VOID: AN AI-POWERED EDUCATIONAL CHATBOT USING RAG, FAISS, AND GOOGLE GEMINI

Abstract

Students preparing for examinations frequently lack efficient tools to extract relevant answers from their own study materials or to generate practice questions from them. This paper presents VOID (Void Online Intelligent Documentassistant), a locally deployable educational AI assistant that combines Retrieval-Augmented Generation (RAG), FAISS vector search, and Google Gemini to deliver documentgrounded question answering, automated MCQ quiz generation, and instant quiz evaluation through a single web application. The system ingests PDF, DOCX, and plain-text study documents, splits their content into overlapping 500- character chunks, embeds each chunk using the all-MiniLML6-v2 Sentence Transformer (384-D vectors), and persists the result in a FAISS IndexFlatL2 index. At query time, the top-5 nearest chunks are retrieved and used to ground a prompt sent to Gemini (gemma-3-1b-it). Evaluation on a 20-query test set yields 90% retrieval recall @5 and 75–85% end-to-end answer accuracy. The quiz generation pipeline achieves 100% structural validity across 100 MCQs through a retry-andvalidate loop. FAISS retrieval completes in under 1 ms; quiz evaluation is entirely local with no LLM involvement.