DocQuiz: RAG-Based Quiz Generator from PDFs

DocQuiz generates quizzes directly from uploaded PDF documents.

It combines:

• Retrieval-Augmented Generation (RAG) for document-grounded context
• Local vector search with Chroma
• Local LLM inference with Ollama (llama3)
• A Flask API backend and Next.js frontend

Problem Statement

Teachers and students often have long study material (slides, notes, books) but limited time to create practice questions manually.

DocQuiz solves this by:

• ingesting a PDF,
• indexing it as embeddings,
• retrieving relevant chunks for a query/topic,
• generating structured quiz questions in JSON format.

Supported quiz types:

• Open-ended
• Multiple-choice (MCQs)
• True/False

Core Concept (How It Works)

1. Upload PDF

   • Endpoint: POST /upload-pdf
   • File is stored in uploads/
   • PDF text is loaded and chunked
   • Chunks are embedded with BAAI/bge-small-en
   • Stored in Chroma under a unique doc_id collection

2. Retrieve Context

   • For quiz requests (/open-ended, /mcqs, /true-false), backend runs similarity search on that doc_id collection.
   • Top chunks are joined as context.

3. Generate Quiz

   • Backend builds a strict JSON prompt.
   • Calls Ollama (llama3) with context + user query.
   • Parses model output into valid JSON response.

4. Render in UI

   • Frontend (quiz-generator/) uploads PDF, stores doc_id, requests quiz questions, and renders quiz flow + results.

Project Structure

Path	Purpose
server.py	Flask backend API, retrieval, prompt construction, quiz generation
vectorization.py	Optional one-time script to pre-index sample PDFs from data/
quiz-generator/	Next.js frontend (App Router)
database/	Local Chroma storage (gitignored runtime data)
uploads/	Uploaded PDFs (gitignored runtime data)
data/	Sample local PDFs for experimentation

Tech Stack

• Python + Flask
• LangChain + Chroma
• HuggingFace BGE embeddings (BAAI/bge-small-en)
• Ollama (llama3)
• Next.js 15 + React + Tailwind CSS

Prerequisites

• Python 3.11+ (3.13 also works in this setup)
• Node.js 18+
• Ollama installed and running locally
• Ollama model pulled:

ollama pull llama3

Reproduce the Project (End-to-End)

1) Clone and setup backend

git clone https://github.com/Raditya0902/rag-project.git
cd rag-project

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2) Start backend

python server.py

Backend runs at: http://127.0.0.1:5000

3) Setup and start frontend

Open a second terminal:

cd quiz-generator
npm install
npm run dev

Frontend runs at: http://localhost:3000

4) Use the app

1. Open http://localhost:3000
2. Upload a PDF
3. Choose quiz type and number of questions
4. Start quiz

API Reproduction via cURL (without UI)

Upload PDF

curl -s -F "file=@/absolute/path/to/your.pdf" http://127.0.0.1:5000/upload-pdf

Important: @ must be present in curl form upload.

Generate MCQs

curl -s -X POST http://127.0.0.1:5000/mcqs \
  -H "Content-Type: application/json" \
  -d '{"query":"neural networks","qs":3,"doc_id":"<DOC_ID_FROM_UPLOAD>"}'

Optional: Pre-index Sample PDFs

python vectorization.py

This indexes files from data/ into default Chroma storage.
Primary flow remains: upload with /upload-pdf and use returned doc_id.

Notes and Good Practices

• database/ and uploads/ are runtime artifacts and should not be committed.
• quiz-generator/.next and quiz-generator/node_modules are build/dependency outputs and are gitignored.
• Use only PDFs you are authorized to use/distribute.