ARCHIVES

Classroom assessment continues to rely heavily on manually created quizzes, which consume a significant time from educators and they fail to reflect how students are actually progressing. General purpose AI tools that attempt to generate quiz questions tend to hallucinate may generate reasonable results but they are not grounded in terms of the material uploaded. This paper aims to addresses both problems by designing and implementing an AI-driven adaptive quiz system that combines Retrieval-Augmented Generation (RAG) with a multi-agent reasoning pipeline to produce assessments. The system implemented uses Retrieval-Augmented Generation to extract content from a teacher-uploaded PDF and generate multiple choice questions from it, making sure that the output stays grounded to the material. The generation pipeline is built along four agents built using LangGraph: an adaptive difficulty agent that adjusts the quiz difficulty based on the session average scores, a quiz generation agent that creates the questions, a review agent that checks each question for issues both through rule-based checks and an LLM evaluation, and an explanation agent that generates a short explanation for each correct answer. Human-in-the-loop interface was also introduced to let the instructor know that the flagged questions are marked in red with the reason and the correct are marked in green. The system was implemented was mainly built using Python, LangGraph, LangChain, FAISS, SentenceTransformers, Groq API, Streamlit, SQLite, and Plotly. It was also tested with 41 unit tests to verify that each component we implemented worked as we expected. When tested it across multiple sessions, the structural review passed about 92% of generated questions, quiz generation took between 18 and 35 seconds, and a clear difference in scores between easy and hard difficulty modes, which confirmed that the adaptive mechanism was actually working.