for_you

backend.api.for_you

For You feed — collaborative filtering over user engagement edges.

Ports the scoring algorithm from grain.social’s foryou.ts (originally tuned by @spacecowboy17.bsky.social), adapted for plyr.fm in two ways:

1. edges are not just likes: a user “engages” with a track via a like OR a track_added_to_playlist event. both edges get weight 1.0 for v1.
2. slower time decay: audio ages slower than photo galleries, so we use a 48h half-life (grain uses 6h).

Scoring recipe per candidate track: score = (sum over paths of 1 / total_edges(coengager) ** DIVISOR_POWER) * paths ** SMOOTHING_FACTOR * 0.5 ** (age_hours / HALF_LIFE_HOURS) / popularity ** POPULARITY_PENALTY

Final output diversifies by artist (MAX_PER_ARTIST hard cap).

Functions

get_for_you_feed source

get_for_you_feed(db: Annotated[AsyncSession, Depends(get_db)], auth_session: AuthSession = Depends(require_auth), cursor: str | None = Query(None), limit: int = Query(30, ge=1, le=50)) -> ForYouResponse

Personalized feed for the authenticated user.

Uses collaborative filtering over engagement edges (likes + playlist adds). Falls back to cold-start (most-engaged last 30 days) if the user has no taste signal yet.

Cursor is an integer offset into the materialized ranked list. Scores may drift between pages — this is fine for v1; if it becomes annoying we can cache the ranked list per-user in Redis.

Classes

ForYouResponse source

Paginated For You feed.