OpenVidya — AI-Powered Learning for India

Built on OpenMAIC. Redesigned for NCERT, CBSE, and how India actually learns.

OpenMAIC is a brilliant open-source multi-agent classroom engine. OpenVidya forks it with one thesis: Indian students learn differently, and their AI teacher should know that.

Developed by Debjyoti

Recent Fixes

• Anti-hallucination grounding — Outline generator now has explicit rules: stay on the exact topic requested, never invent terminology (e.g. "quantum tenacity"), ground all concepts in KB data when available. Applied at both system and user prompt level.
• Language pipeline cleanup — Removed all Chinese (zh-CN) and Hindi (hi-IN) strings from every layer: agent personas, LLM prompts, API routes, quiz grading, default fallbacks, ASR/TTS hooks. Only English and Sanskrit are supported.
• Story Quest mode fixed — educationMode is now stamped on every SceneOutline and flows into Stage 2 slide generation, so story-format layout instructions actually reach the content LLM.

---

What's Different

5 India-First Learning Modes

Each mode is a complete pedagogical pattern — not just a UI theme. The LLM receives mode-specific scene generation rules, slide layout instructions, and a named teacher persona with its own director routing.

Mode	Persona	What it does
🎙️ Teacher Narration	Teacher	1:1 narration — ask anything in chat, slides update live
📖 Story Quest	Didi	Every concept starts with a named Indian scene — auto-rickshaw braking, Rohit Sharma's pull shot, dosa on a tava
⚔️ Exam Dojo	Coach	CBSE/JEE drill: predict → attempt → trap exposed → reinforce. PYQ-pattern questions
🔬 Lab Without Walls	Lab Sir	Virtual NCERT lab — exact apparatus, NCERT experiment numbers, real-time simulation
⚡ Rapid Revision	Flash	Concept chain sprint — memory triggers in ≤8 words, not re-teaching

NCERT/CBSE Knowledge Base

A structured KB feeds verified curriculum data directly into the LLM prompt — grounding generation in actual syllabus content:

• questions_registry.json — CBSE board questions with difficulty, type, trap, and explanation
• concept-graph.json — Concept dependency chains (prerequisite-ordered DAGs per chapter)
• lab-registry.json — NCERT prescribed experiments with apparatus, objectives, and common mistakes

No KB match → graceful fallback to LLM generation from first principles.

---

Architecture

User prompt
    │
    ▼
/lesson-setup  ──► mode picker (5 modes) + PDF + URL inputs
    │
    ▼
Stage 1: Outline Generator
  ├── buildEducationModeContext(mode)    ← scene pattern rules per mode
  ├── buildKBContext(topic, class, mode) ← NCERT/CBSE curriculum grounding
  └── SceneOutline[] (educationMode stamped on each outline)
    │
    ▼
Stage 2: Slide Content Generator
  ├── buildSlideModeContext(mode)        ← layout rules per mode
  ├── teacherContext (agent persona)     ← Didi / Coach / Flash / Lab Sir / Teacher
  └── Slide JSON (visual layout)
    │
    ▼
Classroom Runtime
  ├── Director — mode-specific routing   ← single-agent rules per mode
  └── Named agent narrates live          ← persona-driven, chat-reactive

Each mode is wired at all three levels: outline generation → slide content → live narration. Changing a mode changes what the LLM generates, how it lays it out, and how the teacher speaks about it.

---

Mode-Specific Design

🎙️ Teacher Narration

65% slides (mix of direct explanation, example-first, story/analogy, visual), 20% quiz, 15% conversation. Slides are sparse — the teacher narrates, not the slide. Student questions update the slide live.

📖 Story Quest

Pattern per concept: story slide (named Indian scenario) → "Why?" slide (single question, no answers) → concept slide (formula + story callback) → applications (2–3 images) → quiz (story-referenced questions). Video generation requested for motion scenes.

⚔️ Exam Dojo

Per-concept drill loop: probable Q slide (exam framing, trap flagged) → primary quiz (MCQ + assertion-reason + case-based) → solution slide (answer path, trap exposed) → practice quiz (harder variant). Every quiz includes a PYQ-pattern question. Ends with weak-spot summary.

🔬 Lab Without Walls

Per-experiment pattern: brief slide (NCERT experiment name, objective, apparatus) → interactive scene (full interactiveConfig with sliders, real-time readings, observable changes) → observation quiz. Maximum 3 interactive scenes per session.

⚡ Rapid Revision

Concept chain (prerequisite-first DAG): intro slide → [concept trigger slide (≤8-word bullets) → quiz] × N → weak-spots summary. 3–6 concepts, 8–14 scenes total. No re-teaching — pure memory reactivation.

---

Credits

OpenVidya is a fork of OpenMAIC by THU-MAIC. All multi-agent orchestration infrastructure, scene generation pipeline, and classroom runtime are OpenMAIC's foundational work. OpenVidya adds India-specific pedagogy, NCERT/CBSE curriculum grounding, and mode-specific generation architecture on top.

---

Demo

What's Next

• Local language support — Hindi, Tamil, Telugu, Bengali (generation + narration)
• Full CBSE KB — Class 6–12, JEE Mains, JEE Advanced, NEET question banks
• NCERT MCP server — curriculum as structured context, consumable by any LLM
• India-specific LLM benchmarks — Sarvam AI, Krutrim, and others on CBSE tasks
• Adaptive difficulty — KB tracks per-student weak spots across sessions

---

License

MIT — same as OpenMAIC.