Turn raw error codes, zone names, and device state into plain-language guidance, without
sending a single byte off the machine. The AI is optional, off by default, and runs entirely
on 127.0.0.1.
The managed backend runs a local llama-server bound to 127.0.0.1.
The diagnostic context, your error codes, zone names, and device state, is sent to that
loopback address and nowhere else. No account, no cloud, no telemetry. Once the model is
downloaded it works fully offline, which matters on an air-gapped commissioning bench.
The numbers and safety state come from the sensor and the protocol decoder, not the model. When something looks wrong, an "Explain this" action on the device-health card asks the local model to translate the current state into a readable explanation and next steps. The model never decides whether a zone is intruded or an output is tripped; it only puts words around what the rules already determined.
llama-server (managed) is the private default: the app downloads, launches, and supervises a local server for you, and the data stays on the loopback interface. Ollama (external server) points at an Ollama host you already run; useful if you have a shared GPU box, with the tradeoff that the diagnostic context travels to that host over your network. Pick per workstation in Settings.
From Settings, "Download AI runtime + model" pulls the llama.cpp runtime and a Qwen2.5-7B-Instruct model (about 4.7 GB total) into your user folder, one time. We host only a small manifest; the runtime comes from llama.cpp's GitHub releases and the model from Hugging Face. Every file is verified against a published SHA-256 before it is installed, so a tampered or changed-upstream download aborts rather than landing.
The managed runtime targets Windows x64 today and runs on CPU, so the first answer on a large model is not instant. The feature is opt-in: it ships off, downloads nothing until you ask, and can be disabled at any time. If you prefer your own stack, the Ollama backend lets you choose the model and the hardware.
Install the app, then enable AI diagnostics from Settings when you want it.
