RAG Study Assistant

Instructors answer the same questions over and over: assignment scope, deadline rules, concepts already covered in lecture. Built during my research assistantship at TRU, this system answers those questions automatically, and answers them the way an instructor would.

It is a Retrieval-Augmented Generation (RAG) pipeline grounded in course material. Every response cites the source it retrieved from, and the tone follows the pedagogy of the course rather than generic chatbot phrasing.

Generic LLM chatbots fail in a classroom for two reasons: they hallucinate answers that contradict the course material, and they hand students full solutions when the instructor wants guided hints.

The goal was a system that stays grounded in the actual course content and respects how instructors want students to learn.

Course documents are chunked and embedded into a vector index. At query time, semantic search retrieves candidate passages, a transformer-based re-ranker orders them, and the generator composes a response constrained to the retrieved context.

A pedagogy layer sits on top: prompts are templated per course policy, so a first-year programming course gives hints while a research seminar gives direct references.

Retrieval quality was measured against a hand-labeled set of real student questions. Integrating semantic search with transformer-based retrieval improved response relevance by 40% over the keyword baseline.

Instructor feedback loops flagged weak answers, and those flags fed directly back into chunking and re-ranking improvements.

The system scaled to 3,000+ students, and repetitive instructor questions dropped by 35%. Office hours shifted from restating logistics to actual teaching.