Exam prompt
How do mass number and proton number change in alpha decay?
How certain was your answer?
CorrectUnsureWrong
PDF to flashcards, automatically, with AI
Upload a PDF, photo or lecture slide. Quanta creates document-bound cards with selectable Bloom levels, transparent source status and LaTeX formulas. Recognition quality depends on layout, images and handwriting.
European AI provider (Mistral, Paris) · document-bound · selectable Bloom 1 to 6
Start learning now
Starter is free · one AI run included
„Der größte Vor[t]eil […] ist, dass die KI jede Karte mit Hilfe seriöser Quellen erstellt und diese Quellen auch als Link anzeigt, sodass man dort selber nachlesen kann."- TU Dresden
- Humboldt-Universität zu Berlin
- HTW Dresden
- Fachhochschule Dresden
- Universität Leipzig
- HTWK Leipzig
- Freie Universität Berlin
- IU Internationale Hochschule
- FernUniversität in Hagen
What happens after the upload
No black-box promise. Every step is anchored in the code. PDF, JPG, PNG and WebP files up to 10 MB are supported.
Upload a document
A PDF, photo (JPG/PNG/WebP) or screenshot of notes or a lecture slide. Legible handwriting can be recognised.
Multimodal AI analysis
The European AI provider (Mistral, Paris) analyses text and image at the same time and can capture formulas, diagram labels and structure formulas. The file path runs via the provider's standard endpoint, for which it gives no specific processing location.
Bloom-level extraction
With no selection, the card prompt prefers understanding and application questions. When you select individual Bloom levels from 1 to 6, the cards must follow exactly those cognitive targets.
Document grounding and quote matching
The system instruction tells the model to use only information from the uploaded file. When PDF text can be extracted, the server checks verbatim evidence; a missing match does not discard a card in the visual scan path but leaves it without a verified seal.
Duplicate detection
A limited set of existing cards from your topic is passed to the model as avoidance context. This reduces identical or very close repetitions but is not a complete semantic duplicate guarantee.
FSRS integration
New cards are saved directly in Quanta's study system. After your first rating, FSRS-6 calculates stability, difficulty and the next review date.
Four verifiable quality principles
A versioned Quanta demo document shows the card, formula and evidence in one product state.
Physics · Radioactivity
Flashcard
Evidence linked directly to the card
1 / 4 · Switch subject
Active recall
Question and answer instead of copying
Karpicke & Roediger 2008 (Science 319:966) studied retrieval practice. That supports the methodological context but does not guarantee a particular outcome with Quanta.
Bloom 1 to 6
Freely selectable target levels; otherwise a focus on understanding and application
With no selection, the card scan prefers understanding and application questions. When levels are selected, exactly those become binding. Anderson & Krathwohl 2001 provide the taxonomy, not evidence of product effectiveness.
Document-bound
The file is the model basis, with quote matching where technically possible
The prompt forbids additions from free model knowledge. Verbatim matches mark grounded cards when text is extractable; unmatched visual-scan cards remain without a verified seal.
Bloom taxonomy as a controllable system rule
With no selection, the card scan prefers understanding and application questions. When you select one or more Bloom levels from 1 to 6, exactly those become binding targets in the scan prompt. This supports foundations as well as deliberate application, analysis, evaluation or creation.
Anderson, L. W. & Krathwohl, D. R. (2001). A Taxonomy for Learning, Teaching, and Assessing. Addison Wesley.
Document-bound prompt with transparent verification
Hallucinations are a risk with AI-generated study cards, especially for formulas and mechanisms. Quanta's system instruction restricts generation to the uploaded document. Server-side quote matching verifies evidence where text can be extracted reliably. In the visual file path, unmatched cards are retained because layout, formulas or OCR can diverge, but they are not marked as verified.
McGrew, S. et al. (2021). Breakdowns in AI factual accuracy. Proceedings of AIES '21.
Level context from profile or document
Complete profile details for school type, year and region or degree programme and semester can enrich the scan prompt. When they are missing, generation follows the level and terminology of the uploaded document without assuming an education stage. Pedagogical reference: the Zone of Proximal Development (Vygotsky, 1978).
Vygotsky, L. S. (1978). Mind in Society. Harvard University Press.
LaTeX-native, not bolted on
Quanta renders inline LaTeX formulas ($f(x)$) and block formulas ($$E = mc^2$$) directly via KaTeX. The scan prompt instructs the model to output mathematical expressions in LaTeX. Paivio (1971) provides the theoretical context for visual and textual representations; it does not imply an automatic learning outcome.
Paivio, A. (1971). Imagery and Verbal Processes. Holt, Rinehart & Winston.
What really sits behind the PDF scan
“The scan prompt is designed to create learnable concepts from the uploaded document instead of merely rephrasing exercises. Reading the file is handled by Mistral AI SAS, based in Paris, via the provider's standard endpoint, for which it gives no specific processing location; the basis is the standard contractual clauses. The provider contractually excludes training on the content we transmit. The 30 rolling days the provider states are its general figure for inputs and outputs (as of 27 July 2026); for the file interface it has, to our knowledge (as of 18 August 2026), not published a separate automatic deletion period, so here our own deletion carries the commitment: we delete the uploaded file copy immediately after processing, including when something fails. What gets analysed is the visual file content, including text, formulas and labels. With no Bloom selection, the card scan prefers understanding and application questions; selected levels from 1 to 6 become binding. Where PDF text can be extracted, quote matching checks the evidence. A match can fail because of layout, formulas, images or handwriting; the card then remains document-bound but does not receive a verified seal.”
Frequently asked questions about the PDF flashcard generator
How do I turn a PDF into flashcards?
Open a topic in Quanta, tap "AI Scan", upload the PDF and choose how many cards you want. The European AI provider (Mistral, Paris) analyses the file and streams the generated cards. Processing time depends on document size, file count and current load.
Which document types work best?
PDFs with a clear text structure (textbooks, scripts), photos of notes with legible handwriting and screenshots of presentation slides. PDF, JPG, PNG and WebP files up to 10 MB are supported; multilingual documents are possible.
Can I scan the same text more than once?
Yes. Quanta gives the model a limited set of existing cards as avoidance context. Another scan can therefore focus on different concepts, but full coverage and semantic duplicate freedom are not guaranteed.
What happens if the document has fewer concepts than the cards I asked for?
The scan can return fewer cards than requested and then shows a partial result. The prompt remains restricted to the file. With extractable PDF text, only verbatim-matched cards receive a verified seal; visual cards without a match are stored transparently without that seal.
How does the PDF scan differ from AI Set?
The PDF scan uses an uploaded file as the model basis and instructs the European AI provider (Mistral, Paris) to generate from the document. AI Set instead starts from a topic name and searches for suitable full-text sources by default; if no suitable source is found, no flashcards are created.
Where is my uploaded PDF processed?
AI processing runs through Mistral AI SAS, based in Paris: text functions go via the provider's EU endpoint, while reading files and voice recordings goes via the provider's standard endpoint, for which it gives no specific processing location; the basis is the standard contractual clauses. The provider contractually excludes training on the content we transmit. The PDF scan belongs to the file path, which is why we write standard endpoint here instead of a blanket "processed in the EU". On deletion: the 30 rolling days the provider states are its general figure for inputs and outputs (its privacy policy, as of 27 July 2026), not a retention period for your PDF. For the file interface the provider has, to our knowledge (as of 18 August 2026), not published a separate automatic deletion period. That is exactly why our own deletion is the commitment that applies to your file: we delete the uploaded file copy immediately after processing, including when something fails; in addition, a regular cleanup run removes leftovers from aborted operations. Accounts, study content and results are stored in region europe-west3, that is Frankfurt am Main.
Is the PDF-to-flashcards generator free?
Free forever: 1 topic, no more than 100 cards in total, FSRS, and exactly one AI generation run over the account lifetime. The regular generator offers 40–100 cards from text, a link, or up to 10 files; the guided onboarding entry offers 10–50 from text or files. The onboarding run counts as that one lifetime run; Starter has no monthly AI quota. Further generations require Quanta Essential, which includes 300 AI cards per month shared across generation modes. Eligible users can try the plan for 7 days.
AM
Amos Matzke·Founder & Managing Director, Full-Stack Architect · former MINT-EC student·April 2026
Your first flashcards from a PDF
Upload a document. Let the cards stream in. After your first rating, FSRS schedules the next review.