AI Tutor — RAG Learning Assistant

A working AI tutor that answers learner questions strictly from the course content it is given — built with FastAPI, OpenAI and a pgvector/Postgres retrieval layer, and running inside Pragyanta today.

How the retrieval pipeline works

Course content — lesson text, slide notes, transcripts — is split into chunks and embedded into vectors stored in Postgres via pgvector. When a learner asks a question, the tutor embeds the question, retrieves the most relevant chunks from that specific course (not the whole content library), and passes them to the language model as context alongside the question. The answer is generated from those retrieved chunks, so it is grounded in what the course actually says rather than the model's general training data.

What that buys you over a plain chatbot

• Answers grounded in your own course material — retrieval-augmented generation (RAG), not a general-purpose assistant that might answer from outside the syllabus
• Guardrails so it stays on-topic — the retrieval step only pulls from the course's own content, so a question outside the syllabus returns "that isn't covered in this course" instead of a plausible-sounding but ungrounded answer
• A pattern that drops into an existing LMS — it is a service behind an API, not a platform of its own, so it can sit alongside content already hosted elsewhere rather than requiring a migration

This is the engine behind our AI Tutor & RAG Learning Assistant service. For the interactive content it can be asked about — video, branching scenarios, drag-and-drop — see our H5P interactive samples.

One design choice worth calling out: the tutor is scoped per-course rather than searching a whole content library for every question. That keeps retrieval fast, and it means the same question asked in two different courses can get two different, correctly-scoped answers instead of one generic one borrowed from wherever the model found a partial match.

Want an AI tutor on your courses? Add an AI tutor to your learning platform.