Skip to main content

STEMVentor Mentor — AI LMS Framework

What it is

An AI tutor built on top of all the STEMVentor learning content. It turns the lessons, quizzes and projects of the 12 learning paths into a body of knowledge that a learner can explore, follow and question, and it answers those questions through a web chat that is grounded in the content itself, not in the open web.

What it includes

A web chat that answers from the learning content and cites its sources, a way to browse the whole technology landscape as an interactive graph, and a lesson view with a chat scoped to the lesson in hand. It also supports tutors built around an institution's own material and an advanced path in which learners build a tutor of their own, both described under how an institution can use it.

What makes it different

Grounded, not guessed. A general chatbot answers from what it absorbed in training, and it still guesses. The Mentor answers from the platform's own material: it finds the relevant lessons, checks what is related to them, and only then responds, citing its sources. Because the tutor knows which lesson a learner is in, its answers stay within what is being taught.

Technology Overview

How a question becomes an answer, and the technology used at each step.

  1. Content. Lessons, quizzes and projects are written as MDX files and held in object storage under one flat structure of unique names. Technology: MDX files and object storage.
  2. Ingestion. The content is split into passages and embedded, and the embeddings are stored in Postgres with pgvector. Technology: embeddings, Postgres with pgvector.
  3. Knowledge graph. Topics, lessons and projects become nodes in a Neo4j graph, with "builds on" relationships derived from the content. Technology: Neo4j, with Cytoscape.js to display the graph.
  4. Retrieval. Vector search finds the passages closest in meaning to the question, and the graph adds related and prerequisite material. Technology: pgvector vector search and Neo4j queries, run through a separate FastAPI service.
  5. Context assembly. The retrieved material, the learner's profile and the current lesson are assembled into the prompt. Technology: the FastAPI service, which also holds learner profiles.
  6. Answer. A language model service writes the response from that context and cites the material it used. Technology: a language model service.
  7. Chat interface. A web chat streams the response, keeps the conversation and session, and offers guided actions alongside free text. Technology: a Next.js front end in TypeScript with Tailwind CSS and Zustand, and Auth.js sign-in with Google or an email link.

How an institution can use it

An institution can give its learners the Mentor alongside the learning paths, so that questions are answered from the lesson in front of them.

The Mentor is not tied to STEMVentor's learning paths: the same pipeline can be pointed at the institution's own material, so that the tutor answers from the institution's own content, not from the open web, with learning paths shaped around its context. An institution can start with one subject or course, review how the tutor answers, and extend it over time.

Advanced learners can also build a tutor of their own through Build Your Own AI, an AI learning path suggested for Grades 10 to 12 with about 40 suggested hours. It covers language models and APIs, embeddings, knowledge graphs, retrieval and grounding, and it leads to the AI Course Mentor project (about 30 suggested hours), in which learners build a simplified version of this system. The path can be offered as a stand-alone course, as a follow-on to the AI Landscape path, or as preparation for the AI Under the Hood workshop.

The next phase of the Mentor ties the views together more tightly: starting a lesson from the graph, handing a lesson question over to the open chat, and reflecting a learner's progress back onto the graph.