Getting started

The first decision in this build is to select the engine that powers your AI: the GPU. Whether you’re giving new life to an used card or upgrading to something fresh, there are a few simple things to keep in mind:

We recommend AMD Radeon RX 6000 and 7000 series GPUs for their exceptional price-to-performance ratio and dependable open-source Linux drivers. An 8GB VRAM card is great for getting started, but 16GB or more lets you explore larger, more ambitious models.

You are of course encouraged to try other GPUs from the Raspberry Pi PCIe Database, or share your results for cards not listed, to help grow the community’s knowledge.

Finally, keep the GPU’s idle/maximum power and its physical size in mind, as these will influence your choice of PSU and case.

With your GPU selected, it’s time to house your build in a case that suits both style and substance. If you already have an ATX case, it can easily accommodate a Mini-ITX board. However, if you’re eyeing a new Mini-ITX case, keep these tips in mind:

• GPU fit: Check that your case can handle the size of your GPU, both in length and height. Many modern GPUs are over 25cm long and take up two or more slots, which may exceed the capacity of some Mini-ITX cases.

• PSU type: Determine whether the case requires an ATX or the more compact SFX power supply.

• Cooling options: Look for the number of case fans included or supported. The Sentinel Core can power up to two case fans, so you can keep your setup cool and efficient.

The last piece of the puzzle is your power supply unit (PSU), that suits your GPU and case:

• Case compatibility: Check whether your case requires an ATX or SFX PSU.

• Wattage: Make sure the PSU's 12V output comfortably exceeds your GPU’s maximum power draw. (The Pi’s 5V usage is negligible in comparison.)

• Connections: Confirm there are enough 6+2 PCIe power connectors for your GPU—some SFX models may fall short.

Other features you might want to consider:

• Detachable cables: A modular PSU allows for cleaner cable management.

• Efficiency ratings: 80 Plus or higher certification often means better power efficiency.

• Low idle power: An efficient PSU reduces idle power draw, saving energy when the system isn’t under load.

• Silent operation: Look for 0 RPM fan control to keep things whisper-quiet.

Once your Sentinel Core is up and running, you can customize and enhance it with a few optional extras to suit your needs:
Storage: Planning to store large media libraries or AI models? Consider adding a USB 3 hard drive.

Connectivity: Expand your setup’s capabilities with Raspberry Pi HATs or a ZigBee dongle for seamless smart home integration.

Silent fans: Keep things cool and quiet by upgrading your case fans to ultra-silent models for a more peaceful environment.

Start by going to settings -> Add-ons -> Add-on store. From there install:

• Piper - text to speech

• Get HACS - the installer for the The Home Assistant Community Store

• Llama.cpp - LLM runner

• Whisper.cpp - speech recognition

• Sanctuary Updater - to update our GPU enabled operating system fork

Our fork already comes configured with the HACS add-ons repository and our very own LLM add-on repository.

Start Llama.cpp, Whisper.cpp and Piper, and make sure "start on boot" is checked.

If you want to check and install operating system updates, start Sanctuary Updater.

Start Get HACS and then switch to the logs tab and follow the instructions, which will prompt you to restart Home Assistant.

Now head over to Settings -> Devices & Services -> Add Integration, search for HACS, and install it.

Go to the new HACS icon in the side bar, search for Extended OpenAI Conversation, and install it. Don't worry, we'll change the settings so it connects to llama.cpp instead of OpenAI.

Now head back to Settings -> Devices & Services. Either Wyoming has detected Piper and Whisper.cpp or you need to add them yourself:

• If they are detected, just hit the "Add" button.

• Otherwise, select Wyoming Protocol and from there use "Add service" button to configure the integrations we'll need. Just enter the host and port, and hit "submit":

  • Piper, host: localhost, port: 10200

  • Whisper.cpp, host: localhost, port: 10303

Now configure Extended OpenAI Conversation: Settings > Devices & Services > Hit the "Add intergration" button. Add "Extended OpenAI Conversation" with these settings:

• Name: llama.cpp

• API key: dummy

• Base URL: http://localhost:8765 (so it uses llama.cpp locally on the Sentinel Core)

• Version: leave blank

• Organization: leave blank

• Skip authentication: no

In the configuration (Settings > Devices & Services > Extended OpenAI Conversation > ⚙️), change:

• prompt template: customize the personality as you like

• chat model: a HuggingFace repository name like NousResearch/Hermes-3-Llama-3.1-8B-GGUF

• Maximum tokens to return in response: 300 (increase if you get an error about response length)

• tool use: enable

This is where it all comes together. In Settings -> Voice Assistants select the existing "Home Assistant" and configure the following items:

• Conversation Agent: llama.cpp

• Speech-to-text: whisper.cpp

• Text-to-speech: piper

Then head over to the "Expose" app and select the entities you want you AI to be able to see and control.

That's it! You can now chat with you AI from the assist menu in the top right of the dashboard. For using your microphone it is best to use the mobile app.

From here you can explore different models, different voices, and pair microphones like the Home Assistant Voice, FutureProofHomes Satelite1, or Wyoming Satelite.