-
Notifications
You must be signed in to change notification settings - Fork 0
AI and Privacy
GlucoPilot's AI features — the Companion, cross-domain Overview, pattern/insight narratives, lab-report reading, and the Visit Report — run through a provider you choose on Settings → AI. Three options, all first-class:
The simplest path — paste an API key and you're done.
- Anthropic (Claude) — key from console.anthropic.com.
-
OpenAI — key from platform.openai.com. Use a vision model (e.g.
gpt-4oorgpt-4o-mini) so lab-report reading works.
With a cloud provider, your computed summaries and any uploaded lab images are sent to that provider under their terms. This is the right choice for most people who don't have a local GPU.
Run an OpenAI-compatible server (vLLM, Ollama, LM Studio, llama.cpp) on your own machine. No health data leaves your server. Two roles:
| Role | Used for | Requirement |
|---|---|---|
| Default model | lab extraction + everyday text | must be vision-capable |
| Report model (optional) | Visit Report narrative only | text; larger = better prose |
The report model is loaded on demand and unloaded when idle (e.g. via
Ollama's keep-alive), so a big model doesn't hold GPU memory the rest of the
time. Full setup — including reaching a host model from the container and
graceful loading — is in
docs/LOCAL_MODELS.md.
- Lab-report images are stored on your server. They're sent to the AI provider only for extraction — with a local vision model, they never leave.
- Narratives (patterns, insights, report) send computed summaries (numbers), not raw records.
- Nothing is sent anywhere unless you've selected a provider and triggered a feature. With no key configured, the app still works — those AI features just stay idle.
For a privacy-first setup that still gets a great report: run a local vision model for day-to-day and lab extraction, and either a larger local model or a cloud key just for the occasional Visit Report.