Question Craft is an AI-powered application that automatically generates high-quality multiple-choice questions. Given a sample question, the system uses Llama 2 to create a new question that mirrors the topic, style, and difficulty of the original—making it ideal for practice, assessment design, and educational tooling.

• 🎯 Style- and difficulty-aware question generation
• 🔍 Real-time factual verification using web search
• ⚡ Instant feedback on answer selection
• 🔐 Secure authentication and session management
• 🌐 Fully decoupled frontend–backend architecture

Question Craft combines a lightweight web frontend with a modular, AI-driven backend that handles generation, verification, and answer validation.

The frontend is built with standard web technologies (HTML, CSS, JavaScript) and focuses on clarity, speed, and ease of use.

1. Input
   On sampleInput.html , users submit:

   • Topic
   • Question prompt
   • Four answer choices
   • Correct answer

2. Serialization
   The input is packaged into a structured JSON object and stored temporarily in browser local storage.

3. Backend Routing
   After redirecting to questionPage.html, the frontend:

   • Calls a Vercel API endpoint (/api/getLinks)
   • Retrieves the active backend URL from Upstash Redis
   • Sends the question JSON to the backend for processing

4. Interaction & Feedback
   A loading screen is displayed during generation. Once complete, users can answer the new question and receive immediate correctness feedback.

The backend runs in a Python environment and is orchestrated via a Jupyter Notebook
(langchain-experimentation-2-main.ipynb) that launches a Flask server.

• 🌍 ngrok exposes the local Flask server to the public internet
• 🧭 On startup, the server registers its ngrok URL in Upstash (Redis) for service discovery
• 🔐 Authentication is handled via Vercel serverless functions, using:
  • bcryptjs for password hashing
  • Cookies for session management
  • Redis for persistent user data

The backend uses a multi-step LangChain pipeline to ensure both creativity and correctness:

• A few-shot prompt is constructed using the user’s sample question
• The meta-llama/Llama-2-13b-chat-hf model generates:
  • A new question
  • Four answer choices
  • A proposed correct answer

• The generated question is sent to the Tavily Search API
• Relevant real-world context and supporting facts are retrieved from the web

• A second prompt is built using:
  • The generated question
  • Its answer options
  • Tavily’s retrieved context
• Llama 2 re-evaluates the question and confirms the correct answer based on verified information

✔️ The finalized, verified question is returned to the frontend as a JSON object.

• HTML
• JavaScript
• SCSS / CSS

• Python
• Flask
• LangChain

• Llama 2 (via Hugging Face transformers)

• Vercel — Serverless APIs (authentication & service discovery)
• Upstash — Managed Redis (URLs, sessions, user data)
• ngrok — Secure tunneling for local backend exposure
• Tavily — Real-time search API for factual verification

• Node.js
• TypeScript
• bcrypt.js
• ioredis