Productizing a Music Curation System with ChatGPT
The article discusses the author's experience of using ChatGPT as a music curator and the technical decisions made to productize this pattern into an app called Acetate.
Why it matters
This article showcases how to productize a pattern using large language models like ChatGPT and highlights the technical considerations involved in building a robust and reliable AI-powered music curation system.
Key Points
- 1Using ChatGPT as a music curator by having conversations to get recommendations
- 2Challenges with the context window and lack of structure in the free-form chat approach
- 3Three technical decisions to improve the curation system: no taste profile, MusicBrainz verification, and a directed discovery feature
Details
The author had been using ChatGPT as a music curator, having conversations to get recommendations. However, two issues arose: the context window became too long, causing the model to forget earlier entries, and the free-form chat lacked structure, making it difficult to browse history and access album metadata. To address these problems, the author built an app called Acetate, which persists the feedback history in a database and sends it verbatim to the language model on every selection call. The app also includes a MusicBrainz verification chain to ensure the model doesn't propose non-existent albums, and a directed discovery feature to provide additional guidance to the model.
No comments yet
Be the first to comment