A Subject-Scoped, Retrieval-Augmented Study Assistant for Understanding and Questioning Personal Lecture Notes

Abstract

A student's notes are usually spread across many PDFs, text files, and handouts, and finding one idea inside them can take longer than learning it. Keyword search helps only when the learner remembers the right word. This paper describes a web-based study assistant that treats a learner's own notes as the knowledge source for a language model. Uploaded files are split into short overlapping passages, turned into vectors, and stored with the owner's identity and subject label. A question is answered in three steps: passages from the same owner and subject are ranked by meaning, weak matches are discarded, and the remaining passages are given to the model with an instruction to stay inside them. The reply shows source cards and a simple confidence label, and the same index is used to create flashcards, short questions, and multiple-choice tests. The prototype uses Next.js, Supabase, LangChain text splitting, and OpenAI or Gemini models. No accuracy benchmark was run, so the paper focuses on design, implementation, and a clear plan for testin.