🐙 RAGit

On-device RAG desktop app. Pick a local model, see if your machine can run it, download it, and chat with your files — fully offline, fully yours.

RAGit is an open-source desktop application built with Tauri v2 + React that runs LLMs locally via llama.cpp. It lets you:

• 🔎 Choose a model and instantly see whether your hardware can handle it (🟢 fits / 🟡 tight / 🔴 too big) plus estimated and measured tokens/sec.
• ⬇️ Download GGUF models from Hugging Face with live progress + SHA256 verification.
• 💬 Chat with an OpenAI-compatible local server (llama-server).
• 📚 Build a library of files (text, documents, images, video, audio) and ask questions over them via Retrieval-Augmented Generation.
• 🗂️ Index out-of-band with 5 depth levels, pause / resume / cancel, and persisted per-file metadata (hash, level, progress).
• 👥 Team Mode (optional): expose the app on your LAN with authentication and role-based access control (admin / editor / viewer).

Dual Mode

Mode	Bind	Use case
Local (default)	127.0.0.1	Single user, max privacy
Team/Server	0.0.0.0:PORT	Company hub, employees via browser

⚠️ Team Mode binds 0.0.0.0 by design (LAN share). Tokens are HMAC-signed and passwords are Argon2-hashed, but there is no TLS — only enable it on a trusted network.

Tech Stack

• Tauri v2 desktop shell (Rust)
• React + Vite + Tailwind frontend
• llama.cpp (llama-server) sidecar for chat / embeddings / vision (auto-downloaded per-platform: Windows, Linux, macOS)
• Axum HTTP server for Team Mode
• SQLite (rusqlite, bundled) for users, sessions, libraries, files, and metadata
• Zvec (zvec-rust) for vector storage and hybrid search (HNSW + FTS + scalar filters)
• In-Rust cosine similarity fallback (no external vector extension required)

Getting Started

Prerequisites

• Node.js 18+ and npm
• Rust (stable) with the MSVC toolchain on Windows (Linux/macOS: the platform C toolchain)
• (Optional, for media) ffmpeg, whisper.cpp, pdftotext, tesseract

Build & Run

# install JS dependencies
npm install

# run in dev mode
npm run tauri dev

# build the installer (nsis / msi on Windows, etc.)
npm run tauri build

The first time you launch a model, RAGit downloads the llama.cpp llama-server binary for your platform automatically.

Using RAGit

1. Open Models, pick a quant that fits your RAM/VRAM, Download then Launch.
2. Open Indexing, choose a depth level (L1 raw → L5 rerank), click Index Folder. Watch progress; you can Pause / Resume / Cancel anytime.
3. Open Chat, enable RAG mode (and optionally Rerank), and ask questions — answers cite the source files.
4. (Optional) Team tab → Start Team Server to share the library over your LAN. The first account created becomes admin.

Project Layout

src-tauri/src/
  hardware.rs   detect VRAM/RAM/CPU/GPU
  catalog.rs    model catalog + device fitness
  download.rs   HF download + SHA256 + progress
  engine.rs     llama.cpp sidecar lifecycle (+ embed engine)
  chat.rs       streaming chat completions
  team.rs       Axum server + auth + RBAC
  rag/          parse · embed · store · indexer · media · vision · export
catalog/models.yaml   curated model catalog

See PLAN.md for the full architecture and roadmap.

Contributing

See CONTRIBUTING.md. Issues and PRs welcome!

License

MIT