4fdc6984bc
- Improved search results table with detailed metadata display - Added functionality to preview original MIDI alongside generated motifs - Implemented automatic selection of the best result with visual confidence indicators - Enhanced user experience with a "Try Next Result" button for A/B testing These updates provide users with better insights into MIDI quality and facilitate easier comparisons between original and generated compositions.
2.1 KiB
2.1 KiB
MOTIF — Aims, Scope, and Guidelines (for AI agents)
Aims (what we’re trying to achieve)
- Primary aim (MVP): Let a user search any song name, find real MIDI files online, load one, and play it back in-browser in a way that feels like a proper MIDI performance.
- Secondary aim (next step): From the same MIDI, generate a similar-but-different “Motif” version and play that.
- Future aim (later): Add stylistic transforms (“more ominous”, “dance/remix”), controls, and improved musical intelligence.
Scope (what is in-bounds right now)
- Search + ranking of MIDI candidates from multiple sources
- Fetch + validate + cache MIDI bytes via the backend
- Parse MIDI and show basic metadata (duration, track count, issues)
- Playback of the fetched MIDI (this is the current priority)
- Keep Motif generation working, but don’t block MVP playback on it
Non-scope (avoid for now)
- Perfect matching/similarity, chord/key analysis, advanced arrangement
- Big infra (DB/CDN/monitoring/analytics), large refactors
- New features unrelated to search/fetch/parse/playback reliability
Guidelines (how to work in this repo)
- Bias toward real MIDI: don’t silently replace real results with synthetic/mock in the default user path.
- Make degradation explicit: if mock/synthetic is used, it must be clearly indicated (UI/logs).
- Keep changes incremental: prioritize reliable end-to-end playback over architecture rewrites.
- Respect browser audio constraints: playback must work with autoplay policies (user gesture → resume AudioContext).
- Add practical observability: when fixing issues, add minimal logs/errors that explain which step failed (search vs fetch vs parse vs playback).
- Definition of done for any PR:
- Search returns results for common queries
- Selecting a result fetches bytes successfully (or shows a clear error)
- Playback starts/stops reliably without breaking subsequent plays
“Done means demo-able”
A change is successful if someone can:
- search “Hotel California”
- select a result
- hit Play and hear it
- hit Stop and try another result