Multi-Stage Assist is a highly advanced, local-first (with cloud fallback) conversational agent for Home Assistant. It orchestrates multiple processing stages to provide the speed of standard NLU with the intelligence of LLMs.
| Stage | Purpose | Technology |
|---|---|---|
| Stage 0 | Fast path - instant NLU | Home Assistant built-in |
| Stage 1 | Semantic cache lookup | Cached command replay |
| Stage 2 | Local LLM intent | Ollama (qwen3:4b) |
| Stage 3 | Cloud fallback + chat | Google Gemini |
- Semantic Command Cache - Instant replay of learned commands with reranker validation
- Adaptive Learning - Remembers your custom room/device names
- Temporary Controls - "Turn on light for 10 minutes"
- Natural German Responses - Optimized for German language
- Typo Tolerance - Handles minor spelling mistakes
- Home Assistant 2024.1.0+
- Ollama with
qwen3:4b-instructandmxbai-embed-large - Google Gemini API Key
- Optional: Reranker Addon
- Add via HACS:
https://github.com/kr0ner/multistage-assist - Restart Home Assistant
- Add integration: Settings → Devices & Services → Add Integration
- Pull embedding model:
ollama pull mxbai-embed-large
Power users can add YAML settings to configuration.yaml:
multistage_assist:
# Semantic Cache tuning
reranker_threshold: 0.73
hybrid_enabled: true
hybrid_alpha: 0.7
# Low-hardware mode (cache-only, no LLM in Stage1)
skip_stage1_llm: false
# LLM behavior
llm_timeout: 30
llm_max_retries: 2
# Debugging
debug_cache_hits: false
debug_llm_prompts: falseSee Configuration Reference for all options.
| Topic | Link |
|---|---|
| Architecture | Architecture.md |
| Capabilities | Capabilities-Reference.md |
| Configuration | Configuration.md |
| Development | Development-Guide.md |
Enable debug logging:
logger:
logs:
custom_components.multistage_assist: debugMIT License - see LICENSE
Attribution Required: github.com/kr0ner/multistage-assist