Thanks, those are very thoughtful points. Offline/local approaches make sense, especially for privacy-conscious users or those using DRM-free/physical media. A local database with tags and randomness could definitely handle many cases, and it’s something I’m exploring for future versions. Using an LLM is mainly for flexibility and to handle the variety of moods and contexts without requiring the user to manually tag everything. A specialized, local model trained only on movies/series is an interesting idea, and something I would consider as a lightweight alternative in future iterations. Really appreciate this kind of detailed, constructive feedback.