services/sentiment.py

"""
services/sentiment.py
Model sentimen berbasis IndoBERT / RoBERTa-ID.
Torch di-import secara lazy agar tidak crash saat package belum siap.
"""

import os

LOCAL_MODEL_PATH = "model/final_model"
FALLBACK_MODEL   = "w11wo/indonesian-roberta-base-sentiment-classifier"

# ── RULE-BASED FALLBACK ──
_POS_KW = [
    "bagus","baik","senang","suka","mantap","keren","hebat","oke","setuju",
    "benar","sukses","berhasil","love","good","great","nice","best","amazing",
    "excellent","wonderful","happy","glad","positif","mendukung","bangga",
    "luar biasa","terima kasih","apresiasi","semangat","maju","berkembang",
]
_NEG_KW = [
    "buruk","jelek","benci","kecewa","gagal","salah","rugi","marah","bohong",
    "hoax","fitnah","jahat","tidak setuju","parah","malu","takut","bad",
    "worst","terrible","hate","fail","wrong","poor","awful","negatif","tolak",
    "menolak","turun","jatuh","hancur","krisis","masalah","bahaya","ancam",
]

def _rule_based(text: str) -> str:
    lower = text.lower()
    pos   = sum(1 for k in _POS_KW if k in lower)
    neg   = sum(1 for k in _NEG_KW if k in lower)
    if pos > neg:   return "Positive"
    if neg > pos:   return "Negative"
    return "Neutral"


# ── MODEL LOADING ──
def _load_model():
    try:
        import torch
        from transformers import pipeline

        path  = LOCAL_MODEL_PATH if os.path.exists(LOCAL_MODEL_PATH) else FALLBACK_MODEL
        label = "fine-tuned" if os.path.exists(LOCAL_MODEL_PATH) else "fallback RoBERTa-ID"

        clf = pipeline(
            "sentiment-analysis",
            model=path,
            device=-1,
            truncation=True,
            max_length=512,
        )
        print(f"✅ Sentiment model loaded: {label}")
        return clf

    except ImportError:
        print("⚠️  PyTorch tidak tersedia — rule-based fallback aktif")
        return None
    except Exception as e:
        print(f"❌ Gagal load sentiment model: {e}")
        return None


classifier = _load_model()


# ── LABEL NORMALIZATION ──
def _normalize(label: str) -> str:
    label = label.lower()
    if "positive" in label or label == "label_2": return "Positive"
    if "negative" in label or label == "label_0": return "Negative"
    if "neutral"  in label or label == "label_1": return "Neutral"
    return "Neutral"


# ── PUBLIC API ──
def predict(texts: list) -> list:
    """Return list of label strings."""
    if not texts: return []
    if classifier is None:
        return [_rule_based(t) for t in texts]
    try:
        outputs = classifier(texts, batch_size=8, truncation=True)
        return [_normalize(o["label"]) for o in outputs]
    except Exception as e:
        print(f"❌ predict() batch error: {e} — per-item fallback")
        results = []
        for t in texts:
            try:
                out = classifier(t[:512], truncation=True)
                results.append(_normalize(out[0]["label"]))
            except Exception:
                results.append(_rule_based(t))
        return results


def predict_single(text: str) -> str:
    return predict([text])[0]


def predict_with_score(texts: list) -> list:
    """
    Return list of dicts: {label, score}
    score = confidence dari model (0–1).
    """
    if not texts: return []

    if classifier is None:
        return [{"label": _rule_based(t), "score": 0.5} for t in texts]

    try:
        outputs = classifier(texts, batch_size=8, truncation=True)
        return [
            {"label": _normalize(o["label"]), "score": round(float(o["score"]), 4)}
            for o in outputs
        ]
    except Exception as e:
        print(f"❌ predict_with_score() error: {e} — per-item fallback")
        results = []
        for t in texts:
            try:
                out = classifier(t[:512], truncation=True)
                results.append({
                    "label": _normalize(out[0]["label"]),
                    "score": round(float(out[0]["score"]), 4)
                })
            except Exception:
                results.append({"label": _rule_based(t), "score": 0.5})
        return results