Confidant - Offline Mental Health Companion

Confidant is an open-source, privacy-first AI assistant that provides offline mental health support as a supportive companion. All processing happens locally with no network access for queries.

Note: This project has evolved from an original Python-based Raspberry Pi implementation, then a web app, to the current desktop app (Tauri 2.0). See archive/README.md for information about previous implementations.

Features

• Fully Offline: All AI processing happens client-side, no backend required
• Privacy-First: All processing is local; no data is sent to external servers
• Mental Health Companion: Supports gratitude, mindfulness, mood, stress, anxiety, and depression—with RAG over your knowledge base. Not a substitute for therapy or professional care.
• Local LLM: Runs optimized models locally (Llama-3.2-3B default, Mistral-7B option)
• Streaming responses: Chat replies stream token-by-token for a responsive experience
• RAG System: Retrieval Augmented Generation with ChromaDB and sentence-transformers
• Open Source: MIT license; fully auditable codebase

AI Models

Confidant supports two optimized LLM models for mental health conversations:

Standard Model (Default)

• Llama-3.2-3B-Instruct (~2.5GB)
• Why: Smaller size for resource-constrained systems while maintaining good quality
• Best for: Most mental health conversations; fast and efficient
• Performance: Good quality with lower resource requirements

Enhanced Model (Optional)

• Mistral-7B-Instruct v0.2 (~4.4GB)
• Why: Strong reasoning and contextual understanding
• Best for: Users who want deeper, more nuanced conversations
• Performance: Strong reasoning capabilities, optimized for complex discussions

Both models use GGUF quantization and run completely offline via llama.cpp. The Standard Model is the default for most users; the Enhanced Model is recommended for more complex discussions.

Architecture

This repository contains:

• desktop/ – Desktop app (Tauri 2.0 + React + TypeScript + Rust), primary application
• landing/ – Marketing and download landing page (Next.js, shadcn/ui, Tailark). Built as a static site and deployed to docs/ for GitHub Pages. The deploy workflow (.github/workflows/deploy-landing.yml) builds landing/ and copies output into docs/ on pushes that touch landing/.
• scripts/ – Python scripts for building knowledge bases and downloading models
• docs/ – Design and migration documentation; also the GitHub Pages root (serves the built landing site)
• archive/ – Previous implementations (Python, web) and planning docs

The desktop app uses a Rust backend for Tauri and file operations, and calls Python (llama-cpp-python, ChromaDB, sentence-transformers) via subprocess for LLM inference, embeddings, and vector search. Chat responses stream token-by-token from the LLM to the UI.

Quick Start

Desktop App (Active Development)

Prerequisites:

• Rust (install from https://rustup.rs/)
• Node.js 18+ and npm

Development:

cd desktop
npm install
npm run dev

See desktop/SETUP_INSTRUCTIONS.md for detailed setup and desktop/PYTHON_SETUP.md for Python dependencies (LLM, ChromaDB, embeddings).

Download (Beta)

Pre-built installers for macOS and Windows are available via GitHub Releases. The project includes a landing page (Next.js, in landing/) that is built and served from docs/ via GitHub Pages. Enable Pages with source Branch: main, Folder: /docs to serve the site.

Contributing

We welcome contributions. Please read CONTRIBUTING.md for development setup and guidelines. Use the issue and pull request templates in .github/ when opening issues or PRs.

License

MIT License - see LICENSE file for details.