International Journal of Scientific Research and Engineering Development

Large language models have shown excellent performance in tasks related to general purpose reasoning, but their use in educational institutions is extremely minimal because of hallucinations and it’s inability to maintain the answers within the context of approved course materials. This paper presents Virtual Teaching Assistant, a Retrieval Augmented Generation (RAG) system that is designed to deliver contextually accurate and citation based answers to student queries using the resources provided by institutions. The system combines semantic embedding curated knowledge base which comprises information from course material (as markdown file for each chapter) and discussion forum threads (As JSON for each topic of discussion), with real-time cosine similarity based retrieval and GPT-4o-mini inference. An Optical Character Recognition (OCR) method extends the pipeline to image based inputs, enabling students to attach screenshots of assignments or errors. Deployment is done via containerized FastAPI and a responsive single page front end. Simulations demonstrates that cosine similarity threshold of τ = 0.40 over 384 dimensional all-MiniLM-L6-v2 embeddings gives a high precision retrieval while providing sufficient recall for different query types. The proposed architecture offers reproducible blueprint for institution specific and hallucination resistant academic assistants.

Harsh Choudhary, Gaurav Malik, Deepanshu Sharma, Bersha Kumari, Madhu Choudhary,"Virtual TA: QA Over Course Specific Knowledge Base for Educational Institutions" International Journal of Scientific Research and Engineering Development, V9(2): Page(2190-2195) Mar-Apr 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.