Extraction modes

Seven strategies. The mode is picked per job from the document's page, along with the model that will run it.

Mode key	For	How it works
standard	A document with the answer marked inline after each question	One model pass per 50-page chunk, pulling question, options, answer and explanation together.
questions_only	No answers available anywhere	Extracts stems and options. The answer is left for a person.
ai_answer	Questions without a key, where the model can work the answer out	Two passes: extract the questions, then a second pass determines the correct answer for each.
two_step	A separate answer-key section at the back	Extract the questions, extract the key, then match the two.
regex	A non-standard but consistent answer format	The model analyses the format and writes a regex; the regex is then applied to the whole document with no further model calls. Cheap and fast once it is right.
ai_decide	An unknown format	The model samples the document, picks one of the above, and runs it. Falls back to standard if it cannot decide.
generate	A textbook chapter or study notes — no questions in the source at all	The model writes MCQs from the text. The correct answer comes from the text; the distractors come from the model.

The README lists six

README.md omits ai_answer and merges regex with it under one "AI + Regex" row. The code has seven mode keys.

Choosing between them

• If the PDF says "Correct Answer: C" or "Preferred Response: C" under each question, use standard.
• If the answers are in a block at the end, use two_step.
• If the answers are marked in some house style the model keeps getting slightly wrong, use regex — it pays for one analysis pass and then never calls the model again, so it is the cheapest option for a long document in a consistent format.
• If you do not know, use ai_decide on a small page range first and look at what it chose.
• generate is a different activity from the other six. It is not extraction; it is authoring, and the drafts want reading as such.

Figures

The extraction schema includes a boolean has_figure per question, and the prompt says to set it true only when a figure, radiograph, ECG or chart is essential to answering.

A draft gets a page image attached only if that flag came back true. This is what stops every question on an illustrated page acquiring the illustration.

Where a job goes

POST /drafts/extract → Celery task extract_quiz → a DraftBatch of DraftQuestion rows. Never a live Question, never a live Quiz.

Progress is pushed to a Redis list per job and polled from /jobs.

Configuring the models

Settings → AI Models. Models are registered per task:

Task	Used by
extraction	These modes, and the offline article batch writer
article	POST /articles/ai-draft and ai-refine
flashcard	Generating a deck from an article
teach	The tutor drawer
keyword	Difficulty classification
tool	Reading an image when the primary model cannot see
tts, stt	Voice

get_model_for_task(db, task) falls back to settings.LITELLM_MODEL; get_configured_model(db, task) returns nothing instead, for work that must not run on a model nobody chose.