update retry logic to retry after retr_delay sepcified during error:429

app.py

@@ -66,7 +66,7 @@ class BasicAgent:
         return isinstance(e, genai.errors.APIError) and getattr(e, "code", None) in {429, 503}
 
     def invoke_with_retry(self,question: str, max_retries: int = 5, initial_delay: float = 10.0) -> str:
-        delay = initial_delay
+        current_delay = initial_delay
         for attempt in range(max_retries):
             try:
                 result = self.agent_obj.invoke(
@@ -76,12 +76,22 @@ class BasicAgent:
                 return result['output']
             except Exception as e:
                 if self.is_retriable(e):
-                    print(f"Quota error (attempt {attempt+1}/{max_retries}), retrying in {delay} seconds...",flush=True)
-                    time.sleep(delay)
-                    delay *= 2  # Exponential backoff
+                    # Check if the error object provides a specific retry_delay
+                    if hasattr(e, 'retry_delay') and hasattr(e.retry_delay, 'seconds'):
+                        # Use the specific retry_delay provided by the API
+                        current_delay = float(e.retry_delay.seconds)
+                        print(f"Quota error (attempt {attempt+1}/{max_retries}). API suggested retry after {current_delay} seconds.", flush=True)
+                    else:
+                        # Fallback to exponential backoff if no specific delay is provided
+                        print(f"Quota error (attempt {attempt+1}/{max_retries}). Retrying in {current_delay} seconds with exponential backoff.", flush=True)
+                        current_delay *= 2  # Exponential backoff
+                    
+                    time.sleep(current_delay)
                 else:
+                    # If it's not a retriable error, re-raise it
                     raise
-        raise RuntimeError("Max retries exceeded due to quota errors.")
+        # If all retries fail, raise a RuntimeError
+        raise RuntimeError(f"Max retries ({max_retries}) exceeded due to persistent quota errors or other retriable issues.")
 
     def __call__(self, question: str) -> str:
         """