Building a Playlist Generator with Sentence Transformers

Nima Boscarino wanted to build something fun with Sentence Transformers, and he ended up with a playlist generator that takes a text prompt like "running wild and free" and spits back nine songs whose lyrics actually match the vibe. The write-up on the Hugging Face blog walks through exactly how he did it, and the interesting part isn't the idea itself but the plumbing underneath.

He picked a pretrained model called msmarco-MiniLM-L-6-v3, originally trained on Bing search queries, simply because it's good enough and he didn't need to chase a leaderboard score. That model caps out at 512 word pieces per input, which turned out to be too short for a full song. His fix was simple: chop songs into verses first, then embed each verse separately. Once he did that, the semantic search results got noticeably better.

The search side works by encoding the user's prompt into the same embedding space, then running util.semantic_search against a pickle file of pre-computed verse embeddings. Because multiple verses often belong to the same song, he pulls the top 20 matches instead of just 9, which reliably leaves him with at least nine distinct tracks after removing duplicates. He stored those embeddings, plus song and lyric mappings, as datasets on the Hugging Face Hub, though the actual lyrics stay private for licensing reasons.

The interface runs on Gradio's Blocks API, which let him build a two-stage flow: type a prompt and hit generate, then pick from a Radio list of song names to reveal the lyrics that got matched. Nothing here required custom model training or JavaScript, just Python functions returning component updates. He's already got a wishlist for version two — Spotify integration, highlighting the exact matched verse, maybe a visualization of the embedding space — but even the current version is a solid, hackable demo of what pretrained sentence embeddings can do without much fuss.