DeepTutor

1. Project Overview

The core selling point of DeepTutor is "lifelong personalized counseling". It puts multiple links in the learning process in the same context: students can chat and ask questions, do questions, generate tests, do in-depth research, compile the knowledge base into living books, write in Co-Writer, let the memory system continuously accumulate learning portraits, and can also access sub-intelligence such as Partners or local Claude Code/Codex.

As mentioned in the project paper abstract, DeepTutor try to make up for two kinds of deficiencies: ordinary LLM relies on static pre-training knowledge, which is difficult to adapt to individual learners; traditional RAG can provide knowledge enhancement, but it does not necessarily provide personalized and guided feedback. The paper also puts forward the TutorBench to measure the effect of personalized counseling by interactive evaluation with learner's portrait. The paper reports that DeepTutor improved on average by 10.8 percent on personalized metrics and reasoning by 29.4 percent on five backbone models to enhance general agentic. These figures are from project papers, and pre-sales references should be made to indicate that they are the results of a technical report by the author team, not an independent third-party evaluation.

2. Official Key Schematic

The project itself has relatively complete screenshots and architecture diagrams, which are suitable for direct entry into study notes and pre-sales materials.

2.1 Product Overview

The official website describes the DeepTutor as "Eight surfaces, one running context": Home, Partners, My Agents, Co-Writer, Book, Learning Space, Memory and Knowledge Center share the same agent-native runtime. This point is very suitable for product explanation opening: it is not a bunch of isolated tools, but a workspace connected around learner context.

2.2 System Architecture

This diagram is suitable for explaining the DeepTutor system composition: front-end learning space, back-end agent loop, knowledge base/RAG, tools/capabilities, memory, models and external integration together constitute a scalable learning system.

2.3 Chat Agent Loop

Chat is the default entry, but the underlying is not a simple question and answer, but an agent loop: the model can think, call tools, observe results, continue reasoning, and when necessary, pause through the ask_user and clarify to the user.

2.4 Knowledge Center

Knowledge Center is a DeepTutor RAG center that supports multiple knowledge base/retrieval modes such as LlamaIndex, PageIndex, GraphRAG, LightRAG, LightRAG Server, Obsidian vault, etc.

2.5 Memory Graph

The DeepTutor Memory is a three-tier auditable structure: L1 original event tracking, L2 surface-organized facts, and L3 cross-surface comprehensive portrait. It can be said before sales as a "traceable portrait of the learner", not a hidden black box memory.

2.6 Book Living Book

Book can turn knowledge bases, notes, question banks, or chat history into interactive live books containing text, quizzes, flashcards, timelines, code, graphics, interactive HTML, animations, concept maps, and page-level chats.

3. What can it mainly do

DeepTutor can be understood as an "educational AI agent workspace". It combines capabilities common in learning products:

4. Core Design Understanding

4.1 is not a "question and answer robot", but a unified learning runtime

Many educational AI products have made multiple decentralized functions: a page question and answer, a page generation topic, a page making courseware, and a page reading document. The DeepTutor difference is that Chat, Quiz, Research, Visualize, Solve, and Mastery Path run on the same agent loop. Users switch to learning goals, not to change a set of isolated engines.

Before the sale can say:

the DeepTutor idea is to unify the learning activities into a continuous context: what students have asked, what they have read, what they have practiced, what they have missed and what they have saved can be reused by subsequent learning activities.

4.2 RAG is not just data retrieval, but part of the learning context

The DeepTutor knowledge base supports several search engines:

-LlamaIndex: default, local vector BM25.

-PageIndex: Managed, inferential retrieval of page-level references.

-GraphRAG/LightRAG: Knowledge map search.

-LightRAG Server: Connect to an external LightRAG instance.

-Obsidian vault: direct link Obsidian note library, read and write in-situ content.

Document parsing can also be switched: Text-only, MinerU, Docling, markitdown, PyMuPDF4LLM, etc.

This is of great significance to the educational scene: teaching materials, courseware, lecture notes, exercises, student notes and teacher lesson plans of schools and training institutions can all enter the knowledge base and become the basis for AI guidance.

4.3 Memory is an auditable portrait of personalized learning.

DeepTutor divide memory:

It emphasizes that L2 refers to L1 and L3 refers to L2, so the memory conclusion can be traced back. This is very suitable for educational compliance and home-school communication: it is not just "AI that this student does not understand functions", but should be traceable to specific exercises and dialogue evidence.

4.4 Skills and EduHub: Education Scenario Scalable

DeepTutor using the open Agent-Skills format, the skill is a folder containing 'SKILL.md. EduHub is its built-in educational skills registry, which supports searching, installing and publishing teaching skills, such:

-Socratic tutor.

-Flash card generator.

-Composition feedback.

-Examination blueprints.

-Concept interpreter.

-Review plan generation.

This is very useful for pre-sales, because customers often want to deposit their own teaching methods, course SOPs, problem-solving specifications, and marking rules instead of general AI. Skills offers a lighter-weight approach to "teaching ability encapsulation.

4.5 Partners and multi-channel companion learning

DeepTutor Partners can be understood as "persistent learning partners with their own personality, knowledge base, skills and memory" and can access IM channels. The channels mentioned in README/release include flying book, Slack, Mattermost, enterprise WeChat, DingTalk, Telegram, Zulip, Matrix and other related SDK/channel capabilities.

Pre-sales can map it:

-AI teaching assistants in the class group.

-Q & A partner in the course program group.

-Learning reminder/answering assistant in enterprise training group.

-Code teaching assistant in programming learning group.

5. Applicable Scenario

5.1 K12/University Intelligent Assistant

DeepTutor suitable for a student-oriented "learning workbench":

-Students upload courseware, teaching materials, handouts.

-AI based on course knowledge base to answer questions.

-Generate variant questions and quizzes for wrong questions.

-Record weak points via Memory.

-Generates a staged Mastery Path.

-Put chapter knowledge into Book Living Book.

Customer Value:

-Reduce the pressure on teachers to answer questions repeatedly.

-Give students more continuous personalized tutoring.

-Precipitate learning behavior into analyzable data.

Personalized Learning Products of 5.2 Education and Training Institutions

Teaching and training customers are often concerned about "how to scale to provide an experience similar to one-on-one counseling". DeepTutor can be used as product prototype reference:

-Unified course materials and question bank.

-Students get personalized AI coaching.

-The learning path changes according to the exercise performance.

-Wrong questions, notes, question and answer and question bank linkage.

-The teaching and research team encapsulates their teaching strategies into skills.

Suitable for English, mathematics, programming, postgraduate entrance examination, vocational qualification examination, such as the need for continuous practice and feedback of the scene.

5.3 enterprise training and post learning

In-house training often has a large number of systems, products, sales skills, technical documents, and operation manuals. DeepTutor can be used:

-New employee learning path.

-Product/program knowledge Q & A.

-Sales/pre-sales aptitude test.

-Role-playing Q & A and walkthroughs.

-Compile the knowledge base into an interactive learning manual.

In this scenario, Memory can record the learning progress of employees, the Knowledge Center can connect to enterprise data, and the Co-Writer can assist in writing plans/reports.

5.4 AI programming education and developer training

DeepTutor My Agents can access local or external coding CLI such as Claude Code, Codex, Gemini, Kimi, opencode, and MiMo. Suitable for:

-Programming class AI teaching assistant.

-Step-by-step explanation of code questions.

-Code review and debugging in project-based learning.

-Let students ask concepts, write code, and call coding sub-intelligences in the same learning space.

In such scenarios, DeepTutor can be combined with code warehouses, course documents, question banks, and programming practice platforms.

5.5 teachers' lesson preparation, textbook reprocessing and interactive courseware

Book, Co-Writer, Visualize, Math Animator are also valuable to the teacher side:

-Generate handouts into chapter outlines.

-Generate interactive courseware, concept maps, quiz blocks.

-Convert the content of the textbook into a "living book".

-Expand, rewrite and generate examples of a lesson plan.

-Generate mathematical animation with Manim.

Don't just talk about the student side before sales, the teacher side can also form a clear value.

6. Not quite the scene

7. How to install and use

7.1 PyPI installation

Full Local Web App CLI:

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

Default port:

-Back end: '8001'

-Front end: '3782'

Open after startup:

http://127.0.0.1:3782

Requirements:

-Python '>=3.11,<3.14'

-PyPI package 1.5.4 uploaded on 2026-07-24

-Node.js runtime is also required for complete Web applications. README Chinese version mentions Node.js 20, and source code development suggests Node.js 22 LTS

7.2 Docker Installation

docker run --rm --name deeptutor \
  -p 127.0.0.1:3782:3782 \
  -v deeptutor-data:/app/data \
  ghcr.io/hkuds/deeptutor:latest

Only the '3782' port needs to be published. The front end forwards '/api/' and '/ws/' to the FastAPI back end through the Next.js middleware in the container. Data, configuration, API Key, log, workspace file, memory and knowledge base are saved in deeptutor-data '.

7.3 source code installation

git clone https://github.com/HKUDS/DeepTutor.git
cd DeepTutor
python3 -m venv .venv && source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .
( cd web && npm ci --legacy-peer-deps )
deeptutor init
deeptutor start

7.4 CLI usage

deeptutor chat
deeptutor chat --capability deep_solve --kb my-kb --tool rag
deeptutor run chat "解释傅里叶变换"
deeptutor run deep_research "调研 2026 年 RAG 论文" \
  --config mode=report --config depth=standard

Machine-readable mode:

deeptutor run deep_solve "Find d/dx[sin(x^2)]" --tool reason --format json

'-- format json' outputs NDJSON, which is suitable for being called by other agents or automation systems.

8. Deployment and integration methods

8.1 single-user local learning space

The default certification is closed, suitable for personal learning, teacher lesson preparation, developer trial. The data is placed under 'data/'in the current working directory, which can also be specified with 'DEEPTUTOR_HOME or 'deeptutor start -- home.

8.2 Docker self-hosting

Suitable for pilot within schools/institutions. Both the FastAPI backend and the Next.js frontend run in the container, and the data directory is mounted to the host. Common reverse proxies only need to proxy the frontend '3782', and the API is forwarded by the frontend server.

8.3 multi-user shared deployment

DeepTutor support optional authentication and quarantine workspaces:

data/
├── user/                    # 管理员工作区 + 全局设置
├── users//             # 每个用户的聊天历史、记忆、笔记本、知识库
├── partners//workspace/ # Partner 工作区
└── system/                  # auth/users.json、grants、audit

The first registered user becomes an administrator and can be assigned model, knowledge base, skills, Partner, tools/MCP policies, and code execution permissions. For education clients, this means segregation and empowerment by class, teacher, student. However, multi-user deployments still require strict validation of the permission model and data isolation.

8.4 model and knowledge base integration

The model level supports multiple provider, including OpenAI, Anthropic, DashScope, Perplexity, NVIDIA NIM, Gemini, local Ollama/LM Studio/llama.cpp/vLLM/Lemonade, and other related directions. The specific support varies with the version, and the actual configurable items of Settings → Models shall prevail in PoC.

The knowledge base level can be connected:

-Document upload.

-LlamaIndex local vector BM25.

-PageIndex.

-GraphRAG.

-LightRAG/LightRAG Server.

-Obsidian vault.

-Different document parsing engines.

9. What can I say before sales

9.1 Recommended Opening

DeepTutor is not a single point of AI answer robot, but a AI learning workspace built around the continuous context of the learner. It puts question answering, problem solving, testing, research, courseware generation, knowledge base, learning memory and multi-channel teaching assistants in an agent-native runtime, so that AI can not only answer a question, but also continuously understand a student, a textbook, a course and a learning goal.

9.2 three core selling points

9.3 Demo Story Line

A 30-minute presentation for educational clients:

1. Create a course knowledge base and upload PDF/handouts.
2. Answer student questions based on the knowledge base in Chat and demonstrate the citation/retrieval process.
3. Use Deep Solve to solve a course problem and require step-by-step explanation.
4. Use Deep Question to generate tests for the same knowledge point.
5. Generate the chapter materials into Book, and display quiz block, interactive block, and page chat.
6. Open the Memory Graph and explain how learning portraits can be traced back.
7. Show the question bank, notebook, persona and skill in the Learning Space.
8. Demonstrate models, knowledge bases, tools, and multi-user control in the Settings.

9.4 speech for different customer roles

10. PoC Recommendations

10.1 PoC Target

It is recommended that PoC be defined:

verify whether the DeepTutor can be used as a AI personalized learning space for a certain course/certain type of training, realizing data access, knowledge-based answering, topic explanation, exercise generation, learning memory and teacher review closed loop.

10.2 PoC Data Preparation

10.3 PoC tasks

10.4 evaluation index

10.5 PoC cycle

Recommended 2-4 weeks:

-Week 1: Deployment, Model Configuration, Knowledge Base Import, Core Function Run-through.

-Week 2: Teachers/teaching and research create test questions and explanations, and collect quality feedback.

-Week 3: Small-scale student trial, observation experience, accuracy and personalization.

-Week 4: Security, permissions, operations, costs and secondary development assessment.

11. Risks and Considerations

11.1 Education Compliance and Minor Data

If used in K12 or minor scenarios, focus on:

-Student personal information and learning behavior data.

-Whether the dialogue content, wrong questions and ability portrait can be deleted.

-Whether the parent/school is informed and authorized.

-Whether to call the overseas model API.

-Whether unnecessary external services can be turned off.

-Learning whether the portrait may cause labeling or discriminatory judgments.

11.2 LLM Illusion and Teaching Responsibility

DeepTutor enhances the learning experience, but does not take AI answer as absolutely correct. Formal teaching requires:

-Teachers review key content.

-High risk answers retain references and reasoning process.

-Set boundaries for exam scoring, recommendations for further studies, psychological advice, etc.

-- Show students a "AI can go wrong" tip.

11.3 Deployment Complexity

DeepTutor is a complete learning platform, not a lightweight plugin. It involves:

-Python dependency.

-Next.js front end.

-FastAPI the back end.

-Multi-model configuration.

-Embed models and RAG.

-Document parsing engine.

-Docker/Podman.

-Multi-user permissions.

-Optional code execution sandbox.

Pre-sales to avoid the commitment to "one day online enterprise production". PoC can be fast and production requires engineering.

11.4 code execution and skill security

DeepTutor support Office skills, model generation code execution, skills import, MCP/tools, and more. These are strong extension points and security boundaries:

-Untrusted skills need to be audited.

-Code execution should be sandboxed or sidecar enabled.

-Multi-user deployment to restrict common user tool permissions.

-API Key should not be exposed to students.

-Although EduHub/ClawHub skills import has security gates, it still requires internal review of the organization.

11.5 Project Maturity

DeepTutor star is fast-growing and release-intensive, it is a rapidly evolving open source project. The latest release v1.5.4 is still frequently fixing chat responses, IM tables, LLM JSON parsing, KB forms and other issues. This means that the project is active, but it also means that the interface, experience, and deployment methods may change.

12. Relationship with adjacent products/solutions

13. My Pre-Sales Judgment

DeepTutor is an open source project that fits well as a "product blueprint for educational AI. Its value is not that a certain function is particularly new, but that a group of abilities that educational AI should have are put into the same context: curriculum knowledge base, student memory, topic generation, problem solving explanation, research report, interactive book, teacher writing, skill ecology, IM assistant, multi-user isolation and CLI/Agent interface.

For pre-sales, it is especially suitable for use in the following directions:

-Smart teaching assistant/AI class teacher.

-Personalized learning space.

-AI topic explanation and exercise generation.

-Teacher lesson preparation and interactive textbook generation.

-Enterprise training intelligent learning platform.

-Programming Education Agent Workbench.

But it's more like a self-hosted open source platform/product prototype, not a full-fledged SaaS that can be sold directly to schools. When it comes to the ground, the most important thing is three things:

1. The course content and the quality of the question bank should be better, otherwise the AI is only the packaging layer.
2. The boundaries of teacher audit and teaching responsibility should be clearly designed.
3. Student data, model calls, permissions and code execution security must go through compliance review first.

If the customer has a technical team, clear curriculum resources and personalized learning demands, it is DeepTutor worth doing PoC. If the customer only wants a lightweight AI question-and-answer portal, the DeepTutor may be biased. You can only learn from the design ideas of Knowledge Center, Memory, Book and Skills instead of introducing them completely.

14. REFERENCE

-GitHub repository: https://github.com/HKUDS/DeepTutor

-Official project page: https://deeptutor.info/

-Official Document Entry: https://deeptutor.info/

-arXiv Paper: https://arxiv.org/abs/2604.26962

-PyPI package: https://pypi.org/project/deeptutor/

-Latest Release v1.5.4:https://github.com/HKUDS/DeepTutor/releases/tag/v1.5.4

-Chinese README:https://github.com/HKUDS/DeepTutor/blob/main/assets/README/README_CN.md

-Containerized documents: https://github.com/HKUDS/DeepTutor/blob/main/CONTAINERIZATION.md

-Roadmap:https://github.com/HKUDS/DeepTutor/issues/498

-EduHub:https://eduhub.deeptutor.info/