Update app.py

app.py

@@ -27,19 +27,15 @@ class HuggingFaceInferenceWrapper:
         self.inference_api = inference_api
 
     def generate(self, prompt: str, **kwargs) -> str:
-        # Request raw response and handle JSON decoding
-        response = self.inference_api(inputs=prompt, raw_response=True)
-        result_json = response.json()
-        
-        # Handle both plain text and chat-style model formats
-        if isinstance(result_json, dict):
-            if "generated_text" in result_json:
-                return result_json["generated_text"].strip()
-            elif "outputs" in result_json and isinstance(result_json["outputs"], str):
-                return result_json["outputs"].strip()
-        
-        # If unknown format, fallback
-        return str(result_json)
+        # Call inference API - returns string directly
+        response = self.inference_api(inputs=prompt)
+        if isinstance(response, str):
+            return response.strip()
+        elif hasattr(response, 'text'):
+            return response.text.strip()
+        else:
+            # Handle JSON response
+            return str(response)