Convert PDF to Text for an LLM Prompt

Research and casework rarely arrive as one clean file. You have a handful of PDFs, a Word document of notes, maybe a spreadsheet of figures, and you want to ask a model about all of them together. The chat box only takes text, and a PDF is not text yet.

This guide covers how to convert PDFs to text for an LLM prompt: why the file you drag in is not the text the model reads, how to turn a whole folder of documents into one blob, and where the limits are. There is a spot to try it on your own documents partway down. Nothing is uploaded to us; the extraction runs in your browser.

A PDF, a Word document, or a slide deck is a container. The words are in there, but they are wrapped in layout, fonts, and positioning that a prompt cannot read directly. Before any of it can go into ChatGPT or Claude, the text has to be pulled out of the container.

FileConcat does that extraction for PDFs, Word documents, spreadsheets, and slide decks, then folds the recovered text into one bundle alongside any plain notes or markdown you drop in. Several formats go in; one clean text file comes out.

case-notes/
motion.pdf
exhibits.docx
timeline.xlsx
memo.md

combined.txt
motion.pdf      (text)
exhibits.docx   (text)
timeline.xlsx   (text)
memo.md

You do not have to convert each file one at a time. Give ChatGPT a whole folder of documents by dropping the folder here, and FileConcat extracts every readable one and combines the multiple PDFs into one text file for ChatGPT or Claude in a single pass. If the folder also carries the usual junk, the default filter skips it before it reaches the bundle.

1. 1

   Drop your documents

   A folder or a loose handful works. PDFs, Word documents, spreadsheets, slide decks, and plain text or markdown all go in together.

2. 2

   The text is extracted in your browser

   Each document is opened and its text pulled out on your machine. Nothing is uploaded to do it.

3. 3

   Copy one bundle

   You get a single document with each file labeled by name, plus a live token count so you know it fits before you paste it.

Drop a few PDFs or a folder of mixed documents below. It runs in your browser, there is no account, and nothing is uploaded to us. This is the same engine as the full tool, so what you copy here is ready to paste when you want to feed multiple PDFs to ChatGPT at once.

Extraction works on the text a document actually stores. A PDF exported from a word processor, a Word file, a slide deck: those carry their text, and it comes out clean.

A scanned page is different. If a PDF is really a photograph of a page, there is no text inside it to pull, only an image. Reading words off an image takes recognition, and for every page that comes back with nothing FileConcat runs a recognition pass in your browser and tells you which files it recovered that way. Up to three such documents and 8 MB per drop it starts on its own; past that it asks first. Encrypted or password-locked PDFs are the other story: locked bytes yield no text, and no amount of recognition changes that.

None of that is taken on trust. We built a set of documents designed to break our own extraction and have re-tested against it four times: of the sixteen defects found so far, seven were in PDF handling, five of those are fixed, and two are still open. The two that remain are a superscript footnote marker sorting above its own line, and a page with no text layer being dropped without a note. Since 2026-10-07 such a page is listed and read by recognition; the next re-test will say how well.

Documents add up faster than they look. A few long PDFs can run past a model's context window, and if you are paying per token, a bundle you paste a dozen times is a real line item. This is where a stack of case files or a pile of research papers gets expensive without warning.

FileConcat counts tokens as you add and remove files, and it can show the cost for a given model, so you can trim to what matters first. That helps when you want to summarize multiple research papers with ChatGPT and only a few of them hold the section you actually care about.

Two things move when the bundle grows, and only one of them is the bill. Input is priced per million tokens, so the cost is exactly linear: doubling the documents doubles what the send costs, every time, and a model having room to spare is not a discount. The second is accuracy. Published sweeps find retrieval degrading as the input gets longer, which means the last stretch of a window is not free even when you can afford it. We priced the first half across the whole model catalogue in what it costs to fill a context window, and the sources for the second half are cited there.