Skip to content
- The MLX backend uses an unreleased
mlx-vlm branch (pinned commit) and community-converted mlx-community
weights. Vectors match the original model (cosine 0.9998), but this dependency needs watching.
- Memory (measured on M4): idle daemon ~490 MB, plus ~340 MB for the photo query model while the panel is in use
(released after 5 minutes); the indexing process up to ~2 GB for text, ~3.3 GB for recordings and ~4 GB for photos,
then it exits.
- Photo search was measured on 1,000 Flickr30k photos and an 80-photo test folder; real libraries (many similar
photos, blurred or strongly skewed phone shots of documents) may behave differently.
- Audio search was evaluated on synthetic speech; real recordings (noise, several speakers) may do worse.
Transcripts cover the first 4 hours of a recording; audio embeddings cover all of it.
- Large files (> 64 MB, mostly videos) are recognised by size and content samples, not a full hash.
- CSV/TSV and large number-heavy files are indexed by their head only (when not ignored).
- Polish inflection in keyword search is approximated by stem prefixes, not a real stemmer.
- Full list of open questions: docs/DECISIONS.md.
