Vector Home v2

Offline CPU-only smart home control

Voice/text → Router → GPT-2 Parser → Home Assistant

No cloud · No GPU · No API keys

---

Architecture

Голос / Текст
      ↓
┌─────────────┐    ┌──────────────┐
│   Router v2  │───→│   Parser v2  │
│ regex + Ollama│   │ GPT-2 124M  │
│  95 правил    │   │ 53 инструм.  │
│  100% EN/RU  │   │              │
└──────┬──────┘    └──────┬───────┘
       │                   │
       └───────┬───────────┘
               ↓
       ┌───────────────┐
       │   HA Bridge    │
       │ 53→HA mapping  │
       │ RU→EN translate│
       └──────┬───────┘
               ↓
        Home Assistant

Status

Phase	Description	Status
0	Foundation — EN SFT, validation	✅
1	Router v2 — 95 regex rules, 100% EN/RU	✅
2	Integration — FastAPI, HA bridge, WebSocket	✅
3	Voice — STT + TTS, closed loop	✅
4	Web dashboard — 8 device groups, voice input	✅
5	Real HA connection	⏳

Accuracy

Test	Result
EN commands (52 test cases)	100%
RU commands (43 test cases)	100%
Router total (95 cases)	100%
Cyrillic \b regression	✅ fixed
Rule ordering (query→set, dim→off, thermostat→AC)	✅

Supported Commands

52 tools, 2 languages, 8 domains:

Domain	Tools	EN example	RU example
💡 Light	8	turn on the lights	включи свет в гостиной
🌡️ Climate	9	set temperature to 22°C	установи 22 градуса
🪟 Covers	6	open the curtains	открой шторы
🤖 Vacuum	3	vacuum the office	пропылесось кухню
🔒 Security	7	lock the front door	запри входную дверь
🎵 Media	10	play jazz in the kitchen	включи музыку на кухне
🌿 Garden	3	start irrigation zone 1	включи полив газона
📡 Other	6	wake me up at 07:30	поставь будильник на 7 утра

Quick Start

git clone https://github.com/Osmosy/vector-home.git
cd vector-home
pip install -r requirements.txt

# Tests (130/130 ✓)
python -m pytest tests/ -v

# API server (port 8126) + web dashboard
python -m src.api

# CLI text command
python -m src.pipeline "turn on the lights in the living room"
python -m src.pipeline "включи свет в гостиной"

# Voice pipeline (requires faster-whisper)
python -m src.voice --interactive

# Live Home Assistant mode
export HA_URL="http://homeassistant.local:8123"
export HA_TOKEN="your_long_lived_token"
python -m src.api --live

Web Dashboard

Open http://localhost:8126/panel — dark theme, 8 device group tabs, voice input (Web Speech API), command history, WebSocket real-time updates.

Environment Variables

Variable	Description	Default
HA_URL	Home Assistant URL	http://homeassistant.local:8123
HA_TOKEN	Long-lived access token	(empty = dry run)
VH_PORT	API server port	8126
GPT2_REPO	Path to gpt2-tool-call	../gpt2-tool-call

Requirements

	Minimum	Recommended
CPU	Any x86_64	4+ cores
RAM	600 MB (parser only)	6 GB (with Qwen3 fallback)
Disk	1.5 GB	2 GB
GPU	Not required	—
Python	3.10+	3.12
Internet	Not required	For Ollama fallback

Project Structure

vector-home/
├── src/
│   ├── router.py        # Regex router v2, 95 rules, EN+RU
│   ├── parser.py        # GPT-2 124M, 53 tools, fuzzy match
│   ├── pipeline.py      # Full pipeline: router→parser→HA
│   ├── api.py           # FastAPI + WebSocket + /panel
│   ├── ha_bridge.py     # 53 tool→HA mapping, RU→EN
│   └── voice.py         # STT→router→parser→HA→TTS
├── static/
│   ├── index.html       # Web dashboard — 8 device groups
│   ├── style.css        # Dark theme
│   └── app.js           # WebSocket + voice input
├── data/
│   ├── tools_spec_v2.json    # 52 tool definitions
│   ├── train_dataset_v2.json # 1000 training examples
│   └── test_dataset_v2.json  # Test split
├── tests/
│   └── test_router.py   # 130 pytest tests
├── docs/
│   └── SPEC.md          # Technical specification (Russian)
├── assets/
│   └── logo-original.png
└── models/              # GPT-2 weights (git-lfs)

Testing

python -m pytest tests/test_router.py -v
# 130 passed — router accuracy, Cyrillic regressions,
# rule ordering, HA bridge mappings, parser spec

Difference from barometech/smart-home-gpt2

	Vector Home v2	smart-home-gpt2
Router	Regex+fallback, 100% EN/RU	GPT-2 124M, multi-tool 71.7%
Parser	GPT-2 124M, single-tool	GPT-2 124M, multi-tool
Tools	52	100
Voice	faster-whisper medium	faster-whisper medium
Dashboard	✅ (WebSocket)	❌
HA integration	✅ (53 mapping)	Emulator only
Tests	130 pytest	—
Russian	✅ native	✅ via translate

Documentation

• 📄 SPEC.md — Technical specification (Russian)

License

MIT — see LICENSE

Built by Osmosy · Powered by gpt2-tool-call