Update handler.py

handler.py

@@ -16,50 +16,53 @@ class EndpointHandler:
         )
 
         self.generation_config = {
-            "max_new_tokens": 160,
+            "max_new_tokens": 180,
             "temperature": 0.7,
             "top_p": 0.9,
             "do_sample": True,
             "pad_token_id": self.tokenizer.eos_token_id,
             "eos_token_id": self.tokenizer.eos_token_id,
-            "repetition_penalty": 1.1
+            "repetition_penalty": 1.12
         }
 
-    def __call__(self, data: Dict[str, Any]) -> Any:
+    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
         try:
-            # History kommt von der App (letzte Nachrichten)
-            messages: List[Dict] = data.get("inputs", [])
-            
-            # Nur letzte 10 Nachrichten behalten (Sliding Window)
-            if len(messages) > 10:
-                messages = messages[-10:]
+            # Flexible Input Verarbeitung (Playground + API)
+            raw_input = data.get("inputs") or data.get("messages") or data.get("input", "")
+
+            if isinstance(raw_input, str):
+                messages = [{"role": "user", "content": raw_input}]
+            elif isinstance(raw_input, list):
+                messages = raw_input
+            else:
+                messages = [{"role": "user", "content": str(raw_input)}]
 
             # Prompt bauen
-            prompt = """Du si profesionalni Telesales agent.
+            prompt = """Ti si profesionalni, brzi Telesales agent.
 Govori samo bosanski/srpski/hrvatski.
 Budi direktan, ljubazan i prodajno jak.
 Bez dugih priča. Slušaj kupca i zatvaraj prodaju.
 
 """
 
-            for msg in messages:
+            for msg in messages[-10:]:  # Sliding Window
                 role = "Kupac" if msg.get("role") == "user" else "Agent"
                 prompt += f"{role}: {msg.get('content', '')}\n"
 
             prompt += "Agent:"
 
-            # Token Begrenzung
+            # Tokenisierung
             inputs = self.tokenizer(prompt, return_tensors="pt", truncation=True, max_length=5000).to("cuda")
             
             outputs = self.model.generate(**inputs, **self.generation_config)
             
             response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
             
-            # Nur den neuen Teil nehmen
             if "Agent:" in response:
                 response = response.split("Agent:")[-1].strip()
 
-            return response.strip()
+            # Wichtiges Return-Format für HF Inference
+            return {"generated_text": response.strip()}
 
-        except Exception:
-            return "Žao mi je, došlo je do greške. Možete li ponoviti?"
+        except Exception as e:
+            return {"generated_text": f"Žao mi je, došlo je do greške. Ponovite molim."}