★ Star on GitHub

Blog

Building Offline AI Search

How LUMORA combines full-text search with on-device CLIP embeddings.

Offline search is two problems glued together: structured retrieval (filenames, tags, EXIF, OCR) and meaning (does this look like a beach dog?).

LUMORA uses SQLite FTS5 for the first, and optional CLIP ViT-B/32 embeddings via ONNX for the second. Queries can blend both — text-to-image does not replace keywords; it sits beside them.

Models download only when asked, verify with checksums, and never send your photo pixels to a hosted inference API. After install, search works without the network.

Try the idea on the homepage search playground.

Download LUMORA · All posts