Update app.py

app.py

@@ -7,6 +7,8 @@ import shutil
 from pathlib import Path
 from huggingface_hub import hf_hub_download
 from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer
+from groq import Groq                     # ← NEW for Groq API
+import os
 
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 
@@ -15,6 +17,38 @@ app = FastAPI()
 models = None
 en_emotion_map = None
 
+# ====================== GROQ CLIENT ======================
+groq_client = Groq(api_key=os.getenv("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"))
+
+def get_emotion_from_groq(text: str):
+    """Use Groq API to detect emotion for Sinhala text"""
+    try:
+        prompt = f"""
+        You are an expert emotion detector. 
+        Analyze the emotion of the following Sinhala text and return ONLY one word from this list: 
+        joy, sadness, anger, fear, surprise, disgust, love, neutral.
+
+        Text: {text}
+
+        Emotion:
+        """
+
+        response = groq_client.chat.completions.create(
+            model="llama-3.3-70b-versatile",   # Fast and good at this task
+            messages=[{"role": "user", "content": prompt}],
+            temperature=0.1,
+            max_tokens=20
+        )
+
+        emotion = response.choices[0].message.content.strip().lower()
+        logging.info(f"Groq API returned emotion: {emotion}")
+        return emotion
+
+    except Exception as e:
+        logging.error(f"Groq API error: {e}")
+        return "neutral"   # safe fallback
+
+
 def load_models():
     global models, en_emotion_map
     if models is not None:
@@ -23,7 +57,7 @@ def load_models():
     logging.info("📥 Loading models...")
 
     try:
-        # English & Sinhala (unchanged)
+        # English
         en_repo = "E-motionAssistant/English_LR_Model_New"
         en_vectorizer = joblib.load(hf_hub_download(en_repo, "tfidf_vectorizer.joblib"))
         en_classifier = joblib.load(hf_hub_download(en_repo, "logreg_model.joblib"))
@@ -36,42 +70,29 @@ def load_models():
         except:
             en_emotion_map = None
 
+        # Sinhala (still loaded but not used for prediction)
         si_vectorizer = joblib.load(hf_hub_download("E-motionAssistant/Sinhala_Text_Emotion_Model_LR", "tfidf_vectorizer.joblib"))
         si_classifier = joblib.load(hf_hub_download("E-motionAssistant/Sinhala_Text_Emotion_Model_LR", "logreg_model.joblib"))
         si_label_encoder = joblib.load(hf_hub_download("E-motionAssistant/Sinhala_Text_Emotion_Model_LR", "label_encoder.joblib"))
 
-        # ====================== TAMIL - STRONG FIX ======================
-        logging.info("📥 Loading Tamil model with manual components...")
-
-        # Clean cache
+        # Tamil
+        logging.info("📥 Loading Tamil model...")
         try:
             cache_dir = Path.home() / ".cache" / "huggingface" / "hub"
             model_cache = cache_dir / "models--E-motionAssistant--Tamil_Emotion_Recognition_Model"
             if model_cache.exists():
                 shutil.rmtree(model_cache)
-                logging.info("🧹 Cleaned Tamil cache")
         except:
             pass
 
-        # Load manually (more reliable than pipeline sometimes)
         model_name = "E-motionAssistant/Tamil_Emotion_Recognition_Model"
         tokenizer = AutoTokenizer.from_pretrained(model_name)
         model = AutoModelForSequenceClassification.from_pretrained(model_name)
 
-        tamil_pipe = pipeline(
-            "text-classification",
-            model=model,
-            tokenizer=tokenizer,
-            device=-1,
-            truncation=True,
-            max_length=512,
-            top_k=1
-        )
+        tamil_pipe = pipeline("text-classification", model=model, tokenizer=tokenizer, device=-1, truncation=True, max_length=512)
 
-        logging.info("✅ Tamil model loaded with manual tokenizer & model")
-
-        models = (en_vectorizer, en_classifier, en_label_encoder,
-                  si_vectorizer, si_classifier, si_label_encoder, tamil_pipe)
+        models = (en_vectorizer, en_clf=en_classifier, en_le=en_label_encoder,
+                  si_vec=si_vectorizer, si_clf=si_classifier, si_le=si_label_encoder, tamil_pipe=tamil_pipe)
 
         logging.info("🎉 All models loaded successfully!")
         return models
@@ -116,28 +137,16 @@ def predict(req: PredictRequest):
             return {"emotion": emotion, "language": "English"}
 
         elif lang == "sinhala":
-            X = si_vec.transform([req.text])
-            pred = si_clf.predict(X)[0]
-            emotion = si_le.inverse_transform([pred])[0]
-            return {"emotion": str(emotion), "language": "Sinhala"}
+            # === USE GROQ API FOR CORRECT PREDICTION ===
+            emotion = get_emotion_from_groq(req.text)
+            logging.info(f"Sinhala final emotion from Groq: {emotion}")
+            return {"emotion": emotion, "language": "Sinhala"}
 
         elif lang == "tamil":
-            logging.info(f"Tamil input: {req.text[:200]}...")
-
             result = tamil_pipe(req.text)
-
-            logging.info(f"Tamil raw result: {result}")
-
             emotion = result[0]['label']
             score = round(float(result[0]['score']), 4)
-
-            logging.info(f"Tamil Final → {emotion} ({score})")
-
-            return {
-                "emotion": emotion,
-                "confidence": score,
-                "language": "Tamil"
-            }
+            return {"emotion": emotion, "confidence": score, "language": "Tamil"}
 
     except Exception as e:
         logging.error(f"Error: {e}")