HKU's DeepTutor is now open source! Nearly 30,000 stars — it's not a Q&A bot, but an AI tutor that stays with you for life. Original link: https://github.com/HKUDS/DeepTutor

Right now, if you check the GitHub trending list, DeepTutor has already surged to 29,600 stars — an open-source project from the Data Science Lab at the University of Hong Kong.

💡 What You Will Learn

HKU's DeepTutor is now open source! Nearly 30,000 stars — it's not a Q&A bot, but an AI tutor that stays with you for life. Original link: https://github.com/HKUDS/DeepTutor Right now, if you check t

📜 Table of Contents

• You Feed It Your Materials, It Becomes Your Personal Learning Hub
• Under the Hood: Multi-Engine RAG, Not a Single Monolithic Model
• Memory System: Three-Layer and Fully Traceable, Not a Black Box
• What's the Biggest Difference from ChatGPT?
• How to Get Started? Four Steps from the Official Docs

HKU DeepTutor Goes Open Source! Nearly 30K Stars — It's Not a Q&A Bot, It's an AI Tutor That Learns With You for Life

Original link: https://github.com/HKUDS/DeepTutor

Open GitHub's trending page right now and DeepTutor has surged to 29.6K stars — an open-source project from the Data Science Lab at the University of Hong Kong that hasn't stopped climbing over the past six months.

It's not "just another AI Q&A tool." DeepTutor positions itself as a "Lifelong Personalized Tutoring" system, and its core logic is the complete opposite of ChatGPT: ChatGPT starts from zero with every conversation, while DeepTutor remembers what you've learned, where you got stuck, and continuously adjusts the pace.

You Feed It Your Materials, It Becomes Your Personal Learning Hub

DeepTutor's core use case is simple: import your own materials — textbook PDFs, papers, lecture notes — and let the AI do deep work on those specific resources.

It doesn't search the entire web vaguely; it works based on the materials you specify:

Problem Solving: It doesn't just hand you an answer — it breaks down the full reasoning process into steps, explaining the "why" behind each one.

Quiz Generation & Grading: Upload an exam, and DeepTutor analyzes the question style and difficulty, then auto-generates new practice questions that match. After you complete them, it grades automatically and creates targeted exercises based on your mistakes.

Research: Given a complex topic, it cross-references your knowledge base and the web, then compiles a structured report with citations.

Visualization: It turns abstract concepts into charts, animations, and even interactive content. It has a built-in Math Animator that can turn mathematical formulas into 3Blue1Brown-style visual presentations.

Learning Paths: It tracks your knowledge gaps and continuously adjusts what you should learn next — not a rigid syllabus, but dynamically adapted based on your actual performance.

Under the Hood: Multi-Engine RAG, Not a Single Monolithic Model

DeepTutor's technical architecture is fascinating. It doesn't rely on one model to do everything — instead, it uses an agent-native architecture that chains multiple capability modules together.

Knowledge base retrieval supports 5 engines: LlamaIndex (default local vector + BM25), PageIndex (managed inference-based retrieval), GraphRAG and LightRAG (knowledge graph retrieval), and you can even mount your Obsidian vault directly — letting the AI read and write your notes in place.

On the model side, it supports 25+ LLM providers, including OpenAI, Anthropic, local Ollama, LM Studio, llama.cpp, and even NVIDIA NIM and Lemonade. You can assign different models to different tasks — one for problem solving, another for report writing.

Memory System: Three-Layer and Fully Traceable, Not a Black Box

What makes me think "this was designed by someone who actually understands education" is its memory system.

It doesn't just stuff memories into an invisible vector database. It's a three-layer structure:

L1 is the raw event log (complete records of every conversation); L2 is distilled summary facts organized by scenario; L3 is cross-scenario synthesized judgments.

Every layer is annotated with its source — L2 references which L1 log entries, L3 cites which L2 summaries. You can click into a "memory graph" and trace any generalized conclusion (like "user is weak in linear algebra") all the way back to the specific conversation and the exact problem they got wrong. Auditable, correctable — not a black box.

What's the Biggest Difference from ChatGPT?

I'll sum it up in one sentence: ChatGPT is a great teacher, but it doesn't remember you — every time you ask it something, it's meeting you for the first time.

DeepTutor takes a different path: import your own materials, remember your learning progress, continuously adjust the pace. It's "your personal learning workbench," not a "general-purpose Q&A machine."

Another key difference: it treats AI as an "orchestrator" rather than a "responder." You can have it invoke Claude Code or Codex to help write code, or create AI companions (Partners) with independent personalities and memories that live on Telegram or Feishu, providing long-term tutoring support to students.

How to Get Started? Four Steps from the Official Docs

There are 4 installation paths, and the official recommendation for the smoothest experience is PyPI installation — no need to clone the repo.

Step 1: Create a workspace

mkdir -p my-deeptutor && cd my-deeptutor
pip install -U deeptutor

You'll need Python 3.11–3.13 and Node.js 20+ on your PATH (deeptutor start spins up a built-in Next.js server).

Step 2: Initialize

deeptutor init

This interactively asks for the backend port (default 8001), frontend port (default 3782), LLM provider and API key, and whether to configure an embedding engine. You can skip init too — the app will start with default ports and you can configure everything later in settings.

Step 3: Start

deeptutor start

This launches both backend and frontend. Open your browser and visit http://127.0.0.1:3782 to reach the main interface. Press Ctrl+C to stop both.

Step 4: Play around

Once installed, you don't even need to leave the terminal — run deeptutor chat for an interactive REPL, or deeptutor run chat "explain the Fourier transform" for a one-shot query. Create a knowledge base with deeptutor kb create my-kb --doc textbook.pdf, and inspect memory with deeptutor memory show.

Other installation methods

• Docker one-liner: docker run --name deeptutor -p 3782:3782 -v deeptutor-data:/app/data ghcr.io/hkuds/deeptutor:latest, then open http://127.0.0.1:3782 in your browser
• Source install: Best for developers who want to build on top — clone the repo, run pip install -e ., then npm install for frontend dependencies
• CLI-only mode: No web UI needed — checkout from source and install the packaging/deeptutor-cli package

Licensed under Apache 2.0, no paywalled wrapper versions. Official docs (Chinese): docs.deeptutor.info/zh-cn — installation guides, CLI command reference, and configuration docs are all there. GitHub: github.com/HKUDS/DeepTutor

If you've been using ChatGPT to learn but always feel like "it doesn't remember where I got stuck last time" — DeepTutor might be the answer you're looking for. After all, when it comes to learning, the best teacher isn't the smartest one — it's the one who knows you best.