HKUDS/DeepTutor

What it solves

DeepTutor provides a lifelong personalized tutoring system that transforms static learning materials into interactive, agentic educational experiences. It addresses the need for adaptive learning by combining knowledge retrieval, personalized memory, and guided instruction to help users master complex subjects.

How it works

DeepTutor uses an agentic engine that integrates several core components:

• Knowledge Integration: It utilizes RAG (Retrieval-Augmented Generation) through various engines like LightRAG, GraphRAG, and PageIndex to ingest documents, GitHub repos, and web sources.
• Personalized Memory: A three-layer memory system (L1/L2/L3) tracks learner progress and preferences over time.
• Guided Learning: It creates "Mastery Paths" and "Books" that gate tools and content based on the learner's readiness and progress.
• Tool Ecosystem: The system can execute code in isolated runtimes, perform deep research, and use a plugin-driven visualization catalog to explain concepts.
• Agentic Interface: It supports "Partners" (specialized agent personas with their own libraries and skills) and a CLI for agent-native interaction.

Who it’s for

It is designed for learners seeking a personalized, AI-driven educational path and educators or developers who want to build structured, interactive tutoring environments using their own knowledge bases.

Highlights

• Multi-Engine RAG: Supports LightRAG, GraphRAG, and PageIndex for high-fidelity document retrieval and reasoning.
• Immersive Reading: Features a document-side-by-side view with page-by-page citations.
• Agentic Partners: Allows users to create and consult specialized AI personas with private memories and unique skill sets.
• Dynamic Content Generation: Automatically compiles structured "Books" and quizzes from source materials to track mastery.
• Extensible Tooling: Includes a notebook console, isolated code execution, and integration with MCP (Model Context Protocol) servers.