Ciel：基于 Ollama 与 Qwen 的本地多引擎 AI 桌面助手

🏛️ System Architecture

✨ Key Features & Capabilities

🧠 1. Multi-Engine Intelligence Router

Ciel dynamically delegates tasks to specialized local models running on your hardware:

Primary Intelligence Core (qwen2.5-coder:7b): Handles reasoning, coding, document creation, and tool calling.
Deep Reasoning Engine (qwen3:latest): Engages for complex mathematical logic proofs and multi-step analyses.
Ultra-Fast Speed Engine (qwen2.5-coder:1.5b): Executes instant status inquiries and short commands in under a second.
Text-to-Image Synthesis (sdxl-turbo / Stable Diffusion): Synthesizes digital artwork directly onto canvas and desktop.

🎙️ 2. Hands-Free Voice Perception & Speech

Voice Ear: Offline wake-word detection ("Ciel", "Hey Ciel") powered by faster-whisper (small.en, int8 quantization).
Vocal Core: Smooth, natural speech responses utilizing Microsoft Edge Neural TTS (en-US-AriaNeural).

🖱️ 3. Autonomous Computer Control & Automation

Live App Automation: Controls apps and browsers (Canva, Google Docs, Notion, Sheets, Calculator).
Autonomous MS Paint Sketching: Launches MS Paint and autonomously guides the cursor to draw cats, stars, hearts, and houses in real time.
System Telemetry & Control: Real-time VRAM monitoring, memory purging (/api/models/purge), clipboard analysis, and hotkey listeners.

📁 Repository Structure

Ciel/
├── app/
│   ├── app.py                 # Core Flask application, API router & SSE server
│   ├── agent_tools.py         # 35+ autonomous tools & execution registry
│   ├── launch_desktop.py      # PyWebView native desktop window container
│   ├── launcher.cs            # Native C# launcher source
│   ├── model_router.py        # Dynamic LLM routing engine
│   ├── primitive_engine.py    # Autonomous canvas and drawing routines
│   ├── security.py            # Tool safety guardrails and permissions
│   ├── task_manager.py        # Asynchronous background task scheduler
│   ├── requirements.txt       # Python dependencies
│   ├── static/                # Cyberpunk HUD styles, icons & animations
│   └── templates/             # HUD interface template (index.html)
├── docs/
│   └── ciel_architecture.svg  # Complete system architecture diagram
├── plans/
│   └── ciel-improvement-roadmap.md   # Architectural roadmap and hardening plan
├── tests/                     # Automated test harness & security validations
├── CIEL_CAPABILITIES.md       # Detailed operations & commands reference
├── CONFIG_GUIDE.md            # Hardware & model configuration guide
├── INITIAL_PROJECT_ANALYSIS.md# Technical audit and API reference
├── INSTALL_DEPENDENCIES.bat   # One-click dependency installer
├── RUN_CIEL.bat               # One-click desktop launcher
├── RUN_CIEL_COMPANION.bat     # Compact taskbar companion launcher
├── LICENSE                    # MIT Open Source License
└── README.md                  # Project documentation

🚀 Getting Started

Prerequisites

Windows 10 / 11 (64-bit).
Python 3.10+ (ensure Add Python to PATH is checked during install).
Ollama installed and running locally.
```
ollama pull qwen2.5-coder:7b
```

1. Installation

Clone the repository:

git clone https://github.com/N1412N/Ciel.git
cd Ciel

Install required Python dependencies:

Run INSTALL_DEPENDENCIES.bat
or manually:

pip install -r app/requirements.txt

2. Launching Ciel

Simply double-click RUN_CIEL.bat (or Ciel.exe).

The desktop HUD interface will open maximized with hands-free voice detection active!

🧪 Running Tests

The test suite covers backend security, model routing, safety guardrails, and APIs:

pytest tests/

🙏 Acknowledgements & Open-Source Credits

Manas Ciel is built upon and inspired by incredible open-source projects, models, and libraries. Sincere gratitude and full credit to the following creators and repositories:

🧠 Local AI & Intelligence Engines

Ollama (ollama/ollama) — High-performance local LLM execution engine and inference server.
Qwen 2.5 & Qwen (QwenLM/Qwen2.5) — Primary intelligence, code reasoning, and logic models (qwen2.5-coder:7b, qwen2.5-coder:1.5b, qwen3) by Alibaba Cloud.
Diffusers (huggingface/diffusers) — Pretrained diffusion pipelines for text-to-image artwork synthesis (SDXL-Turbo, Stable Diffusion) by Hugging Face & Stability AI.
Real-ESRGAN (xinntao/Real-ESRGAN) — Deep learning image super-resolution and anime restoration algorithms by Xintao Wang.

🎙️ Voice Perception & Audio Synthesis

Faster-Whisper (SYSTRAN/faster-whisper) — Fast, quantized OpenAI Whisper implementation using CTranslate2 by SYSTRAN.
CTranslate2 (OpenNMT/CTranslate2) — Ultra-fast inference engine for Transformer models by OpenNMT.
OpenAI Whisper (openai/whisper) — Foundational automatic speech recognition (ASR) architecture by OpenAI.
Edge-TTS (rany2/edge-tts) — Python interface for Microsoft Edge's neural text-to-speech voice synthesis by rany2.
SpeechRecognition (Uberi/speech_recognition) — Library for microphone audio capture and processing by Anthony Zhang (Uberi).

🖱️ Desktop GUI & Computer Automation

PyWebView (r0x0d/pywebview) — Lightweight, native cross-platform desktop GUI window container by Roman Sirokov.
Flask (pallets/flask) — Python WSGI micro web framework and SSE streaming core by Pallets.
PyAutoGUI (asweigart/pyautogui) — Cross-platform GUI automation for programmatic mouse/keyboard control by Al Sweigart.
mss (BoboTiG/python-mss) — Ultra-fast cross-platform screen capture module in pure Python by Mickaël Schoentgen.
Winotify (versa-sy/winotify) — Native Windows 10/11 toast notification engine by Versa-sy.
Send2Trash (arsenetar/send2trash) — Safe, native Windows Recycle Bin deletion module by Arsen Tarverdyan.
Keyboard (boppreh/keyboard) — Global Windows keyboard hooks and hotkey listeners by Boppreh.

🎨 Creative Tools & Core Utilities

Primitive (fogleman/primitive) — Geometric primitive art reproduction algorithm utilizing hill-climbing optimization by Michael Fogleman.
OpenJarvis (open-jarvis/OpenJarvis) — Stanford OpenJarvis cognitive extension abstractions and agent tools.
DuckDuckGo Search (deedy5/duckduckgo_search) — Live web search integration without API keys by deedy5.
BeautifulSoup4 — HTML/XML parsing for web inspection tools by Leonard Richardson.
OpenCV & Pillow — Computer vision, color analysis, and image manipulation.