DARIA

A local AI companion with memory, voice, plugins, and a desktop-like web interface

🇷🇺 Русский  |  🇬🇧 English

What is this • Features • Quick Start • Architecture • Configuration • Plugins • Contributors

---

💡 What is this

DARIA (Desktop AI Reactive Intelligent Assistant) is a local AI companion built to run on your own machine with Ollama. It is designed to feel like a persistent desktop assistant rather than a generic chat window.

Out of the box, DARIA ships with Dasha: a warm Russian-speaking persona with her own tone, mood, preferences, and long-term memory.

It combines:

• 🧠 Long-term memory that keeps track of facts, preferences, and prior conversations
• 💬 Emotion-aware dialogue with mood shifts, emotion classification, and softer response shaping
• 🖥️ Desktop-style web UI with windows, widgets, and notifications in the browser
• 🎙️ Voice support with speech-to-text and text-to-speech backends
• 📚 RAG over local files, URLs, and RSS sources
• 🔌 Plugins for extending behavior without touching the core app
• 🖱️ Optional desktop control for screenshot capture and system interaction
• 🔧 CLI mode for a lighter workflow without the web UI

In short, DARIA is a customizable local companion project: part assistant, part character system, part personal AI sandbox.

---

✨ Features

Feature	Description
Local LLM	Ollama-powered, fully offline. Supports any model (gemma, llama, mistral…)
Emotional system	17 moods, ML emotion classifier (scikit-learn), natural transitions
Long-term memory	SQLite + vector search. Remembers facts, preferences, past events
Memory Gatekeeper	LLM-based gate — retrieves memories only when relevant
Thinking pipeline	Evaluator → Dasha → Checker for higher-quality responses
Voice	STT via faster-whisper, TTS via QwenTTS/Silero, optional voice-clone integration
RAG	Files (.md/.txt/.pdf), URLs, RSS. Sentence-transformers + TF-IDF + BM25
Desktop control	Screenshots + mouse/keyboard via mss/pyautogui/xdotool (permission-gated)
Plugin system	Lifecycle hooks, isolated venvs, GitHub catalog, ZIP installs
MCP connectors	SQLite / PostgreSQL via Model Context Protocol
Multi-user mode	Per-user memory isolation for shared deployments
Streaming	SSE streaming for web + CLI
Mobile UI	Responsive layout for phones
SSH tunnels	Cloudflare / localhost.run / serveo / VPS reverse tunneling

---

🚀 Quick Start

Requirements: Python 3.10+, Ollama running locally

# 1. Clone
git clone https://github.com/your-org/daria.git
cd daria

# 2. Install
python install.py

# 3. Pull a model
ollama pull gemma3:4b

# 4. Run
python run.py

Open http://localhost:7777 in your browser.

CLI mode (no browser)

python cli.py
python cli.py --model gemma3:4b
echo "Привет!" | python cli.py   # pipe mode

Dependency sync (after updating)

python run.py --update        # interactive
python run.py --update-yes    # non-interactive / CI

---

🏗️ Architecture

daria/
├── core/
│   ├── brain.py              # DariaBrain — central response engine
│   ├── memory.py             # WorkingMemory, LongTermMemory, MemoryGatekeeper
│   ├── llm.py                # Ollama provider (streaming, vision, embeddings)
│   ├── config.py             # DariaConfig — YAML + env loading
│   ├── emotion_classifier.py # scikit-learn ML emotion classifier (14 classes)
│   ├── desktop.py            # Physical desktop capture & control
│   ├── plugins.py            # Plugin loader, lifecycle, venvs
│   └── debug.py              # Per-module debug (DARIA_DEBUG_*)
├── web/                      # Flask server, SSE streaming, REST API
├── voice/                    # STT (faster-whisper), TTS (QwenTTS/Silero)
├── rag/                      # KnowledgeStore, Ingester, Retriever
├── mcp/                      # Model Context Protocol connectors
├── plugins/                  # Built-in plugins (tamagotchi, telegram-bot, …)
├── config/
│   ├── config.yaml           # Main settings
│   ├── persona.yaml          # Dasha's system prompt template
│   ├── character_memory.yaml # Dasha's self-knowledge
│   └── prompts.yaml          # User-editable prompt variables (hot-reloaded)
├── cli.py                    # CLI REPL
├── run.py                    # Unified entry point
└── tests/                    # Regression suite (pytest)

---

⚙️ Configuration

config/config.yaml is safe to publish as a template, but keep real secrets and deployment-specific values out of git.

Public-safe example:

web:
  host: "127.0.0.1"
  port: 7777

llm:
  provider: "ollama"
  model: "llama3.2-vision:11b"

voice:
  stt_enabled: false
  tts_enabled: false

rag:
  enabled: false

server:
  enabled: false
  admin_username: "admin"
  admin_password: ""
  secret_key: ""

tunnel:
  enabled: false
  backend: "cloudflare"

Use environment variables or local-only config for secrets like DARIA_ADMIN_PASSWORD, API keys, and tunnel credentials.

config/prompts.yaml — editable without touching code, hot-reloaded:

extra_rules: ""            # Additional behavior rules
forbidden_topics: []       # Topics Dasha won't discuss
response_style: warm       # warm | neutral | professional
use_emoji: false
prompt_prefix: ""          # Prepended to every system prompt
user_context_hint: ""      # Context about the user

Debug modes

DARIA_DEBUG=1 python run.py              # all modules
DARIA_DEBUG_BRAIN=1 python run.py        # brain only
DARIA_DEBUG_MEMORY=1 python run.py       # memory only
DARIA_DEBUG_FALLBACK=1 python run.py     # fallback events

---

🔌 Plugins

Plugin	Description
tamagotchi	Virtual pet with 5 needs + XP system
telegram-bot	Telegram interface for DARIA
voice-call	Voice call interface

Install from catalog or ZIP:

# In-app: Settings → Plugins → Catalog

Plugin API hooks: on_load, on_chat_message, on_chat_response

---

📄 Docs

• Changelog (EN) — full version history
• Changelog (RU) — русская версия истории изменений
• tests/ — regression test suite (pytest tests/)

---

👥 Contributors

🤍 Thanks

Special thanks to @qtbn5jsrk9-hash for giving Dasha her voice and helping shape her personality, profile, and instructions.

---

Made with care. Runs offline. Remembers you. 🌿

Читать на русском →