中文版

DeepTutor（GitHub：https://github.com/HKUDS/DeepTutor ，2025-12-28 创建，已 46,977★，Apache-2.0）是香港大学数据科学实验室（HKUDS）出品的开源 AI 家教平台，口号是”终身个性化辅导”。它不是一个问答机器人，而是围绕”系统地学一门东西”设计的完整学习系统：自托管部署，模型自备，还在高频更新（最新 v1.6.15）。

它能干什么

• 知识库问答：把教材、PDF、EPUB 导入建成知识库，内置 LightRAG、GraphRAG、PageIndex 等多种检索引擎；回答带出处引用，阅读时文档就摆在对话旁边、逐页标注来源
• 引导式学习：内置”精通路径”，按知识点分阶段推进、每关有测评卡关，而不是你问一句它答一句
• 出题与题库：能生成练习题与测验，答题情况沉淀进题库和学习日志，薄弱环节会被记住
• 生成学习书：把一个主题编译成结构化的书，边生成边读，可导出 Markdown
• 深度解题：专门啃难题的 deep_solve 模式，外加数学动画把公式动起来讲
• 记忆与伙伴：它记得你的学习进度；还能把本地的 Claude Code、Codex 等 agent 接进来当陪练，通过微信等十多个 IM 渠道也能对话
• 技能生态：有自己的技能社区，技能可一键安装且带安全检查

怎么用

• 本地安装：pip install deeptutor 后初始化再启动（需要 Python 3.11+ 和 Node 20+），或 Docker 一条命令拉起，也有纯 CLI 版本
• 模型自备：填自己的 API Key，或接本地模型服务（Ollama、LM Studio、llama.cpp 等都支持），官方还配套了站点 deeptutor.info 和一篇 arXiv 论文

适合谁

在职自学新技术、备考、或者想给孩子/学生搭一个”真带学习路径”的 AI 老师的人。相比直接拿聊天模型当老师，它的优势在于知识库有出处、进度有记忆、学习有路径和测评——把逆向工程、编程这类需要长期啃的主题丢进去让它带着学，比零散提问有效得多。

English Version

DeepTutor (GitHub: https://github.com/HKUDS/DeepTutor , created 2025-12-28, now 46,977★, Apache-2.0) is an open-source AI tutoring platform from the University of Hong Kong’s data science lab (HKUDS), billed as “lifelong personalized tutoring”. It is not a Q&A chatbot but a complete learning system designed around systematically learning a subject: self-hosted, bring-your-own-model, and updated at a rapid pace (latest v1.6.15).

What it does

• Knowledge-base Q&A: import textbooks, PDFs, and EPUBs into knowledge bases powered by multiple retrieval engines (LightRAG, GraphRAG, PageIndex, and more); answers carry citations, and immersive reading keeps the document beside the chat with page-level sources
• Guided learning: built-in Mastery Paths advance topic by topic with gated assessments, rather than one-off question-and-answer exchanges
• Quizzes and a question bank: it generates practice questions and tests, and your results settle into a question bank and learning journal — weak spots are remembered
• Book generation: a topic can be compiled into a structured learning book, readable while it generates and exportable to Markdown
• Deep problem solving: a deep_solve mode for hard problems, plus math animations that bring formulas to life
• Memory and partners: it tracks your learning progress, can pull in local agents like Claude Code and Codex as study partners, and can chat through a dozen-plus IM channels including WeChat
• A skills ecosystem: its own skills community, with one-command installs behind a security check

How to run it

• Local install: pip install deeptutor, then init and start (needs Python 3.11+ and Node 20+); or a single Docker command; a CLI-only build also exists
• Bring your own model: your own API key, or a local model service (Ollama, LM Studio, llama.cpp, and others are supported). The project also maintains a website (deeptutor.info) and an arXiv paper

Who it’s for

Anyone self-studying a new technology on the job, preparing for exams, or setting up an AI teacher with a real learning path for a student. Compared with using a chat model as a tutor, its edge is grounded knowledge bases, persistent memory of your progress, and structured paths with assessments — drop a long-haul subject like reverse engineering or programming in, and it teaches far more effectively than scattered questions.