EduRAG: A Multi-Format Retrieval-Augmented Generation Assistant for Academic Question Answering

Large Language Models (LLMs) can produce fluent responses, but they often provide unsupported information when answers rely on specific documents. This paper introduces EduRAG, a Retrieval-Augmented Generation (RAG) assistant designed for answering student questions based on uploaded academic materials. The proposed system combines multiple file formats, semantic chunking, dense embedding generation, vector similarity retrieval, and context-based response synthesis. EduRAG supports various educational formats, including PDF, DOCX, PPTX, TXT/MD, CSV/XLSX, and image-based text through OCR. This allows practical use across different classroom resources. The backend is built using Flask and includes APIs for health monitoring, document upload, indexing, querying, file listing, and deletion. Sentence-Transformers are used for creating semantic embeddings, and FAISS provides efficient nearest-neighbor retrieval. For generation, the architecture supports both cloud-hosted and local LLM options through configurable providers. Index persistence and metadata storage allow for reusable sessions and quicker follow-up queries. Experimental results on a mix of academic documents show that EduRAG enhances answer relevance and grounding compared to direct LLM prompting without retrieval. It maintains acceptable latency for interactive educational use. The system shows that retrieval-based prompting significantly reduces the risk of generating false information and increases trustworthiness in academic assistants. EduRAG offers a low-cost, flexible foundation for institutional learning support and can be further improved with reranking, citation tracing, and hybrid retrieval methods.