Why most AI flashcard generators produce cards you can't trust
A general chatbot asked to "make flashcards from this PDF" will happily produce fifty cards. Some of them come from the PDF. Some come from what the model already believed about the topic, and a few are simply wrong. You will not notice which is which — that is the entire problem — because you are studying the cards precisely because you don't yet know the material.
Getting a fact wrong in an essay costs you marks. Getting a fact wrong in a flashcard costs you far more, because spaced repetition is a memorisation machine: it will patiently drill the error into you over weeks, and re-teach it every time you start to forget it. A wrong flashcard is worse than no flashcard.
Thaxas is built around that risk. Cards are generated only from the chunks of your upload that the retrieval step actually returned, every card is stamped with its source page, and when your material doesn't cover something, Thaxas says so instead of filling the gap from memory. You can read more about why in why AI study tools hallucinate.
How it works
- 1
Upload your material
PDFs, PowerPoint, Word, plain text and Markdown, photos of handwritten notes (JPG, PNG, WebP, HEIC), and audio lectures — up to 50MB a file. A whole module's readings can go into one space.
- 2
Thaxas reads and indexes it
The file is split into passages of roughly 500 tokens with the page numbers kept attached, then embedded so the generator can retrieve the right passage for each card rather than skimming the whole document.
- 3
Generate a deck
Ask for cards on the whole space or on one chapter. Each card is written from retrieved passages and labelled with the page it came from.
- 4
Review on the FSRS schedule
Rate a card Again, Hard, Good or Easy and FSRS sets the next interval from your actual memory state. Cards you find easy disappear for months; cards you keep failing come back tomorrow.
Real spaced repetition, not a two-day reminder
Most "AI flashcard" apps show you the deck in a loop and call it spaced repetition. Thaxas runs FSRS-4.5 — the same free-spaced-repetition scheduler the Anki community adopted — with the published seventeen weights and a power-law forgetting curve. Your review history produces a memory state for each card, and the next interval is computed to catch the card just as your recall probability decays to the target. The implementation is tested against the reference values, because a scheduler that is quietly wrong is indistinguishable from one that is right until your exam.
If you want the theory rather than the marketing, we wrote it up: spaced repetition, explained properly and FSRS vs SM-2.
What you can turn into flashcards
- Lecture slides — the deck the lecturer actually taught from, not a generic summary of the topic.
- Textbook chapters and readings — including scanned PDFs; the text is extracted from the image.
- Photos of handwritten notes — your own handwriting, straight from your phone.
- Recorded lectures — audio is transcribed first, then carded, with timestamps kept as the reference.
- A quiz you just took — convert the questions you got wrong straight into a deck, which is the single highest-yield thing on this list.