feat: recommend from Deezer and YouTube Music radios
CI / chart (push) Successful in 2s
CI / test (push) Successful in 1m57s
CI / chart-push (push) Successful in 5s
CI / image (push) Successful in 3m39s

- Deezer: /artist/{id}/radio for collection artists, affinity 2.0
- YouTube Music: RDAMVM watch playlist via yt-dlp, affinity 2.0
- seed radios from the 2 newest songs of the mood's own catalog
- dedupe candidates by recording, not only by external id
- clean raw YouTube titles and Topic suffixes when indexing
- tests, smoke marker for the RDAMVM path, docs
This commit is contained in:
TapTap committed 2026-11-17 15:22:05 +00:00
1 parent 62bcefad20
commit 26c41cc7d9
12 files changed
+296 -44

No files matched your search

+17 -9
View File
@@ -260,13 +260,18 @@ States: `new` (hidden candidate pool), `downloading`, `ready`, `failed`,
1. **Text queries**: `query`, then up to 3 collection artists ordered by how
many songs the mood has, then up to 2 genres, capped at 4 queries.
2. **Catalog similarity graph** (`MusicCatalog.discover`): up to 5 collection
artists -> their top tracks; up to 2 of those artists' related artists ->
their top tracks; up to 2 collection albums -> album deep cuts. Deezer
implements this; YouTube Music has no such API and returns nothing.
artists -> their **radio** (`/artist/{id}/radio`, the platform's own mix)
and top tracks; up to 2 of those artists' related artists -> their top
tracks; up to 2 collection albums -> album deep cuts. YouTube Music runs
the `RDAMVM<videoId>` watch-playlist radio for the 2 newest collection
songs, which is the same queue its player builds (`yt-dlp
--flat-playlist`), so it also returns real recommendations.
- Scoring: exact collection artist `3.0 + 0.5` per extra song (capped `+1.5`),
partial artist `1.5`, album already in the collection `+1.0`, catalog
affinity (related artist `1.5`, album deep cut `1.0`, artist top `0.5`),
mood words in the title `+1.0`, over 8 minutes `-1.0`.
affinity (platform radio `2.0`, related artist `1.5`, album deep cut `1.0`,
artist top `0.5`), mood words in the title `+1.0`, over 8 minutes `-1.0`.
- A recording is deduplicated by `(artist, title)` as well as by external id, so
two uploads of a song already in the mood do not both enter the pool.
- Stores up to a bounded pool of `new` candidates (60). Excludes external ids
already in `mood_songs`, `added` or `rejected`.
- `topUp(mood, force)`:
@@ -322,14 +327,17 @@ What this server reproduces without telemetry:
| Signal | Here | There |
|---|---|---|
| Taste profile | whole mood collection, artists weighted by song count + albums | per-user history and playlists |
| Similarity | catalog graph (related artists, artist top tracks, album deep cuts) + text queries | learned embeddings + co-occurrence |
| Similarity | the platform's own radio (Deezer artist radio, YouTube `RDAMVM`) + catalog graph + text queries | learned embeddings + co-occurrence |
| Feedback | add / not interested (permanent) | skips, saves, play counts, survey models |
| Context | mood hint, genres, refresh window | session, time of day, device, activity |
| Ranking | weighted score + per-artist diversity penalty | deep rankers with hundreds of features |
Not reproducible here: cross-user collaborative filtering, audio analysis and
sequence models. The add/reject lifecycle and the collection itself are the
only behavioural signals, which is why discovery leans on the catalog graph.
only behavioural signals, which is why discovery leans on the platforms' own
radio endpoints. Spotify is not an option: its `/v1/recommendations`, related
artists, audio features and audio analysis were removed for new apps on
2024-11-27, and its terms forbid training on Spotify content.
### 4.4 Download pipeline
@@ -711,8 +719,8 @@ Phases 1-8 are implemented in both repositories.
- Deezer is the default for new moods; every mood stores its own source and
can switch it, which clears the active recommendation pool and refills it.
- Chart 0.4.0 adds `DEFAULT_SOURCE`, `DEEZER_*` and an optional ARL secret.
- Tests: 94 server tests, 66-step smoke test with fake `yt-dlp`, `rip` and a
fake Deezer API (including the artist graph).
- Tests: 97 server tests, 67-step smoke test with fake `yt-dlp`, `rip` and a
fake Deezer API (including the artist graph and both radio paths).
### Client