How FTC Works

How FTC Works

FTC means Files To Cards. It is ONCard's pipeline for turning source material into study questions and then into saved cards.

You may refer to:

• What FTC accepts
• How sources are normalized
• OCR on vs OCR off
• How question generation works
• How generated questions become cards
• Limits and stress protection
• What happens on cancel or failure

What FTC accepts

FTC currently accepts:

• Images: .png, .jpg, .jpeg, .webp, .bmp, .tiff, .tif
• PDFs
• PowerPoint files: .pptx

All files in one FTC run must belong to the same source family. You cannot mix images and PDFs in the same run.

FTC works in units:

• For images, 1 file = 1 unit
• For PDFs, 1 page = 1 unit
• For PPTX, 1 slide = 1 unit

How sources are normalized

Before any AI stage starts, ONCard converts the selected files into normalized PNG pages inside the runtime folder.

• Images are loaded and resized into normalized PNGs.
• PDFs are rendered page-by-page into PNGs.
• PPTX files are not rendered as full visual slides. ONCard extracts the slide text from the PowerPoint XML, then draws that text onto a clean image. This means FTC on .pptx is mostly text-driven, not a screenshot of the original slide layout.

This normalization step gives later stages one common page format to work with.

OCR on vs OCR off

FTC has two main routes.

OCR on

When OCR is enabled:

1. Every normalized page is sent through the OCR model.
2. The OCR model returns markdown with analysis and plain text.
3. ONCard cleans the OCR output and strips prompt echoes or junk text.
4. The collected page text is merged into one study paper.
5. The paper is used to generate study questions.

This route is more text-faithful and is safer when the source is dense with readable text.

OCR off

When OCR is disabled:

1. ONCard skips page transcription.
2. The paper model reads the page images directly.
3. It writes revision material from what is visibly present.
4. If there are many pages, ONCard builds partial paper sections in batches and then merges them.
5. The merged paper is used to generate study questions.

This route is more direct and avoids OCR, but it depends more on the paper model's visual reading quality.

How question generation works

After the paper is ready, FTC moves into question generation.

• Questions are generated in batches of up to 4.
• The model is asked for strict JSON first.
• If the model returns unusable structured output, ONCard falls back to extracting questions from plain text.
• New questions are compared against already generated ones so duplicates and close repeats are skipped.
• The run keeps retrying until it reaches the requested count or it hits the retry limit.

Questions are meant to be standalone. FTC explicitly tells the model not to write things like "according to the passage" or "from the notes above".

How generated questions become cards

FTC itself does not directly write full cards. It first produces questions, then sends each question into the normal autofill flow.

That autofill step fills in:

• title
• subject
• category
• subtopic
• hints
• search terms
• answer
• difficulty
• optional MCQ answers

Each saved card also receives the FTC run id. That lets ONCard track which cards came from the same Files To Cards run.

If valid MCQ answers are returned during autofill, ONCard also builds and caches the shuffled MCQ payload for later study use.

Limits and stress protection

FTC has hard limits to stop the pipeline from overloading.

Standard mode

• Max input units: 12
• Max questions when units are 1-6: 14
• Max questions when units are 7-9: 18
• Max questions when units are 10-12: 20

Force mode

• Max input units: 24
• Max questions when units are 1-6: 19
• Max questions when units are 7-9: 21
• Max questions when units are 10-24: 29

If a PDF or PPTX is too large for standard mode, ONCard can offer force mode. Force mode increases capacity, but the app warns that it slows the process and puts more stress on the pipeline.

What happens on cancel or failure

FTC is designed to clean up after itself.

• If you stop an FTC run, ONCard deletes cards that were already saved under that run id.
• It also removes queued FTC jobs and clears the runtime files for that run.
• If question generation fails after producing some valid questions, ONCard can offer to keep the partial question set instead of throwing everything away.

So FTC behaves more like a managed pipeline than a one-shot prompt.