Getting started with the Aubora beta

You're probably missing the drive ID. Reconnect with it (the number in your kDrive URL) in the Drive ID field. Indexing and search work without it; only fetching originals needs it.

The EU/US region split is handled automatically, so it's usually not that. Check the folders you expect are in the account you authorized, and that you've added them under your library's Content / folders — connecting an account doesn't pick folders for you.

Thumbnails are made during indexing, which needs the on-device AI models installed (download is opt-in). Install them when prompted (Settings → AI models) and let the scan run — photos fill in as they're processed.

That's the models loading on first use — the panel shows "Loading AI models…". It's once per launch; per-photo progress starts once they're warm.

No source means no provider is connected to that library yet — open its Source to pick kDrive / pCloud / Dropbox / local. No content means a provider is connected but you haven't added any folders — open Content to add them.

Aubora keeps your cloud access tokens in the macOS keychain. Choose Always Allow when prompted so it doesn't ask again.

Yes — photos become searchable, mappable, and grouped as the scan runs; you don't wait for it to finish. Pause and resume from the Library panel, and it resumes on its own after a restart.

No. Indexing and search run entirely on your Mac. The only traffic leaving your machine is to your own storage provider, to read your photos. Nothing about your library is sent to Aubora.

Search uses CLIP image–text embeddings: each photo becomes a vector once, stored locally, and your query is matched against them on your Mac. Models run through ONNX Runtime — CPU by default, with optional GPU acceleration (CoreML) on Apple Silicon (see below); the Intel build is CPU-only.

No — it reads only. It never writes back to kDrive, pCloud, Dropbox, or your folder.

The first pass reads and embeds every photo — a one-time cost (paced by kDrive's rate limit there). After that, only new or changed photos are processed. Aubora picks up where it left off and asks the source only for what's changed, matching each photo by file id and content hash: unchanged is skipped, changed is re-indexed. (Local folders compare modification time and size.)

Under ~/.aubora/libraries/default/: a SQLite database (store.db, plus its -wal / -shm companions) for the search index, on-disk caches of display thumbnails, larger previews, and face crops, and a one-time models/ folder. The vectors themselves are tiny — the space is almost all image cache.

Roughly 120 MB per 1,000 photos by default (a ~9,000-photo library ≈ 1.1 GB), mostly the previews cache. See the breakdown and reclaim space under Settings → Cache — thumbnails, previews, and face crops, each with a size and a Clear button. Turning off Fast lightbox previews drops the previews cache (about three-quarters of the total), down to ~30 MB per 1,000 photos; the trade is that opening a photo then fetches the full-resolution original — crisper, but slower and re-downloaded from the cloud.

To start over: Settings → Danger zone → Reset index (type RESET to confirm), or quit Aubora and delete ~/.aubora/libraries/default/. The next scan rebuilds it.

To embed a photo it fetches the original, then discards it — it never keeps a second copy of your library. What stays on disk is the search index (embeddings + coordinates) and a small downscaled thumbnail per photo, so the gallery and map stay fast without re-reading your source. The full-resolution original is never kept.

Yes, as an experiment — on Apple Silicon only. By default Aubora runs on the CPU (the supported mode); under Settings → Advanced → Acceleration you can turn on "Use GPU acceleration (CoreML) — experimental" to move the search model (CLIP) onto the GPU. The first time, it compiles the model for your GPU once — a few minutes, with a progress note, and you can keep working — then caches it so later launches are fast. (The Intel build is CPU-only: CoreML caching needs a newer runtime than Intel Macs have, so the option isn't offered there.)

It's not faster — wall-clock time is about the same. What changes is load: the CPU path pushes several cores hard, the GPU path drops to roughly one, so the rest of the Mac stays responsive and runs cooler while it works. Because compiling adds startup cost, Aubora only uses the GPU for large scans (default ≥2,000 photos, adjustable under the toggle) — smaller incremental scans stay on the CPU, where they're quick anyway. So it mainly helps your big initial import or a large re-index.

Only the search model moves; people and faces run on a separate engine, and your search results don't change (the GPU's 16-bit math is too small to affect ranking). A fresh install has the model it needs; a library carried over from a very early build may need its models refreshed first. If anything goes wrong — the compile fails, your chip computes incorrectly, or the model can't be cached — Aubora falls back to the CPU and tells you; and if results ever look off after you enable it, switch the toggle back off to return to the supported path.

It reads the GPS the camera saved in each photo. It also reads JSON companion files that sit next to your photos — the kind Google Takeout exports — so a location survives even when a service has stripped it from the image. Companion files in a different format? Email contact@aubora.app and we'll look at adding support.

Turning coordinates into place names is separate and off by default (Settings → Privacy, online geocoding). Switched on, Aubora looks names up from OpenStreetMap's Nominatim; the map itself — plotting photos by GPS — works fully without it.

Faces are detected and grouped into people entirely on-device. On the free tier People is browse-only — look through the groups and open a person's photos; naming, merging, and splitting people come with your beta license.

Apple Silicon (M-series) is preferred — it's where the optional GPU acceleration (CoreML / Neural Engine) keeps local indexing lightest on your machine. An Intel (x86) build for older Macs is on the way; it does everything the Apple Silicon build does — indexing, search, map, people — on the CPU, just without the GPU option. Beyond macOS, nothing's planned: Windows, Linux, and mobile aren't supported and aren't on the beta roadmap.

A beta license is time-limited but renews itself automatically while you're online. If our licensing server is unreachable, it keeps working through a 14-day grace period, so an outage on our side never locks you out. If a license genuinely expires or is revoked, you still aren't locked out — Aubora drops to the free tier, capped at 2,000 photos of your already-indexed library. Re-activate, or paste a fresh beta key, under Settings → License.

Email contact@aubora.app with what you did and what happened, which source you connected (kDrive / pCloud / Dropbox / local), and the log file — a rotating log at ~/.aubora/logs/aubora.log (it keeps the few most recent). It makes almost every connect / first-index problem diagnosable.