Feat/h5p review

docs/feature-ideas.md

@@ -0,0 +1,348 @@
+# TLEF-CREATE Feature Ideas & Product Roadmap
+
+> Brainstorm date: 2026-02-26
+> Status: Ideation — prioritize and spec out individually before implementation
+
+---
+
+## Current Product Snapshot
+
+| Capability | Status |
+|---|---|
+| 8 question types (MC, TF, flashcard, matching, ordering, cloze, summary, discussion) | Done |
+| AI generation from materials (RAG + LLM) | Done |
+| H5P export + real H5P preview | Done |
+| PDF export (questions / answers / combined) | Done |
+| Chat Mode (AI-guided conversational workflow) | Spec'd |
+| LMS integration | None |
+| Quiz import | None |
+| Collaboration | None |
+| Analytics | None |
+| Question bank | None |
+
+---
+
+## Tier 1 — High Impact, Core Differentiators
+
+### 1.1 Quiz Import & Convert
+
+**Pain point**: Educators have years of existing quizzes in Word, PDF, Google Forms, Canvas, Quizlet. Asking them to recreate from scratch is a non-starter.
+
+**Feature**:
+- Drag in a Word/PDF exam → AI extracts question structure (type, stem, options, answer)
+- Structured extraction via LLM (question text → JSON schema) → batch `Question.create()`
+- Support formats: `.docx`, `.pdf`, `.txt`, plain paste
+- Future: Quizlet CSV import, Canvas QTI XML import
+
+**Why it matters**: Turns the adoption pitch from "create new content" to "bring your existing content and make it better." Drastically lowers the barrier to first value.
+
+**Technical path**: Reuse material upload pipeline + new LLM prompt strategy for structured extraction. Output maps to existing Question model.
+
+**Effort estimate**: Medium (new prompt engineering + parsing layer, reuse existing upload/question infra)
+
+---
+
+### 1.2 AI Quality Dashboard
+
+**Pain point**: Educators generate 20 questions but can't tell if they're pedagogically sound — Bloom's coverage? Difficulty balance? Ambiguous wording?
+
+**Feature**:
+```
+Quiz Quality Report
+├── Bloom's Taxonomy Coverage (bar chart)
+│   Remember ████████░░ (3)
+│   Understand ██████████ (4)
+│   Apply ██████░░░░ (2)
+│   Analyze ████░░░░░░ (1)
+│   Evaluate ░░░░░░░░░░ (0) ⚠️
+│   Create ░░░░░░░░░░ (0) ⚠️
+├── Difficulty Distribution (easy/medium/hard pie chart)
+├── Potential Issues
+│   ⚠️ Q3: "All of the above" — weak distractor
+│   ⚠️ Q7: Negative phrasing — may confuse
+│   ⚠️ Q12: Only 2 distractors — add more
+│   ✅ No duplicate concepts detected
+├── Content Coverage Map (material sections vs questions)
+└── [Fix Issues Automatically] button
+```
+
+**Why it matters**: Elevates the tool from "question generator" to "teaching design consultant." No competitor does this well. Makes instructors feel the tool understands pedagogy, not just content.
+
+**Technical path**: Bloom's keyword matching (client-side) + LLM deep review for ambiguity/distractor quality. New tab in Review & Edit or standalone page.
+
+**Effort estimate**: Medium (Bloom's analysis is lightweight; LLM review prompt + UI)
+
+---
+
+### 1.3 Quiz Variants — Anti-Cheat Multi-Version
+
+**Pain point**: 3 sections of the same course need 3 different exams. Manually creating variants is tedious and error-prone.
+
+**Feature**:
+- **Variant A**: Same questions, shuffled option order
+- **Variant B**: AI generates equivalent questions (same LO, different wording/scenario/numbers)
+- **Variant C**: Same LO, different question type (MC → TF → Cloze)
+- Export as separate H5P packages or one randomized Question Set
+- Side-by-side comparison view of variants
+
+**Why it matters**: Solves a universal pain point that AI is uniquely good at. "30 seconds to generate 3 versions of a midterm" is a genuine wow moment.
+
+**Technical path**: Variant A = shuffle (trivial). Variant B/C = new LLM prompt: "Given this question, generate an equivalent that tests the same concept with different context." Reuse existing `convertQuestionToH5P`.
+
+**Effort estimate**: Low–Medium (prompt engineering + new "Variants" UI panel)
+
+---
+
+## Tier 2 — Strong Value Add
+
+### 2.1 Cross-Quiz Question Bank
+
+**Pain point**: After 3 semesters, an instructor has generated 200+ questions across 6 quizzes. Finding and reusing good ones is impossible.
+
+**Feature**:
+- All generated questions auto-indexed into a personal Question Bank
+- Semantic search by topic, concept, LO text
+- Filter by type, difficulty, course, date, usage count
+- "Pull from Bank" when creating a new quiz — search and add existing questions
+- Usage tracking: how many times each question has been used, in which quizzes
+
+**Why it matters**: Transforms the product from a "one-shot tool" into an "appreciating asset." The more you use it, the more valuable your library becomes. This is the strongest retention mechanism possible.
+
+**Technical path**: Already have Qdrant for RAG. Generate embeddings for each Question, store in a new `question-bank` collection. Frontend: new Question Bank page + "Add to Quiz" flow.
+
+**Effort estimate**: Medium (embedding pipeline + new search UI + "add to quiz" integration)
+
+---
+
+### 2.2 LMS Direct Publish (Canvas / Moodle)
+
+**Pain point**: Export H5P → download → open Canvas → navigate to assignment → upload → configure → repeat for every edit. 5 minutes of tedium per quiz update.
+
+**Feature**:
+- "Publish to Canvas" button alongside Export
+- OAuth2 flow to connect Canvas account (one-time setup)
+- Auto-create Canvas Assignment + upload H5P
+- "Re-publish" after edits — one-click update
+- Future: Moodle support via similar REST API
+
+**Why it matters**: Eliminates the last-mile friction that makes instructors dread updating quizzes. Turns a 5-minute manual process into 1 second.
+
+**Technical path**: Canvas REST API supports file upload + assignment creation. New `canvasIntegrationService.js` backend + OAuth flow + "Publish" button in export section.
+
+**Effort estimate**: High (OAuth integration, Canvas API, error handling, token management)
+
+---
+
+### 2.3 Multimedia Source Support
+
+**Pain point**: A lot of teaching happens via lecture recordings and YouTube videos, not just PDFs and docs.
+
+**Feature**:
+- Paste a YouTube URL → auto-transcribe via Whisper API
+- Upload lecture recording (mp3/mp4) → transcribe → use as material
+- Timestamps preserved: each generated question links back to the specific moment in the video
+- "Watch the relevant clip" link on each question for student study
+
+**Why it matters**: Unlocks an entirely new content source that text-only tools can't touch. Particularly valuable for flipped classrooms.
+
+**Technical path**: Whisper API / yt-dlp for transcription. Transcripts feed into existing RAG pipeline. Store timestamp metadata alongside chunks.
+
+**Effort estimate**: Medium (transcription service + timestamp tracking)
+
+---
+
+### 2.4 Study Guide Auto-Generation
+
+**Pain point**: Students want study materials, not just quizzes. Instructors don't have time to create both.
+
+**Feature**:
+- "Generate Study Guide" button alongside Export
+- AI creates a structured study document from the same materials:
+  - Key concepts per learning objective
+  - Summary paragraphs
+  - "Try this" practice prompts
+  - Links back to source material sections
+- Export as PDF or Markdown
+- Complements the quiz: "Here's what to study, and here's how to test yourself"
+
+**Why it matters**: Doubles the value of uploaded materials. Instructors get two outputs (quiz + study guide) from one input (materials). Students love it.
+
+**Technical path**: New LLM prompt chain using existing RAG context. New export format alongside H5P/PDF.
+
+**Effort estimate**: Low–Medium (mostly prompt engineering + PDF template)
+
+---
+
+## Tier 3 — Innovative / Exploratory
+
+### 3.1 Smart Distractor Generation
+
+**Feature**: When generating MC questions, AI specifically crafts distractors based on **common student misconceptions** rather than random wrong answers.
+
+- Analyzes the topic to identify typical mistakes students make
+- Generates distractors that test real understanding vs surface recall
+- Tags each distractor with the misconception it targets
+- "This student chose B — they likely confused X with Y"
+
+**Why it matters**: The quality of distractors is what separates a good MC question from a trivial one. This makes auto-generated questions rival hand-crafted ones.
+
+---
+
+### 3.2 Content Coverage Map
+
+**Feature**: Visual heatmap showing which parts of the uploaded material are covered by questions and which are gaps.
+
+```
+Chapter 1: Introduction ████████████ (8 questions)
+Chapter 2: Core Concepts █████████░░ (6 questions)
+Chapter 3: Applications  ███░░░░░░░░ (2 questions) ⚠️ Under-covered
+Chapter 4: Case Studies  ░░░░░░░░░░░ (0 questions) ⚠️ No coverage
+```
+
+- Click on a gap → "Generate 3 questions for Chapter 4"
+- Ensures comprehensive assessment
+
+**Why it matters**: Addresses a real concern: "Did I cover everything?" Currently requires manual cross-referencing.
+
+---
+
+### 3.3 Collaborative Quiz Authoring
+
+**Feature**: Invite TAs or co-instructors to collaboratively build a quiz.
+
+- Share a quiz with edit permissions via email/link
+- Real-time presence indicators (who's editing what)
+- Comment/suggest mode: TA suggests a question change, instructor approves
+- Activity log: who added/edited/deleted what
+
+**Why it matters**: Large courses have 5-10 TAs who all contribute to assessments. Currently no tool bridges the gap between "solo authoring" and "LMS quiz editor."
+
+**Effort**: High (auth model changes, real-time sync, permission system)
+
+---
+
+### 3.4 Spaced Repetition Export (Anki)
+
+**Feature**: Export flashcards and Q&A pairs as Anki decks (.apkg format).
+
+- Auto-convert flashcard and MC questions into Anki cards
+- Include difficulty tags for Anki's algorithm
+- Students import into Anki for long-term retention study
+
+**Why it matters**: Anki has millions of active users. Being the bridge between "instructor creates content" and "student studies with spaced repetition" is a unique value chain.
+
+---
+
+### 3.5 Multi-Language Quiz Generation
+
+**Feature**: Generate quizzes in multiple languages from the same source material.
+
+- "Translate this quiz to French" → AI translates all questions, options, feedback
+- Maintains correct answers and question structure
+- Side-by-side bilingual preview
+- Useful for multilingual programs
+
+---
+
+### 3.6 Rubric Auto-Generation
+
+**Feature**: For discussion and essay questions, auto-generate grading rubrics.
+
+- AI creates criteria, performance levels, and point allocations
+- Aligned to the learning objective
+- Export as PDF rubric alongside the quiz
+- Rubric preview in the Review tab
+
+---
+
+### 3.7 Quiz Difficulty Simulation
+
+**Feature**: Before deploying, simulate how a class of N students might perform.
+
+- AI estimates expected score distribution based on question characteristics
+- Identifies questions that are "too easy" (>95% predicted correct) or "too hard" (<20%)
+- Suggests rebalancing: "Replace Q7 (predicted 98% correct) with a harder variant"
+
+---
+
+### 3.8 Source Citation on Every Question
+
+**Feature**: Each generated question automatically links back to the exact passage/page in the source material.
+
+- "This question was generated from Material X, page 3, paragraph 2"
+- Click to highlight the source passage
+- Useful for: verifying accuracy, student study references, academic integrity
+
+**Technical path**: RAG already retrieves relevant chunks with metadata. Store the chunk reference on each Question document.
+
+---
+
+### 3.9 Voice-Based Quiz Creation
+
+**Feature**: "Create 5 multiple choice questions about photosynthesis, medium difficulty"
+
+- Voice input in Chat Mode
+- Natural language commands for quick quiz creation
+- Useful for instructors who think out loud
+
+---
+
+### 3.10 Template Marketplace
+
+**Feature**: Community library of quiz templates.
+
+- "Intro to Psychology — Midterm Template (30 MC + 5 Essay)"
+- Instructors can publish anonymized quiz structures
+- Others can fork and customize with their own materials
+- Rating system for templates
+
+---
+
+## Priority Matrix
+
+| Feature | Impact | Effort | Differentiation | Priority |
+|---|:---:|:---:|:---:|:---:|
+| Quiz Import & Convert | High | Medium | Medium | **1** |
+| AI Quality Dashboard | High | Medium | Very High | **2** |
+| Quiz Variants | High | Low | High | **3** |
+| Question Bank | High | Medium | Medium | **4** |
+| Study Guide Generation | Medium | Low | High | **5** |
+| Content Coverage Map | Medium | Low | High | **6** |
+| Smart Distractors | Medium | Low | Very High | **7** |
+| Source Citation | Medium | Low | Medium | **8** |
+| Multimedia Sources | High | Medium | High | **9** |
+| LMS Direct Publish | Very High | High | Medium | **10** |
+| Rubric Generation | Medium | Low | Medium | **11** |
+| Spaced Repetition / Anki | Medium | Medium | High | **12** |
+| Multi-Language | Medium | Medium | Medium | **13** |
+| Difficulty Simulation | Medium | Medium | Very High | **14** |
+| Collaborative Authoring | High | Very High | Medium | **15** |
+| Voice Creation | Low | Medium | Low | **16** |
+| Template Marketplace | Medium | High | Medium | **17** |
+
+---
+
+## Recommended Implementation Sequence
+
+### Phase A — Quick Wins (1-2 weeks each)
+1. Quiz Variants (low effort, high wow factor)
+2. Study Guide Generation (reuse existing RAG, new prompt)
+3. Smart Distractors (enhance existing MC generation prompt)
+4. Source Citation (store RAG chunk refs on Question model)
+
+### Phase B — Core Differentiators (2-4 weeks each)
+5. Quiz Import & Convert (new parsing pipeline)
+6. AI Quality Dashboard (Bloom's analysis + LLM review)
+7. Content Coverage Map (visualization of RAG coverage)
+8. Question Bank (Qdrant embedding + search UI)
+
+### Phase C — Platform Features (4-8 weeks each)
+9. Multimedia Sources (Whisper integration)
+10. LMS Direct Publish (Canvas OAuth + API)
+11. Rubric Generation + Anki Export
+12. Multi-Language Support
+
+### Phase D — Advanced (future)
+13. Difficulty Simulation
+14. Collaborative Authoring
+15. Template Marketplace