AI suggestions
AI recognition is on by default — no setup needed. New photos get suggested names, descriptions, and tags in the background. It only ever suggests: nothing lands in an item’s real fields without your confirmation.
What it does
Section titled “What it does”- Suggests names, descriptions, and tags for new photos
- Reads text in photos (labels, serials, receipts)
- Split pile — breaks a photo of many things into separate items you approve
- Visual similarity — CLIP embeddings power search-by-photo (in-process, local)
Providers
Section titled “Providers”Built-in (local) — bundled Florence-2, runs offline after a one-time ~400 MB download into /data on first use. Nothing leaves your network.
Remote — any OpenAI-compatible /chat/completions endpoint with a vision model: Ollama, LM Studio, or a hosted provider. This is an upgrade path, not a requirement — handy if you already run a stronger model. Photos a remote provider processes do leave your server, but private items are never sent.
Configure in Settings → AI suggestions. The compose file already maps host.docker.internal to your Docker host, so a service running outside the container (e.g. Ollama at http://host.docker.internal:11434/v1) is reachable at that address.
Turning it off
Section titled “Turning it off”Prefer no AI at all? Turn it off in Settings → AI suggestions, or set STACHE_NO_ML=1 to disable ML entirely on constrained hosts — everything else keeps working.
Resource notes
Section titled “Resource notes”- CLIP (photo similarity) loads lazily on first embed — ~200 MB idle becomes ~450–550 MB warm.
- AI is non-blocking: filing works with no models loaded, and failures land in the Review queue for a one-tap retry.