The research behind Lumyo

Students keep notes, slides, readings and lab instructions across different places. The information they need is often already in their own notes, but finding and connecting it while studying takes effort.

General-purpose chatbots can explain and summarize, but their answers aren't necessarily based on the student's own materials. Retrieval-augmented generation (RAG), where a system first looks up relevant documents and then writes an answer, helps with that. The paper argues that in education students also need control over the scope: one lecture, one course, or a chosen set of notes.

The question the paper asks is how a mobile AI learning assistant can help students retrieve, summarize and organize their own notes while letting them choose which sources the AI model uses.

The paper presents the design and implementation of Lumyo as a mobile app with a server back end. Notes are split into chunks and indexed in the background in a vector database. For each conversation, the student picks the notes that may be used as sources.

When the student asks a question, the search is restricted to chunks from those selected notes that belong to that student. The five most relevant fragments, each labelled with its note's title, are passed to a large language model together with up to ten earlier messages from the conversation. The model is instructed to answer from that context, to say when the context doesn't contain enough information, and to refer to the note titles it used. The answer comes back with a list of the source notes behind the retrieved fragments.

The same app uses AI to write a summary for each note and to suggest categories, to help with organizing and reviewing.

The paper describes three contributions: the design and implementation of Lumyo, a retrieval workflow in which students choose which of their notes can be searched, and an analysis of how the architecture supports learner control, transparency of the retrieval process, source provenance and graceful handling of failures.

Its main point is that the student's choice of sources is kept consistent through the whole process, from selecting notes in the app to the search, the generated answer and the source list returned with it. The paper states that it does not propose a new embedding or ranking method.

The paper describes the research system as it was built for the study. The Lumyo app in development grew out of it and keeps the same core idea: you choose the notes for each chat, and answers show their sources.

The paper describes that research setup; it isn't a promise of how the app will perform.