Semantra Files

(Following text is A.I.-generated) Semantra 0.2 is a big rewrite. It now uses Google's EmbeddingGemma 2, which runs entirely on your machine through Apple MLX. You can search text, images, PDF pages, audio and video together, and you can search with text, an image, an audio file or your voice.

Requirements: an Apple Silicon Mac (M1 or later). For Windows, Intel Macs, or the older smaller model, use v0.1.3.

Download

• Semantra_0.2.1_aarch64.dmg: open it and drag Semantra to Applications. The app is signed and notarized. The download is large (~900 MB) because the model weights are bundled.

What's new in 0.2

• Multimodal search: one embedding space covers text, images, rendered PDF pages, audio (15 s windows) and video (frames plus soundtrack).
• Multimodal queries: search with text, an image or audio file, or your voice.
• New viewers: an image viewer with zoom and pan, and an audio/video player with a waveform, hit markers and seeking. Results show media thumbnails.
• Show filter: limit results to Text, Pages, Images, Video or Audio. Text results are excerpted and show page badges. Reader highlights are narrowed to the matching sentences.
• Indexing: 256-token chunks, batched GPU encoding, and video decoding at about 12× real time at 1080p. Failed files can be retried.
• Reliability: PDF handling is more robust. A bad page no longer fails the whole PDF. Startup errors now show an alert and are logged to ~/Library/Logs/Semantra/startup-error.log.

Fixed in 0.2.1

• 0.2.0 crashed at launch on every Mac except the one that built it, because MLX's Metal kernels weren't bundled. They are now bundled. 0.2.0 was never published, so 0.2.1 is the first 0.2 release.

Full Changelog: https://github.com/freedmand/semantra/compare/v0.1.3...v0.2.1