractual, numeric or regulatory, because a paraphrase hides small wording changes while a quote does not.

When the model cannot find the answer, insist it says so: "if this is not in the document, reply exactly: not in the document" turns a hallucinated clause into a fast, obvious failure. Two habits make the difference between a usable document pass and a dangerous one: cite everything you intend to act on, and verify the two or three citations that matter most before you forward the summary to anyone.

Scanned versus native files

A scanned PDF has no characters to extract, only pictures of characters, so the model is effectively answering from nothing and will produce fluent, plausible, wrong text. The tell is easy: if a summary of page one contains none of the exact numbers, names or dates you can see on that page, extraction failed. The fix is preprocessing — run the file through OCR first, or re-export it from the source application while choosing the text-layer option.

If OCR is not available to you, upload only the pages you need as images in a vision-capable image model rather than uploading the whole scan and hoping. And never mix: a partly scanned document will answer well for the native pages and invent for the rest, which is worse than knowing the file is unusable because the failure looks selective and therefore trustworthy.

Which model reads which document

Claude Sonnet is the careful reader: long contracts, dense reports, anything where a missed clause has a cost, and anything with tables and footnotes to track. GPT-5 is the fast generalist — better for a first pass over a familiar document type, for comparing two drafts of the same thing, and for reasoning across sections rather than retrieving one fact.

DeepSeek V4 Pro is the economical middle for bulk reading, like forty partner contracts hunting for one clause, where the value is coverage rather than the quality of any single answer. Practical routing: read with the cheap model to find the relevant pages, then re-ask those pages with the premium model. That pattern costs a fraction of a full premium pass and gives better answers, because the premium model no longer has to guess which pages matter.

Prompt patterns for document questions

The shape of the question decides the shape of the answer. For extraction, be literal: "list every date, amount and currency in the payment terms section, as a table with the page number." For comparison, give it both sides: "compare sections 4 and 7 on termination notice and tell me only where they disagree." For compliance checks, state the rule rather than the topic: "we are a Delaware company with 50 employees; does this contract permit assignment without consent, and quote the clause."

For a summary with a purpose, say what it is for: "summarise the technical limitations section for a non-technical buyer." What you should not write is "tell me about this contract", which reliably produces a generic paragraph about contracts. When you need several angles, ask for them as numbered questions in one message.

Where document answers go wrong

Five failure modes account for most bad document answers. Silently skipped pages: the extraction dropped a table or an appendix and the model never knew it was missing, so it answers confidently from what survived. Merged numbers: a column of figures gets flattened into a run-on sentence and one digit changes. Attribution drift across a long thread: by question ten the model is answering from its own earlier summary rather than the file, and the summary is where the first mistake entered.

Definitional drift: the model uses a term differently from the document's definitions section, which matters most in contracts. And confident absence: asked about something outside the file, it produces an answer anyway. Every one of these is cheap to catch if you require page citations and re-ask the important question in a fresh thread with the same file attached.

Long documents and context limits

Long does not mean the whole file is read equally well. The middle of a very long document is where answers get vaguest, because the model's attention is spread across everything that came before. Work around it rather than fighting it: split a 300-page document into logical chunks, ask the same question of each, and compare the answers, because the chunk with a notably different answer is usually the interesting one.

For a long read, run a cheap model first to produce a one-page map of sections and page numbers, then work only inside the range you care about. State what the document is at the top of the thread — "this is a 2024 master services agreement between us and a German customer" — because that context alone removes a surprising number of misreadings, and re-state it if you start a new thread.

What a chat with PDF document pass costs

Document work is billed like any other text job, by the tokens of the extracted text that go into the request, and a big file is expensive to carry through every turn of a long thread. The practical control is scoping: upload the twenty pages that matter instead of the two-hundred-page manual, and you have usually removed most of the spend without losing the answer.

The second control is model routing, since a cheap model reading forty documents to find one clause costs a fraction of the premium pass that finally interprets it. The third is not repeating yourself: once a question is answered, record the answer in your own notes rather than asking a follow-up reformulation that returns the same text at full price. On Ofrai that same wallet pays for images and video, so a document afternoon competes directly with the campaign work you scheduled after it.

Who should preprocess the file instead

Some documents are not worth fighting with in chat. Regulatory filings where every number carries legal weight, contracts where a paraphrase could change meaning, anything under privilege or a confidentiality obligation, and anything that will be quoted externally should be read by a tool built for that job, or by a person who can be held responsible.

The same goes for very large batches: if you need to read hundreds of documents, a script with an extraction API is faster and cheaper than a human-in-the-loop conversation, and it produces a consistent record. Chat with PDF is strongest exactly where a person would otherwise read the whole thing once and then talk about it — one contract, one report, one long manual, a question you need answered before lunch.

Anything where an unchecked paraphrase could cost money or credibility belongs outside the chat window, not inside it.

Open studio

FAQ

Which models can see the file?
Listed text models in Text Studio. Switch models in the same thread when models allows.
Is my file used to train Ofrai?
Use the product for generation; see the privacy policy for how uploads are handled.
How much does a generation cost?
Each job lists Ofrai credits in the studio. Catalog credits are starting points; resolution and duration can change the final spend.
How do I start on Ofrai?
Open Image or Video Studio from this page, pick a recommended model, paste or adapt the starter prompt, and generate from your library.
Can I use chat-with-PDF output commercially?
You can use the answers the way you use any model output, subject to each model's usage rules. Read the model page and your distribution terms before commercial release.

Related models

Compare