🎓 NCERT AI Tutor

An AI-powered RAG chatbot for Class 10 NCERT — Mathematics & Science.
Built with Groq (Llama 3.1), ChromaDB, LangChain, and Streamlit.

🚀 Live Demo

Try it here → https://ncert-ai-tutor-appdhcgh4agx3ppbvxt4y7j.streamlit.app

The live demo runs on Streamlit Cloud — no local setup needed, open and start using instantly.

🧠 What This Project Does

Students type questions in English or Hindi → the app retrieves relevant chunks from actual NCERT Class 10 PDFs → Llama 3.1 generates a grounded answer.

No hallucinated answers from the LLM's training data. No random internet content. Just your NCERT textbook — explained like a friend or formatted for board exams.

✨ Features

Feature	Details
📚 RAG Pipeline	Retrieves answers strictly from NCERT Class 10 PDFs
🔍 Subject Filter	Post-retrieval chunk validation — Maths (`jemh`) or Science (`jesc`)
🌐 Bilingual	English and Hindi answer modes
🧠 Answer Modes	Easy Explanation (friend-style) or Exam Preparation (structured points)
💬 Memory	Remembers last 3 conversation turns (`ConversationBufferWindowMemory`)
🛡️ Fallback Defence	3-layer system prevents out-of-syllabus answers
💡 Dynamic Suggestions	Groq-generated follow-up questions after each answer
📝 Export Summary	Download full session as a `.txt` study summary
🎨 Gen Z Dark UI	Purple gradient, WhatsApp-style animated chat bubbles, glassmorphism
⏱️ Timestamps	Every message shows HH:MM AM/PM

🏗️ Tech Stack

Layer	Technology	Version
UI	Streamlit + Poppins + Custom CSS	1.58.0
LLM	Groq API → `llama-3.1-8b-instant`	groq 0.37.1
Embeddings	`sentence-transformers/all-MiniLM-L6-v2` (local, no API)	5.5.1
Vector DB	ChromaDB (local, persisted to `chroma_db/`)	1.5.9
PDF Parsing	PyMuPDF via `langchain_community`	1.27.2
Memory	`langchain_classic.ConversationBufferWindowMemory(k=3)`	1.0.8
Orchestration	LangChain core + classic + community	1.3.9

📁 Project Structure

ncert-ai-tutor/
│
├── app.py                  # Streamlit UI — Gen Z dark theme redesign
├── rag_pipeline.py         # RAG engine — retrieval, filtering, LLM calls
├── ingest.py               # One-time PDF → ChromaDB ingestion script
├── requirements.txt        # All Python dependencies with exact versions
├── .env                    # Your Groq API key (NOT committed to Git)
├── .gitignore              # Excludes .env, chroma_db/, __pycache__, venv/
│
├── data/
│   └── ncert_pdfs/         # Place your NCERT PDF files here
│       ├── jemh101.pdf     # Class 10 Maths chapters  (jemh prefix)
│       ├── jemh102.pdf
│       └── jesc101.pdf     # Class 10 Science chapters (jesc prefix)
│
└── chroma_db/              # Auto-created after running ingest.py (gitignored)

🚀 Getting Started

1. Clone the repo

git clone https://github.com/ayush22cp008/ncert-ai-tutor.git
cd ncert-ai-tutor

2. Create and activate a virtual environment

python -m venv venv

# Windows
venv\Scripts\activate

# macOS / Linux
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Add your Groq API key

Create a .env file in the root folder:

GROQ_API_KEY=your_groq_api_key_here

Get a free key at console.groq.com

5. Add NCERT PDFs

Download Class 10 NCERT PDFs from ncert.nic.in and place them inside data/ncert_pdfs/.

Naming matters! The subject filter reads the filename prefix:

Maths PDFs → must start with jemh (e.g. jemh101.pdf, jemh105.pdf)

Science PDFs → must start with jesc (e.g. jesc101.pdf, jesc202.pdf)

6. Run ingestion (one time only)

python ingest.py

This reads all PDFs, chunks them, embeds with all-MiniLM-L6-v2, and saves to chroma_db/.
Re-run only when you add new PDFs.

7. Launch the app

streamlit run app.py

Open http://localhost:8501 in your browser.

🛡️ How the 3-Layer Fallback System Works

Prevents the LLM from answering outside the NCERT Class 10 syllabus:

User Question
     │
     ▼
ChromaDB: similarity_search(k=10) — retrieves top-10 candidate chunks
     │
     ▼
[Layer 1] Post-retrieval subject filter
          os.path.basename(source).startswith("jemh" / "jesc")
     │
     ├── 0 valid chunks → STATIC FALLBACK returned immediately
     │                    No LLM call. No API cost.
     │
     └── Valid chunks found ↓

[Layer 2] STRICT RULE in prompt (all 4 templates)
          "DO NOT use training knowledge"
          "DO NOT answer outside this context"
          "If not found → say fallback phrase exactly"
     │
     ▼
LLM generates answer using ONLY provided NCERT chunks
     │
     ▼
[Layer 3] Sentinel detection in answer text
          if "not covered in your Class 10" in answer → is_fallback=True
          Effect: dynamic suggestions skipped, memory not updated

⚠️ Known Limitations

Honest limitations — tested and acknowledged:

1. Subject filter is post-retrieval, not pre-retrieval

The ChromaDB where filter was unreliable because it stored full absolute Windows paths (e.g. C:\Users\...\jemh101.pdf) at ingest time, but the filter compared against relative paths. The current fix uses os.path.basename() after retrieval — which works in practice, but retrieves all 10 candidates first before filtering. A proper fix would be to store a subject metadata field (e.g. "Mathematics") at ingest time and filter on that.

2. LLM can still answer from training data on very generic questions

The STRICT RULE prompt reduces this significantly, but questions like "What is AI?" or "Who is the Prime Minister?" may still get answered from Llama's training data instead of triggering the fallback. Adding a query classification step (is this question answerable from Class 10 NCERT?) before calling the RAG chain would fix this completely.

3. Memory is session-only (not persistent across restarts)

ConversationBufferWindowMemory lives in st.session_state — it resets when the browser tab is closed or Streamlit restarts. Persistent cross-session memory would require saving to a file or database.

4. No authentication / multi-user support

This is a single-user local app. All users on the same Streamlit instance share the same global ChromaDB embedding cache (_vector_store). Not designed for production deployment as-is.

5. Hindi support depends on PDF content language

The Hindi answer mode changes the LLM's response language, but the NCERT PDFs ingested are in English. Retrieving chunks in English and responding in Hindi works, but chunk relevance scoring may be slightly lower for Hindi queries.

🔧 Key Engineering Decisions & Bug Fixes

PyMuPDF instead of PyPDF
PyPDF threw LimitReachedError on large NCERT PDFs (>200 pages). PyMuPDF handled all files without issues.

Post-retrieval filter instead of ChromaDB where filter
ChromaDB stores the exact path passed at ingest time. On Windows, this was the full absolute path. The $in filter compared relative paths — they never matched, so filtering was silently disabled. os.path.basename() works regardless of path format.

langchain_classic.memory instead of langchain.memory
langchain.memory import path was deprecated and raises ModuleNotFoundError. The maintained module is langchain_classic.memory.

base_prompt.partial() for injecting chat_history
RetrievalQA only passes {context} and {question} to the chain internally — extra keys like {chat_history} caused ValueError: Missing some input keys. Pre-filling via .partial() before passing the prompt to RetrievalQA fixed this.

@import url() instead of <link> tag for Google Fonts
A <link> tag placed before <style> inside st.markdown() was rendered as visible text by Streamlit. Moving the font import inside <style> as @import url(...) fixed the raw-text CSS display.

Direct prompt .replace() instead of RetrievalQA
The final refactor removed RetrievalQA entirely. We now call vector_store.similarity_search() directly, filter results, build the prompt string with .replace(), and call ChatGroq.invoke(). This gives full control and eliminated all chain-level key errors.

📦 Full Requirements

langchain==1.3.9
langchain-core==1.4.7
langchain-community==0.4.2
langchain-classic==1.0.8
langchain-groq==1.1.3
chromadb==1.5.9
sentence-transformers==5.5.1
pymupdf==1.27.2.3
streamlit==1.58.0
python-dotenv==1.2.2
groq==0.37.1

🙋 Built By

Ayush Halpati — GenAI Developer
B.Tech Computer Engineering, BVM Engineering College, Gujarat
LinkedIn · GitHub

📄 License

MIT License — free to use, modify, and distribute.

"Not just another chatbot. Your NCERT textbook — finally explained the way you actually learn."

Built With

python

Updates

ayush halpati started this project — Aug 02, 2026 04:52 AM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.