"fix_backend_retry_logic"

app/server/chat.py

@@ -786,8 +786,7 @@ async def _send_with_split(
         content: str, file_list: list | None
     ) -> AsyncGenerator[ModelOutput, None]:
         """Manual retry logic for streaming."""
-        # Reduced retries to 1 to minimize ban risk
-        max_retries = 1
+        max_retries = 2 # Increased from 1
         for attempt in range(max_retries + 1):
             try:
                 gen = session.send_message_stream(content, files=file_list)
@@ -798,19 +797,18 @@ async def _send_with_split(
                 return
             except Exception as e:
                 if has_yielded:
+                    # If we already started yielding data, we can't easily retry the stream
+                    # without sending duplicate content to the client.
+                    logger.error(f"Stream interrupted after yielding data: {e}")
                     raise e
                 
-                # Safety check: Do not retry if error indicates auth failure or rate limit
                 error_str = str(e).lower()
-                if "429" in error_str or "403" in error_str or "quota" in error_str:
+                # Safety check: Do not retry if error indicates auth failure or rate limit
+                if any(code in error_str for code in ["429", "403", "401", "quota"]):
                     raise e
 
                 if attempt < max_retries:
-                    # Increased delay with randomness to avoid pattern detection
-                    base_delay = 3.0
-                    jitter = random.uniform(0.5, 1.5)
-                    delay = base_delay + jitter
-                    
+                    delay = (attempt + 1) * 3.0 + random.uniform(0.5, 1.5)
                     logger.warning(f"Stream failed to start (attempt {attempt+1}/{max_retries}). Retrying in {delay:.2f}s. Error: {e}")
                     await asyncio.sleep(delay)
                 else:
@@ -818,18 +816,26 @@ async def _send_with_split(
 
     @retry_with_backoff(
         max_retries=3,
-        base_delay=1.0,
+        base_delay=2.0,
         max_delay=30.0,
         exponential_base=2.0,
-        retryable_exceptions=(ConnectionError, TimeoutError, OSError),
+        retryable_exceptions=(ConnectionError, TimeoutError, OSError, Exception), # Catch all for APIError
     )
     async def _send_with_retry(
         content: str, file_list: list | None, is_stream: bool
     ) -> AsyncGenerator[ModelOutput, None] | ModelOutput:
         """Internal function with retry logic."""
-        if is_stream:
-            return _stream_with_retry(content, file_list)
-        return await session.send_message(content, files=file_list)
+        try:
+            if is_stream:
+                return _stream_with_retry(content, file_list)
+            return await session.send_message(content, files=file_list)
+        except Exception as e:
+            # If it's a known "Stream interrupted" error, we want to retry
+            error_msg = str(e)
+            if "Stream interrupted" in error_msg or "truncated" in error_msg:
+                logger.warning(f"Gemini stream interrupted, triggering retry: {e}")
+                raise e # Let decorator handle retry
+            raise e
     
     if len(text) <= MAX_CHARS_PER_REQUEST:
         try: