PDFs and OCR
You can attach a PDF to the AI generator and make flashcards from it. Textbooks, lecture slides, papers and scanned handouts all work, and you choose which chapters are actually used rather than feeding in the whole book.
Attaching a PDF
Section titled “Attaching a PDF”In the AI generator, use the PDF button and pick either From vault or From computer. A PDF from your computer is read in memory — it is not copied into your vault.
PDFs are limited to 10 MB each. If yours is larger, split it or export the chapters you need.
You can attach more than one PDF. Each gets its own tab, and cards can come from all of them in a single run.
Choosing chapters
Section titled “Choosing chapters”Once a PDF is attached, the Chapters panel lists its outline with the page range each entry covers. Tick the chapters you want.
This is the part worth spending a moment on: only the pages in your selection are read and sent. Selecting one chapter of a 400-page textbook costs you that chapter, not the book. If a PDF has no embedded outline you will see No chapters detected — in that case the whole document is used.
A parent chapter covers all of its sub-sections, so ticking it includes everything underneath.
Text extraction vs OCR
Section titled “Text extraction vs OCR”There are two ways a page becomes text, and which one you get depends on the AI provider you have configured.
| Provider | How pages are read |
|---|---|
| Decks Pro | Every selected page is transcribed by an OCR model |
| Your own API key (OpenAI, Claude, Gemini, local) | Text is extracted directly from the PDF |
Direct text extraction is free and instant, but it only returns what the PDF actually stores as text. A scanned chapter, a photographed page, or a slide exported as an image contains no text layer, so extraction returns nothing useful from it.
OCR reads the rendered page image instead, which is why it works on scans. It also preserves structure — headings, lists and tables — and writes maths as LaTeX, so formulas survive into your cards rather than arriving as mangled characters.
While pages are being read you will see Reading pages 3/12….
What happens to the pages
Section titled “What happens to the pages”Pages you send for OCR are transcribed and discarded — we do not store them. Every request is sent with a “do not train” flag set and routed only to providers whose stated policy is not to train on or retain what they receive; we enforce that on our side, and their handling is governed by their own published policies. The transcription that comes back is written only to your own vault, in the cache folder below. See what happens to your content for the full picture.
The OCR cache
Section titled “The OCR cache”Transcribed pages are cached to disk, so a PDF is never read twice.
The cache lives in .obsidian/plugins/decks/pdf-ocr/ by default; you can move it under Settings → AI → PDF OCR cache folder. Pages already cached are marked cached and cost nothing on a repeat run.
The cache is readable Markdown
Section titled “The cache is readable Markdown”Each page is stored as its own Markdown file, laid out as <pdf>/<model>/<page>.md. These are not raw text dumps — they are clean Markdown with the headings, lists and tables the page had, and maths written as $…$ and $$…$$ so formulas render properly in Obsidian.
That makes the by-product genuinely useful: a scanned textbook chapter comes out the other side as readable notes.
There is one thing to know before you rely on it. The default location is inside .obsidian/, and Obsidian does not index its config folder — files there will not appear in search, in the graph, or in [[ autocomplete.
If you want to use the transcriptions as notes, point PDF OCR cache folder at an ordinary folder in your vault, for example PDF transcripts. Every transcribed page then becomes a normal note you can link to with [[PDF transcripts/<pdf>/<model>/12]], embed with ![[…]], search, and edit — while still serving as the cache, so nothing is transcribed twice.
Each entry is keyed by the PDF’s content, the page number, and the OCR model. Two consequences worth knowing:
- Re-running the same chapters, or coming back to a PDF next week, is free.
- Switching quality tier re-transcribes, because it is a different model.
Renaming or moving the PDF does not invalidate anything — the key is the file’s content, not its path. To clear the cache, delete the folder.
Getting good cards from a PDF
Section titled “Getting good cards from a PDF”- Select narrowly. A chapter at a time produces better cards than a whole book, and the model has less to lose track of.
- Say what you want in the prompt. “Focus on the definitions and the enzyme names” steers the output far more than the source alone.
- Use the chapter labels. Cards record which section they came from, so you can tell at a glance whether coverage is even.
- Generate again to go deeper. The first run covers a source broadly; running it again with the cards already produced asks for what was missed. When the model has nothing substantive left it says so.
Limits
Section titled “Limits”- 10 MB per PDF.
- Encrypted or password-protected PDFs cannot be opened.
- A PDF with no outline is treated as a single unit — there are no chapters to pick from.
- OCR requires Decks Pro. With your own API key, scanned PDFs will produce little or nothing, because there is no text layer to extract.