Skip to main content

​
Overview

AI scanning is the core workflow of Flashcard AI. Point the desktop app at a folder of images or a video file, and Gemini extracts every multiple-choice question as a structured flashcard — no manual typing required. Images are batched into PDFs (up to 50 pages per batch) and sent to Gemini with a structured extraction prompt. Gemini returns one JSON object per page containing the question, answer options, correct answers, question type, and whether it inferred the answer.

​
Supported input types

TypeExtensions
Images.jpg, .jpeg, .png, .webp, .bmp
Video files.mp4, .avi
Direct PDFSent as-is in batch mode
Video files are processed by extracting frames first. Each distinct frame is treated as an image page before being batched.

​
How scanning works

1

Select a folder

Choose a folder containing your exam images (or a video file) from the desktop app. The app reads all supported image files in order.
2

Images are merged into PDFs

Images are grouped into batches of up to 50 pages each. Each batch is merged into a single in-memory PDF and sent to Gemini in one API call, which is faster and cheaper than sending images one by one.
3

Gemini extracts questions

Gemini processes every page and returns a JSON array — one object per page. The extraction prompt instructs Gemini to capture the question stem, all answer options, the correct answer(s), question type, and whether the answer was inferred.
4

Flashcards are created

Each valid JSON object becomes a Flashcard. Pages with no question (NOT_A_QUESTION) are silently skipped. Cards with inferred answers get an automatic warning note.
5

Deck is saved

All extracted cards are saved to decks.json as a new deck. The deck is immediately available for study and quiz mode.

​
Gemini JSON output format

Each page Gemini processes returns one object with this structure: 
{
  "question": "Which layer of the OSI model is responsible for end-to-end communication?",
  "options": ["A. Network", "B. Transport", "C. Session", "D. Data Link"],
  "correct_answers": ["B"],
  "type": "single_choice",
  "inferred": false
}
For multi-answer questions: 
{
  "question": "Which of the following are valid HTTP methods?",
  "options": ["A. GET", "B. SEND", "C. POST", "D. FETCH", "E. DELETE"],
  "correct_answers": ["A", "C", "E"],
  "type": "multiple_choice",
  "inferred": false
}

​
Model fallback chain

If a model is unavailable (HTTP 404) the app automatically falls back to the next model in the list:
  1. gemini-2.5-flash (default, recommended)
  2. gemini-2.5-flash-lite
  3. gemini-3-flash-preview
  4. gemini-3.1-flash-lite-preview
  5. gemini-flash-latest
  6. gemini-flash-lite-latest
gemini-2.5-flash is the recommended model as of 2026. It offers the best balance of speed and accuracy for exam image extraction.

​
Inferred answers

When no explicit answer clue is visible in an image (no highlight, checkmark, filled bubble, or solution section), Gemini reasons using its domain knowledge and marks the card with "inferred": true. These cards receive an automatic note:
⚠ Đáp án do AI suy luận (không có đáp án rõ trong ảnh)
Inferred answers may be incorrect. Review cards with the warning note carefully and correct them before using the deck in a graded context.

​
Skipped pages

Gemini automatically skips pages that do not contain a question:
  • Blank or mostly blank pages
  • Title pages and course headers
  • Logo or watermark-only pages
  • Diagrams or charts with no question stem
  • Answer explanation pages without a question
These pages return "question": "NOT_A_QUESTION" and are not added to the deck.

​
Multi-key parallel mode

Adding multiple Gemini API keys enables parallel processing. The image list is divided into equal-sized packs — one pack per key — and each pack runs on a dedicated worker thread simultaneously. 
120 images + 3 API keys
→ Pack 1 (40 images) → Key 1 [thread 1]
→ Pack 2 (40 images) → Key 2 [thread 2]
→ Pack 3 (40 images) → Key 3 [thread 3]
This reduces total scan time proportionally to the number of keys. Each key rotates through models independently and handles its own rate-limit backoff.
Keys are validated in parallel before the scan starts. Dead or invalid keys are excluded automatically, and the scan begins from a key that has had recovery time after validation.

​
Real-time log output

The scanning UI streams live progress messages: 
🚀 PDF Batch mode: 60 images → 2 batch(es) of up to 50 pages each
── Batch 1/2: images 1–50 ──
🔧 Merging 50 images into PDF...
✔ PDF ready (2048KB)
📤 Sending PDF batch 1/2 (50 pages, 2048KB) | Key 1 [...abcd1234] | Model: gemini-2.5-flash
⏳ Waiting for response on batch 1/2...
✅ Batch 1/2 done — 48/50 cards extracted
⏱ Waiting 7.5s before next batch (rate limit buffer)...

​
Scanning tips

Gemini reads the pixel data directly. Blurry, low-resolution, or heavily compressed images reduce extraction accuracy. Aim for at least 150 DPI for scanned documents.
The batch-PDF mode assumes one question per page. If an image contains multiple questions, only the first will be reliably extracted.
The default safe rate is 8 requests per minute per key. With one key this means roughly 7.5 seconds between batches. Add more keys to speed up scanning proportionally.
Gemini preserves code syntax, math formulas, Greek letters, and special symbols exactly as they appear. Indentation, operators (==, !=, >=), and arrows (→, ≥) are retained in the output.

​
Next steps

Study mode

Review your extracted deck with keyboard shortcuts and track mastery.

Quiz mode

Practice with multiple-choice questions and save your progress.

Build docs developers (and LLMs) love

Get started for freeTalk to us