Update handler.py

handler.py

@@ -17,23 +17,30 @@ class EndpointHandler:
         print("Model Ready!")
 
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
-        # 2. This runs every time your website sends a new message
-        
-        # Extract the user's message from the incoming API request
+        # Get the raw message from the website
         user_input = data.pop("inputs", "")
         
-        # Convert text to numbers for the AI
-        input_ids = self.tokenizer(user_input, return_tensors="pt").input_ids.to(self.model.device)
+        # 1. THE FIX: Format it EXACTLY like your university training data!
+        system_prompt = "You are a helpful, accurate, and friendly Ormuri language assistant. Only provide the exact translation. Do not invent words."
+        formatted_prompt = f"### Instruction:\n{system_prompt}\n\nUser Question: {user_input}\n\n### Response:\n"
+        
+        # Convert text to numbers
+        input_ids = self.tokenizer(formatted_prompt, return_tensors="pt").input_ids.to(self.model.device)
         
-        # Generate the response!
+        # 2. THE FIX: Turn down the temperature so it stops hallucinating fake words!
         output_ids = self.model.generate(
             input_ids, 
             max_new_tokens=150, 
-            pad_token_id=self.tokenizer.eos_token_id
+            pad_token_id=self.tokenizer.eos_token_id,
+            temperature=0.1, # Strictly factual, no guessing
+            do_sample=True
         )
         
-        # Convert the numbers back into Ormuri text
-        full_response = self.tokenizer.decode(output_ids[0], skip_special_tokens=True)
+        # Slice the output to ONLY grab the brand new words after "### Response:\n"
+        new_tokens = output_ids[0][input_ids.shape[1]:] 
+        
+        # Decode back to text
+        final_answer = self.tokenizer.decode(new_tokens, skip_special_tokens=True)
         
-        # Send it back to your website exactly how Hugging Face expects it
-        return [{"generated_text": full_response}]
+        # Send ONLY the clean answer back
+        return [{"generated_text": final_answer.strip()}]