Language: English | 简体中文 | 繁體中文 | 日本語
A research assistant that runs on your own machine. Your data stays put.
Unlike general-purpose AI chat, Yanmo is purpose-built for researchers: deconstruct vague advisor directions, track papers across multiple sources, cross-validate formula derivations, evaluate project proposals objectively, and assist with thesis writing. Five purpose-built tools, each with its own interface — not crammed into a single chat box.
Three core principles: data stays local, formulas must not be wrong, plugin system for extensibility. Plus AI-powered dedup, research gap finder, and AI-text humanizer.
Desktop application, not a web app. Runs on Windows, Linux, and macOS.
Requires Python 3.11+, Node.js 18+, and Ollama.
pip install -e ".[dev]"
python -m backend.mainBackend runs at 127.0.0.1:8000. Then start the frontend:
cd frontend
npm install
npm run devOpen http://localhost:5173 in your browser.
Note (2026-05): The primary supported frontend is the React Web app (
frontend/). The Tauri desktop shell (frontend/src-tauri/) and the Qt frontend (frontend_qt/) are deprecated and will be removed in a future release. Maintaining three frontends is unsustainable for a small team. If you rely on the Tauri or Qt frontend, please open an issue to discuss migration to the Web frontend.
Five tabs across the top (+ chat panel on the right): Advisor Decoder, Paper Tracker, Project Reviewer, Formula Validator, Thesis Writer.
Paste your advisor's words. Auto-decomposes into executable tasks with ordering and time estimates. Each task expands to show sub-steps and suggested resources. Advisor data is classified as confidential, forced to local model only.
Set research keywords. Parallel fetches from ArXiv, Semantic Scholar, and DBLP. Papers displayed in a feed with one-click Chinese summaries. Auto-updates every hour in the background.
Triple dedup (title similarity + arxiv_id + DOI) ensures no preprint appears twice. Click "Find Ideas" to analyze the current paper list across 10+ dimensions (theory, efficiency, robustness, fairness, etc.), flagging research gaps sorted by confidence.
Submit a project name and description. Scores on three dimensions — novelty, soundness, methodology — with specific weaknesses and improvement suggestions.
Split-pane LaTeX source and rendered preview. Submit for dual-channel validation: basic checks (bracket matching, division by zero, domain) + SymPy symbolic computation. Both channels must pass for a green light. No LLM involved — pure local symbolic computation.
Generates structured outlines with estimated word counts. Supports BibTeX paste-and-parse. Bottom panel has an "AI-text humanizer" tool that detects AI writing traces and auto-cleans.
Plugins live in ~/.yanmo/plugins/. Load them via Settings → Plugin Manager. Minimum requirement: plugin.toml + plugin.py. See plugin_schema/API.md for the development guide. Supports lifecycle management, hot-load/unload, and event bus.
| Component | Description |
|---|---|
| Plugin Engine | Lifecycle management, hot-load/unload, event bus |
| Task Scheduler | Auto-fetch papers hourly |
| WebSocket | Real-time push of paper updates to frontend |
| Security Layer | 3-tier data classification + cloud operation interception + audit logs |
| Storage | SQLite (structured) + ChromaDB (vectors) + local files |
Default: all operations use local Ollama. Cloud API keys are optional — if you don't fill them in, nothing leaves your machine.
- Confidential data is forced local; cloud operations are auto-blocked
- Cloud outbound traffic has audit logs
- Data stored in
~/.yanmo/, backup anytime
| Layer | Stack |
|---|---|
| Backend | Python, FastAPI, SQLite, ChromaDB, SymPy, SciPy |
| Frontend | React 18, TypeScript, Vite, KaTeX |
| Desktop Shell | Tauri (Rust) |
| Fonts | Noto Sans SC, Noto Serif SC |
- myBlog — Author's blog, more research-related articles
MIT