DocuSage

Private AI PDF chat and local RAG, built for your desktop.
Search documents locally, chat with offline GGUF models, and call the assistant from anywhere with Alt+Space.

Website   |   Download   |   Assistant mode

---

The idea

DocuSage started with a familiar research problem: the answer was already inside the document, but finding it meant breaking focus, searching across tabs, and often uploading private material to a hosted service.

It is designed as the assistant that stays beside the work instead. DocuSage starts hidden, waits in the system tray, and appears as a compact panel when you press Alt+Space. PDF parsing, embeddings, vector retrieval, local model files, and session data remain on your computer. You can use a local GGUF model for a fully offline workflow or deliberately choose a remote provider for final answer generation.

Current dashboard

The dashboard brings document ingestion, general chat, RAG mode, model selection, chat history, folders, index status, and assistant controls into one desktop workspace.

What DocuSage can do

• Ask grounded questions across private PDF documents.
• Parse, chunk, embed, and retrieve document context locally.
• Run offline chat with local GGUF models through mistral.rs.
• Store semantic document vectors in a local LanceDB index.
• Launch hidden and toggle the assistant globally with Alt+Space.
• Switch between compact, medium, and full window modes.
• Hide to the system tray without losing drafts, sessions, documents, or model state.
• Configure Local, Gemini, OpenAI, Anthropic Claude, OpenRouter, Ollama, LM Studio, and custom OpenAI-compatible providers.
• Keep API credentials in the platform keyring where available.
• Download, connect, disconnect, restart, and remove local models in the app.
• Tune RAG chunk size, overlap, top-k retrieval, and source context display.
• Render streamed Markdown and line-numbered code blocks with copy controls.
• Persist multiple chats, folders, drafts, selected documents, provider choice, sidebar state, and window mode.

Architecture

flowchart LR
    UI[React 19 desktop UI] --> CMD[Tauri commands]
    CMD --> RUST[Rust backend]

    PDF[Private PDF] --> PARSE[pdf-extract]
    PARSE --> CHUNK[Local chunking]
    CHUNK --> EMBED[fastembed embeddings]
    EMBED --> DB[(LanceDB index)]
    QUESTION[User question] --> SEARCH[Semantic retrieval]
    DB --> SEARCH
    SEARCH --> CONTEXT[Selected local context]

    CONTEXT --> LOCAL[mistral.rs + GGUF]
    CONTEXT --> REMOTE[Chosen cloud or remote provider]
    LOCAL --> STREAM[Streamed grounded answer]
    REMOTE --> STREAM
    STREAM --> UI

    TRAY[System tray] --> LIFE[Assistant lifecycle]
    HOTKEY[Alt+Space] --> LIFE
    LIFE --> UI

Loading

Component map

Layer	Technology	Responsibility
Desktop shell	Tauri v2	Native window, tray, global shortcut, single instance, hidden startup, and installer bundle.
Frontend	React 19, TypeScript, Vite	Chat, document controls, settings, provider management, model catalog, RAG controls, and marketing site.
Backend	Rust 2021 + Tokio	Tauri commands, streaming, local model lifecycle, persistence, provider requests, and assistant behavior.
Local inference	mistral.rs	On-device GGUF model execution for offline AI chat.
PDF pipeline	pdf-extract	Local text extraction before chunking and indexing.
Embeddings	fastembed + BAAI/bge-small-en-v1.5	Local vector generation for document chunks and questions.
Vector database	LanceDB	Persistent semantic search across locally ingested PDFs.
Provider layer	reqwest + platform keyring	Optional cloud/remote inference and protected credential storage.
Assistant lifecycle	Tauri tray and global shortcut plugins	Hidden launch, close-to-tray, window modes, focus, and explicit quit.

How document Q&A works

1. DocuSage extracts text from the selected PDF on-device.
2. The text is split using the configured chunk size and overlap.
3. fastembed creates local vectors for each chunk.
4. LanceDB stores the vectors in a persistent local index.
5. The question is embedded and matched against relevant document chunks.
6. DocuSage sends the selected context to the local GGUF model or the provider you explicitly selected.
7. The grounded answer streams back into the desktop conversation.

Privacy boundary

Local mode keeps the full pipeline on your computer: documents, extracted text, embeddings, retrieval, prompts, generation, model files, and history.

Provider mode still performs PDF ingestion and semantic retrieval locally. Only the active prompt, conversation context, and selected document excerpts required for synthesis are sent to the configured provider. The full PDF and LanceDB index are not uploaded by the retrieval pipeline.

Desktop lifecycle

Event	Behavior
App launch	Starts hidden, registers the global shortcut, creates the tray menu, and keeps backend services ready.
Alt+Space while hidden	Opens and focuses the compact assistant panel.
Alt+Space while visible	Hides the assistant while preserving state.
Escape or close	Hides the window to the tray instead of quitting.
Expand control	Cycles compact -> medium -> full -> compact.
Tray Quit	Explicitly stops the app and background services.

Keyboard shortcuts

Shortcut	Action
Alt+Space	Toggle the assistant globally
Ctrl+Space	Fallback global toggle
Escape	Hide the assistant
Ctrl/Cmd + ,	Open settings
Ctrl/Cmd + N	Start a new chat
Enter	Send a message
Shift+Enter	Add a new line
Ctrl+Shift+C	Clear the current chat

Installation

Download the Windows app

Get the newest installer from GitHub Releases. After installation, open DocuSage once and press Alt+Space or use the tray icon to show the assistant.

Build from source

Required dependencies

• Windows 10 or Windows 11
• Node.js 18+ and npm
• Rust stable toolchain
• Microsoft C++ Build Tools with Desktop development with C++
• WebView2 Runtime on systems where it is not already installed
• Git
• Enough disk space for local GGUF model files; individual models commonly require 1.5 GB or more

git clone https://github.com/Waqar-743/DocuSage.git
cd DocuSage/DocuSage
npm install
npm run tauri dev

Build a release installer:

cd DocuSage
npm run tauri build

Local model setup

Use the in-app model catalog or place a .gguf file in the default model directory:

Platform	Default path
Windows	Documents\DocuSage\models\
macOS	~/Documents/DocuSage/models/
Linux	~/Documents/DocuSage/models/

Override model discovery with DocuSage/src-tauri/.env:

MODEL_PATH=D:\DocuSage\models
USE_GPU=0

Provider setup

Open Settings -> AI Providers to configure and test one of the supported profiles:

• Local
• Google Gemini
• OpenAI
• Anthropic Claude
• OpenRouter
• Ollama Local or Remote
• LM Studio Local or Remote
• Custom OpenAI-compatible API

Each profile can define its model, base URL, timeout, temperature, organization/project metadata, enabled state, and protected API credential.

Project structure

DocuSage/
|-- .github/workflows/
|   `-- pages.yml                 # GitHub Pages deployment
|-- DocuSage/
|   |-- docs/
|   |   `-- hidden-assistant-mode.md
|   |-- public/
|   |   |-- media/               # Website and README imagery
|   |   |-- robots.txt
|   |   `-- sitemap.xml
|   |-- src/
|   |   |-- App.tsx              # Desktop application interface
|   |   |-- App.css
|   |   |-- MarketingSite.tsx    # Public website
|   |   |-- MarketingSite.css
|   |   |-- main.tsx             # Web/Tauri entry selection
|   |   `-- lib/api.ts           # Typed Tauri command bridge
|   |-- src-tauri/
|   |   |-- capabilities/
|   |   |-- icons/
|   |   |-- src/
|   |   |   |-- assistant.rs     # Tray, shortcuts, and window lifecycle
|   |   |   |-- commands.rs      # App and model Tauri commands
|   |   |   |-- providers.rs     # Provider profiles and remote inference
|   |   |   |-- rag.rs           # PDF, embeddings, and LanceDB pipeline
|   |   |   |-- lib.rs           # Tauri setup and command registration
|   |   |   `-- main.rs
|   |   |-- Cargo.toml
|   |   `-- tauri.conf.json
|   |-- package.json
|   |-- vite.config.ts
|   `-- index.html
|-- README.md
`-- LICENSE

Main dependencies

Frontend

Package	Purpose
React 19	Desktop and marketing interfaces
TypeScript 5.8	Typed frontend code
Vite 7	Development server and production bundling
Tailwind CSS 4	Utility styling used by the desktop application
Lucide React	Interface icon set
Tauri JavaScript APIs	Native dialogs, openers, and desktop bridge

Rust backend

Crate	Purpose
tauri 2	Native desktop runtime and tray support
mistralrs 0.7	GGUF local model inference
lancedb 0.26	Local vector storage and search
fastembed 4	Local text embeddings
pdf-extract 0.10	PDF text extraction
tokio 1	Async runtime
reqwest 0.12	Cloud and remote provider requests
keyring 3	Platform-backed credential storage
Tauri plugins	Dialog, file system, opener, global shortcut, and single-instance behavior

Environment variables

Variable	Default	Purpose
MODEL_PATH	Auto-detected	Directory containing local .gguf files
DOCUSAGE_MODEL_PATH	Unset	Alternate local model directory override
USE_GPU	0	Enables GPU acceleration where the runtime supports it
CHAT_TEMPLATE	Auto-detected	Overrides the selected model chat template
TOK_MODEL_ID	Auto-detected	Overrides tokenizer model discovery

Useful commands

cd DocuSage

npm install          # Install frontend dependencies
npm run dev          # Run the website/browser interface
npm run build        # Type-check and build the web bundle
npm run tauri dev    # Run the desktop app in development
npm run tauri build  # Build the desktop installer

License

DocuSage is released under the MIT License.