📄 TestGen – AI-Powered Question Paper Generator

TestGen is a smart, AI-driven question paper generation tool designed to create customized exam papers based on previous year questions, topic weightage, user-defined patterns, and difficulty levels. It enables educators and learners to generate personalized, balanced, and efficient assessments using the power of LLMs, FAISS, and RAG (Retrieval-Augmented Generation).

screen-recording-2025-06-26-113855_4TgM8mko.mp4

🚀 Features

• 🔍 Smart Retrieval: Uses FAISS to fetch the most relevant questions from a large database based on topic, subtopic, and difficulty.
• 🧠 LLM Integration: Generates new questions using Gemini or OpenAI based on topic input and weightage distribution.
• ⚖️ Weightage Handling: Allows custom weightage input (e.g., 40% PYQ, 60% new) and aligns questions accordingly.
• 🧾 Pattern Configuration: Users can define question types (e.g., 5 marks, 10 marks), number of questions, and total marks.
• 🧮 Marks & Topic Mapping: Dynamically assigns marks, subtopics, and serial numbers to each question.
• 📊 Personalized Content Recommendation: Suggests YouTube videos and learning resources based on weak or uncovered areas, identified by gaps in the user’s selected question set.
• 📦 Structured Output: Generates clean JSON or dictionary-like output, easy to integrate into UI or export to PDF.

---

🛠️ Tech Stack

Component	Tech Used
Backend	Python, Flask
LLM Integration	Gemini API / OpenAI GPT-4
Vector Search	FAISS + Sentence-BERT
Retrieval Logic	RAG (Retrieval-Augmented Generation)
Data Handling	Pandas, JSON
YouTube Recommender	YouTube Search API, Custom Matching
Web App (optional)	Streamlit / Flask Frontend

---

📚 How It Works

1. User Input:

   • Select subject, topics, question type (5/10 marks), number of questions, total marks.
   • Set topic-wise weightage and PYQ vs AI-generated split (e.g., 60:40).

2. Question Pool Building:

   • Questions are fetched from past year PDF dumps (parsed and indexed).
   • FAISS retrieves topic-relevant PYQs.
   • LLM generates new questions on uncovered areas based on topic distribution.

3. Paper Assembly:

   • Questions are selected and mapped to marks.
   • Subtopics are auto-assigned using semantic understanding.
   • A final dictionary (or JSON) is created: {"Sr.No": x, "Question": ..., "Subtopic": ..., "Marks": ...}

4. Content Recommendation:

   • Topics with insufficient coverage trigger a YouTube content search.
   • API returns video links to strengthen weak areas.