spoti

Vibe-based playlist generator over your Spotify liked songs. NLP coursework project — describe a mood, get a playlist drawn from your library.

Live: https://spoti.phirios.com

What it does

• Logs you in with Spotify OAuth, syncs your full liked-songs library in the background with progress + ETA.
• Enriches every track with Last.fm genre tags, GetSongBPM tempo, and LRCLIB lyrics (all cached in Postgres so each track only hits upstream once globally).
• Embeds each track with a local sentence-transformer (all-MiniLM-L6-v2, 384-d) using "<title> by <artists>. Tags: <last.fm tags>".
• Vibe playlists — type a free-form mood prompt; backend embeds it, cosine-sims against your library embeddings, returns top N. One click saves the result as a real Spotify playlist.
• Smart playlists — k-means over your embeddings, named by the most frequent Last.fm tags in each cluster. One click saves any cluster to Spotify.
• Click any track for a subpage with BPM, genres, lyrics, and similar tracks in your library (cosine sim within your own collection).
• Search bar + infinite scroll over the cached library.

Stack

• backend — Rust, axum, sqlx (Postgres), reqwest. OAuth flow, library sync engine, vibe matcher, auto-clusterer.
• embedder — Python, FastAPI, sentence-transformers, scikit-learn. CPU-only PyTorch. Two endpoints: /embed, /cluster.
• web — Next.js 16 + React 19 + Tailwind 4. URL-routed track and playlist subpages, polling-based sync progress UI.
• infra — k8s namespace nlp-project, plain Postgres deployment, nginx ingress with cert-manager, manifests in kubernetesmanifests.

Architecture

┌────────────────┐  OAuth  ┌────────────────┐
│ user (browser) │ ──────▶ │   spoti-web    │ ◀─ static landing + dashboard
└────────────────┘         └────────┬───────┘
                                    │ /api/* (Next rewrite)
                                    ▼
                           ┌────────────────┐    /embed    ┌─────────────────┐
                           │ spoti-backend  │ ───────────▶ │ spoti-embedder  │
                           │     (Rust)     │ ◀─ embeddings (sentence-       │
                           └────────┬───────┘                 transformers)  │
                            ┌───────┴───────┐                └────────────────┘
                            ▼               ▼
                       ┌─────────┐    ┌──────────┐
                       │postgres │    │ Spotify, │
                       │ (cache) │    │ Last.fm, │
                       └─────────┘    │ LRCLIB,  │
                                      │ GetSong- │
                                      │ BPM      │
                                      └──────────┘

Why per-track caches are global

Tracks (tracks, track_tags, track_embeddings, track_bpm, track_lyrics) are keyed by spotify_track_id, not per-user. The first user who triggers enrichment for a track pays the upstream cost; every subsequent user sees a cache hit. Only the /me/tracks enumeration itself is per-user (Spotify gives no other way to know what's in a specific user's library).

Project layout

spotiList/
├── backend/        Rust axum service
│   ├── src/        modules: auth, sync, library, playlists, auto, …
│   ├── migrations/ sqlx-migrate SQL files
│   └── Dockerfile
├── embedder/       Python FastAPI sidecar (sentence-transformers + sklearn)
│   └── app/main.py
└── web/            Next.js dashboard
    └── src/app/dashboard/
        ├── page.tsx
        ├── components/  Column/section/panel components
        ├── types.ts
        └── dummy.ts     localhost-only mock data for layout previews

Running locally

The web app has a DUMMY mode that activates when served from localhost so you can preview the layout without backend access.

cd web
bun install
bun dev   # → http://localhost:3000/dashboard with mock data

For the full stack (backend + embedder + Postgres) you'll want the k8s manifests or a docker-compose you assemble from the Dockerfiles. The embedder pre-downloads the model into the image; first build is heavy (~1.5 GB image with PyTorch CPU + model weights).

Notes

• Spotify deprecated the Audio Features, Audio Analysis, track popularity, artist genres, 30s previews, and recommendations endpoints for new apps in November 2024. We replace what we can: tempo via GetSongBPM (correct host api.getsong.co), genres via Last.fm tags, lyrics via LRCLIB.
• GetSongBPM has a 3000 req/hour limit. We cache hits and misses in track_bpm so we never re-query the same track.
• Last.fm tags are filtered to drop noise (year-only, decade, artist-name-equal tags) before being saved or used in embedding text.

Data attributions

• BPM data: GetSongBPM
• Lyrics: LRCLIB
• Genre tags: Last.fm

License

Coursework project; source is here for reading. No license attached yet — assume "all rights reserved" and ask if you want to use anything.