File: kolom Saran diambil. File besar (11.000 baris) bisa butuh 3-10 menit tergantung model & hardware.
Alur:
1. Filter noise (kosong, "-", "TA", "Tidak ada")
2. Filter apresiasi murni ("Bagus", "Baik")
3. Dedup kalimat identik
4. Embedding semantik multilingual
5. HDBSCAN clustering
6. Label otomatis per cluster via Ollama