refactor: render from output and derive content at save

Author: Algorithm5838

backend/open_webui/utils/middleware.py

@@ -4069,7 +4069,6 @@ async def flush_pending_delta_data(threshold: int = 0):
 
                                     processed_data = {
                                         'output': full_output(),
-                                        'content': serialize_output(full_output()),
                                     }
 
                                     # print(data)

src/lib/components/chat/Chat.svelte

@@ -1805,6 +1805,12 @@
 			message.output = output;
 		}
 
+		// When the response is done and content is provided, set it for copy/TTS
+		// (the frontend renders from output during streaming, content is the final cache)
+		if (done && content) {
+			message.content = content;
+		}
+
 		if (error) {
 			await handleOpenAIError(error, message);
 		}
@@ -1813,7 +1819,9 @@
 			message.sources = sources;
 		}
 
-		if (choices) {
+		// Only accumulate content from choices when there's no structured output
+		// (output-based rendering takes priority — avoids redundant serialize/parse round-trip)
+		if (choices && !output) {
 			if (choices[0]?.message?.content) {
 				// Non-stream response
 				message.content += choices[0]?.message?.content;