helixnow/deep-student

---

Why DeepStudent

Learning workflows are spread across too many tools — read here, take notes there, build mind maps elsewhere, review in yet another app. PDF readers, mind-mapping tools, translation apps, note-taking tools, LMS platforms, arXiv, flashcard apps, AI assistants… every tool is its own silo. Once your learning data is scattered, you spend more energy shuttling between tools than actually learning.

DeepStudent’s answer: give AI native read-write access to all your learning data. One sentence from you, and it generates a mind map from your textbook, creates questions from your materials, turns key points into flashcards, searches and downloads papers, or researches the web and writes conclusions into your notes — all without leaving the workbench.

---

What You Get

The core idea isn’t “more features” — it’s the unified data layer. The same material can be read, queried, turned into a mind map, used to generate quizzes, made into flashcards, researched, and written back — no data shuttling between apps.

---

Core Capabilities

1. Study with AI Chat

Study around your materials, not just general chat.

• Multi-modal input (drag & drop images / PDF / Word) with multi-turn conversation
• Reference panel for injecting knowledge base notes or textbooks into context, with real-time token estimation
• Deep reasoning mode (chain-of-thought), showing the full thinking process
• Multi-tab sessions & session branching — explore different approaches
• Authority modes Ask / Plan / Craft (default Craft): read-only · plan-then-approve · execute with tool approvals
• Learning desktop (OS / Workbench mode): multi-window study shell around chat and resources
• Multi-model comparison (experimental): side-by-side answers from multiple models
• Session grouping, group-level System Prompt, default skill configuration
• Sub-agent execution (experimental): automatic task decomposition, background completion

2. Learning Hub

Organize materials, notes, questions, mind maps, translations, and flashcards in one place.

• Full-format management: notes / textbooks / question banks / mind maps
• Auto-vectorization pipeline on import (OCR → chunking → embedding → indexing), with real-time status
• Built-in PDF / DOCX reader with dual-page view and bookmarks
• Reading mode toggle — prevents keyboard popup on mobile during scrolling
• Content search across sessions and resources with session tagging
• Resource export with format-specific adapters
• Unified data source for downstream Q&A, mind maps, question generation, and flashcards

3. Knowledge Mind Maps

Structure your knowledge, not just get answers.

• Generate a complete knowledge structure from a single sentence (e.g., “generate a high school biology mind map”)
• Multi-round conversational editing of nodes
• Toggle between outline view and mind map view, right-click menu editing
• Node masking for recitation practice

4. Question Sets & Practice

Turn textbooks and exam papers into practice-ready question banks.

• Upload textbooks / exam papers, AI auto-extracts or generates question sets
• Daily practice, timed practice, mock exams with auto-grading
• Question history view — review past practice sessions and track progress over time
• AI deep analysis of knowledge points and problem-solving approaches
• Mastery tracking by knowledge point (feeds weak-point profile & bounded FSRS review bias)

5. Anki Smart Flashcards

Push understanding into long-term memory.

• Trigger card creation via natural language in chat (e.g., “turn this document into flashcards”), with batch generation
• Visual template editor (HTML / CSS / Mustache) with real-time preview
• Task board for batch card creation progress tracking with checkpoint resume
• 3D flip preview, one-click sync to Anki

6. PDF / DOCX Smart Reader

Study around your documents, not just open them.

• Full format support: PDF, DOCX
• Split-screen: chat on the left, read on the right
• Select pages or passages to auto-inject into chat context
• AI responses can include page number references

8. AI Essay Grading

Chinese and English essay grading and polishing.

• Multi-scenario: Gaokao / IELTS / TOEFL / CET-4/6 / Postgraduate entrance exam
• Multi-dimensional AI scoring (vocabulary, grammar, coherence, etc.) with iterative grading
• Revision suggestions with highlights
• Sentence-by-sentence polish comparison
• Customizable scoring dimensions and grading settings

9. Deep Research

Multi-step, long-chain research agent.

• Interactive confirmation of research depth and format preferences before starting
• Automatic task decomposition: define objectives → web search → local retrieval → analysis → report generation
• 7 search engines supported (Google CSE / SerpAPI / Tavily / Brave / SearXNG / Zhipu / Bocha)
• Reports auto-saved as notes

10. Academic Paper Search & Management

One-stop paper retrieval, download, and citation.

• Search via arXiv / OpenAlex with structured metadata
• Batch PDF download, auto-saved to VFS, multi-source fallback (arXiv → Export mirror → Unpaywall)
• SHA256 deduplication
• BibTeX, GB/T 7714, APA citation formats
• DOI auto-resolution to open-access links

11. Smart Memory

Gets smarter the more you use it.

Inspired by mem0 and memU, implementing a complete memory lifecycle on desktop.

• Auto-extracts user facts after each conversation (identity / preferences / goals / subject status)
• Vector comparison of new vs. existing memories, LLM decides ADD / UPDATE / APPEND / DELETE / NONE
• Batch memory write with write idempotency for data integrity
• Aggregated into user profile, auto-injected into subsequent conversations
• Tag system: 90-day inactivity → downweight; frequent hits → upweight; search hits auto-rehabilitate
• Browse, edit, batch delete, export
• Privacy mode: one-click disable of all external API calls

12. Skill System & MCP Extensions

An extensible workbench, not a closed feature set.

• Skills load AI capabilities on demand — tools only loaded when activated, saving tokens
• 40+ built-in skills/tool groups: Cards · Research · Paper · Tutor · Literature Review · Exam Analysis · Mind Map · Q-Bank · Memory · Session Manager · Office Suite · Todo · Canvas Notes · Image Generation · Web Fetch · Subagent Workspace, and more
• community skill marketplace skill marketplace: browse / verify / install community skills with risk scan
• Three-tier loading (Built-in → Global → Project-level), custom skills via SKILL.md
• MCP protocol compatible (OAuth for supported presets), connecting external tools like Arxiv, Context7
• 12 built-in provider templates (incl. Gemini), plus custom multi-protocol endpoints (Anthropic / Grok / Ollama, etc.)
• Adapted for Gemini 3, GPT-5.5 / GPT-5.2 Pro, GLM-5, DeepSeek V4, Seed 2.0, Kimi K2.5, and more

13. Local-First & Data Governance

Your learning data stays under your control.

• All data stored locally (SQLite + LanceDB + Blob)
• Full ZIP backup & recovery (incremental backup retired), data import/export
• AES-256-GCM encryption for sensitive data, dual-slot A/B switching
• Audit logs for full traceability
• Cloud sync (experimental): backup-style sync via S3-compatible storage & WebDAV; FTP backend experimental; not real-time collaboration

Installation

Download the latest version from GitHub Releases:

iOS: local source build only via Xcode (no App Store package). See Build Configuration Guide.

Getting Started

After your first launch, try this path:

1. Import a PDF / textbook / paper
2. Start a conversation around the material
3. Generate a mind map
4. Create a question set or flashcards
5. Use translation / close reading to deepen understanding

This path best demonstrates DeepStudent’s core value: not isolated features, but a complete learning chain.

---

What Makes It Different Technically

If you’re a developer, this section is for you.

• Unified learning data layer — One material can be read, searched, structured, practiced, memorized; upper-layer apps are different views of the same data
• Local-first — Metadata (SQLite), vector indices (LanceDB), file content (Blob) all stored locally
• Skill-driven architecture — Capabilities load on demand, combined with MCP protocol and multi-search-engine integration
• End-to-end loop — Import → understand → research → structure → practice → flashcards → memory

---

Architecture Overview

DeepStudent
├── Learning Materials: PDF / DOCX / textbooks / questions / notes / mind maps / translations
├── Unified Data Layer: VFS + SQLite metadata + LanceDB vector index + Blob file storage
├── Workflow Layer: chat / research / mind map / question sets / translation / essay / memory
├── Extension Layer: Skills / MCP / multi-search engines / custom model providers
└── Interface Layer: Desktop (macOS · Windows · Linux) & Mobile (Android; iOS source-build only)

---

Tech Stack

---

Development

Prerequisites

Local Development

git clone https://github.com/helixnow/deep-student.git
cd deep-student

npm ci
npm run tauri dev

For more build and packaging info, see BUILD-CONFIG.md

---

Documentation

---

Roadmap

On the way to v1.0. Near-term focus:

• User experience & stability improvements
• Desktop & mobile UI/UX optimization
• Cloud sync & backup enhancements
• Resource full lifecycle management optimization
• Skill & workflow expansion
• More model integrations & adaptations

---

Project History

DeepStudent started as a Python demo in March 2025 and has evolved through nearly a year of continuous iteration:

---

Contributing

Help make DeepStudent better.

1. Read CONTRIBUTING.md for development workflow
2. Ensure npm run lint and type checks pass before submitting a PR
3. Bugs & suggestions via Issues

---

License

AGPL-3.0

---

Acknowledgments

DeepStudent would not be possible without these outstanding open-source projects:

Frameworks & Runtimes Tauri · React · Vite · TypeScript · Rust · Tokio

Editors & Content Rendering Milkdown · ProseMirror · CodeMirror · KaTeX · Mermaid · react-markdown

UI & Styling Tailwind CSS · Radix UI · Phosphor Icons · Framer Motion · Recharts · React Flow

Data & State LanceDB · SQLite / rusqlite · Apache Arrow · Zustand · Immer · Serde

Document Processing PDF.js · pdfium-render · docx-preview · docx-rs · umya-spreadsheet · Mustache · DOMPurify

Internationalization & Toolchain i18next · date-fns · Vitest · Playwright · ESLint · Sentry

---

Made with ❤️ for Lifelong Learners