First AI Model Specially trained to Control Home Assistant
AI Summary
Summary: Home LLM AI Model for Home Assistant
- Introduction
- Home LLM is the first AI model designed to control Home Assistant.
- It’s based on Microsoft’s F2 model and trained on a specific dataset for home automation interaction.
- Integrates with Lama conversation, splitting AI responses into user answers and service/entity commands.
- Operation and Limitations
- Can run on a Raspberry Pi 4, but not recommended due to slow response (~60 seconds).
- Better performance on a PC, with or without a GPU, enabling real-time control without internet.
- Requires a model trained on the project’s dataset, which is available for download.
- Installation on Home Assistant
- Install the text generation web UI as an add-on.
- Configure the interface via the web UI.
- For PC installation, download from GitHub, unzip, and edit the CMD flags file.
- Follow the installer prompts, selecting GPU or CPU usage.
- Configure the model in the text generation web UI, choosing the appropriate model file for your hardware.
- Using the AI Model
- Install the Lama conversation integration on Home Assistant.
- Restart Home Assistant and add the new integration.
- Configure voice assistants and exposed devices in Home Assistant settings.
- Test the AI with voice commands for various tasks.
- Improving AI Responses
- Edit the AI prompt to improve response accuracy.
- Use input Booleans and automations to work around limitations, like triggering scripts.
- Current Capabilities and Future Developments
- Controls multiple devices but has limitations (e.g., cannot change light colors or use scripts directly).
- Basic information shared with the model includes entity name, friendly name, and state.
- Ongoing work to enhance AI capabilities and train larger models for more complex operations.
- Support and Updates
- Project by Alex Okono, with continuous feature additions.
- Encouragement to support the creator and stay tuned for more updates.
For detailed instructions and support, visit the creator’s website or GitHub page.