I built this project as an alternative to cloud-based media services. The system uses embedding models (CLAP for audio, SigLIP for images, Jina v3 for text) to enable semantic search across your local files. You can search for "relaxing instrumental music" or "machine learning papers" and it finds relevant content based on meaning, not just filenames.
The recommendation engine works by fine-tuning models on your ratings using PyTorch. Everything runs locally in Docker with CUDA support. The metadata system uses simple .meta text files alongside your original files, so you're never locked into a proprietary format.
Built with Flask and designed to work with existing folder structures without requiring reorganization. The trained models learn your preferences without any data leaving your machine.
Currently supports music, images, text, and video. All models download automatically on first run.
1 comment
[ 0.25 ms ] story [ 17.5 ms ] threadThe recommendation engine works by fine-tuning models on your ratings using PyTorch. Everything runs locally in Docker with CUDA support. The metadata system uses simple .meta text files alongside your original files, so you're never locked into a proprietary format.
Built with Flask and designed to work with existing folder structures without requiring reorganization. The trained models learn your preferences without any data leaving your machine.
Currently supports music, images, text, and video. All models download automatically on first run.
Github: https://github.com/volotat/Anagnorisis