Update app.py

app.py

@@ -711,17 +711,24 @@ Provide clear, actionable advice while emphasizing the need for professional med
     
     def _generate_response(self, query: str, context: str) -> str:
         """Enhanced response generation using model.generate() to avoid DynamicCache errors"""
-        if self.llm is None or self.tokenizer is None:
-            return self._generate_fallback_response(query, context)
-            prompt = f"""{self.system_prompt} 
-            MEDICAL KNOWLEDGE CONTEXT:
-            {context}
-            PATIENT QUESTION: {query}
-            RESPONSE (provide practical, Gaza-appropriate medical guidance):"""
-        
-        try:
-            inputs = self.tokenizer(prompt, return_tensors="pt").to(self.llm.device)
+    if self.llm is None or self.tokenizer is None:
+        return self._generate_fallback_response(query, context)
+
+    # 🧠 Build prompt (this was in the wrong place before)
+    prompt = f"""{self.system_prompt}
+
+MEDICAL KNOWLEDGE CONTEXT:
+{context}
+
+PATIENT QUESTION: {query}
+
+RESPONSE (provide practical, Gaza-appropriate medical guidance):"""
+
+    try:
+        # ✅ Tokenize and move to correct device
+        inputs = self.tokenizer(prompt, return_tensors="pt").to(self.llm.device)
 
+        # ✅ Generate the response
         outputs = self.llm.generate(
             **inputs,
             max_new_tokens=800,
@@ -730,11 +737,22 @@ Provide clear, actionable advice while emphasizing the need for professional med
             do_sample=True,
             repetition_penalty=1.15,
         )
-        
-        except Exception as e: 
+
+        # ✅ Decode and clean up
+        response_text = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
+        lines = response_text.split('\n')
+        unique_lines = []
+        for line in lines:
+            line = line.strip()
+            if line and line not in unique_lines:
+                unique_lines.append(line)
+        return '\n'.join(unique_lines)
+
+    except Exception as e:
         logger.error(f"Error in LLM generate(): {e}")
         return self._generate_fallback_response(query, context)
 
+
         # Decode and clean up
         response_text = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
         lines = response_text.split('\n')