So Troi is implemented in the backend with a content resolver using typesense. It’s working. But, we match between 5 and 20% of the recommendation, so we end-up with tiny playlists.
We’ve been speaking about how to mitigate this in https://community.metabrainz.org/t/sort-troi-patches-by-stability-and-use/. What seems to be the better solution is to import into fw the data that troi needs to generate the recommendation. This way we have less server load on listenbrainz side, and better recommendation in our side.
Some of this data will be useful for other things like tracks
tags (tags are in our roadmap anyway)
Some of this data will only be used by troi to generate recommendation :
- Tracks popularity
- User recommendations tracks
- User already listened tracks
- User’s feedback about tracks (liked or disliked afaik)
Question : do you have any though about this ?