QuestGen.

This is a simplified conceptual diagram of the components documented in the repository; it is not a claim that all operations use a single synchronous pipeline.

1. 01

   Ingest and index the course

   Create a course, upload supported material and extract ordered content. PDF pages and presentation slides can receive optional visual transcription. Persist chunks and searchable metadata in SQLite; report coverage and warnings when visual analysis is incomplete.

   backend/ingestion.py
2. 02

   Retrieve under the chosen source scope

   Use course-scoped SQLite FTS5/BM25 search, optionally fused with local Ollama embeddings. When files are explicitly selected, pass the allowed document IDs into retrieval and exam context preparation instead of searching the whole library.

   backend/documents.py
3. 03

   Consult in Agent Mode

   Discuss educational goals and question mix rather than applying a fixed MCQ checklist. A streaming tool-calling agent can inspect course material and perform bounded web research, then save a versioned question plan that requires an explicit UI approval.

   backend/agent_mode.py
4. 04

   Draft and revise questions

   After approved consultation, store complete question drafts with answers and explanations. The professor can edit, revise, select and carry chosen questions into Exam Mode without starting a second agent drafting pipeline.

   backend/agent_drafts.py
5. 05

   Plan and draft a full paper in Exam Mode

   Validate section counts, marks and source scope. Build slot-specific evidence and infer difficulty/Bloom demand when unspecified; optional Deep Agents planning remains advisory. Generate mixed-format questions in batches, splitting invalid or oversized batches for retries.

   backend/exam_batch.py
6. 06

   Review, finalize and export

   Keep or revise selected drafts, maintain question versions and ordering, then explicitly approve the complete paper. Student and instructor exports are available as PDF or DOCX.

   backend/exam_builder.py

Source grounding, consultation and approval

found = await asyncio.to_thread(
    documents.retrieve, session['course_id'], section['topic'], 12, material_ids)
if not found:
    found = await asyncio.to_thread(
        documents.sample_course_sources, session['course_id'], 12, material_ids)
if not found:
    raise ValueError(f'No indexed evidence for section {section["id"]}. Reprocess the course files.')
if any(s.get('document_id') and s['document_id'] not in material_ids for s in found):
    raise ValueError('Retrieval returned an unapproved document')

Implementation evidence, not an accuracy benchmark

The uploaded code contains FastAPI and frontend tests for consultation approval, Agent-to-Exam handoff, source restrictions, ingestion resilience, asynchronous batch generation, question editing/versioning, and paper export. These automated tests check behavior and constraints, not independent learning outcomes or model-generated question correctness. The source archive does not establish a measured hallucination rate, teacher time saved, production uptime or human-rated assessment quality; no such metric is claimed here.