From b2a9be27c5146a49c7389db22846eb13b803fef6 Mon Sep 17 00:00:00 2001
From: Classic298 <27028174+Classic298@users.noreply.github.com>
Date: Thu, 20 Aug 2026 22:11:09 +0200
Subject: [PATCH] perf: stop recopying the streamed message text on every delta

Every streamed token appended to the response text through a dict slot, which allocates a new string and copies everything received so far. A long answer pays that copy thousands of times on the event loop, so every other request on the worker waits behind it.

The appends now clear the slot before growing the string, which leaves a single reference and lets CPython resize it in place. This covers the six accumulation points in the Chat Completions streaming path. The Responses API transport copies its items before appending, so it holds two references and is unchanged.

Measured on CPython 3.12 with 10-character deltas:

| accumulated text | before | after |
|---|---|---|
| 100 KB | 5.97 ms | 0.98 ms |
| 250 KB | 29.53 ms | 2.83 ms |
| 1 MB | 659 ms | 15.5 ms |

Results match the replaced expressions for every input where the key exists and holds a string, key order and exception type included. Buffering into a list and joining at the end was the alternative, but each delta is read back immediately by the stream save, and joining per read measured 666 ms at 250 KB.
---
 backend/open_webui/utils/middleware.py | 21 +++++++++++++++------
 1 file changed, 15 insertions(+), 6 deletions(-)

diff --git a/backend/open_webui/utils/middleware.py b/backend/open_webui/utils/middleware.py
index ebb51cb559..06b420f6c7 100644
--- a/backend/open_webui/utils/middleware.py
+++ b/backend/open_webui/utils/middleware.py
@@ -303,6 +303,15 @@ def tool_result_content(tool_result: Any) -> str:
     return str(tool_result)
 
 
+def append_text(container: dict, key: str, value: str) -> None:
+    text = container.get(key, '')
+    container[key] = ''  # dropping the dict's reference lets CPython extend the str in place
+    try:
+        text += value
+    finally:  # a non-str value raises above and would leave the slot cleared
+        container[key] = text
+
+
 def merge_streamed_reasoning_details(target: list, details) -> None:
     items = details if isinstance(details, list) else [details]
     for item in items:
@@ -321,7 +330,7 @@ def merge_streamed_reasoning_details(target: list, details) -> None:
 
         for key, value in item.items():
             if key in ('text', 'summary') and isinstance(value, str) and isinstance(existing.get(key), str):
-                existing[key] += value
+                append_text(existing, key, value)
             else:
                 existing[key] = value
 
@@ -5136,7 +5145,7 @@ async def streaming_chat_response_handler(response, ctx):
                                             # Append to reasoning content
                                             parts = reasoning_item.get('content', [])
                                             if parts and parts[-1].get('type') == 'output_text':
-                                                parts[-1]['text'] += reasoning_content
+                                                append_text(parts[-1], 'text', reasoning_content)
                                             else:
                                                 reasoning_item['content'] = [
                                                     {
@@ -5234,11 +5243,11 @@ async def streaming_chat_response_handler(response, ctx):
                                         if inside_tag_block:
                                             # Append to the existing tag-based item
                                             if last_item_type == 'open_webui:code_interpreter':
-                                                last_item['code'] = last_item.get('code', '') + value
+                                                append_text(last_item, 'code', value)
                                             elif last_item_type == 'reasoning':
                                                 parts = last_item.get('content', [])
                                                 if parts and parts[-1].get('type') == 'output_text':
-                                                    parts[-1]['text'] += value
+                                                    append_text(parts[-1], 'text', value)
                                                 else:
                                                     last_item['content'] = [
                                                         {
@@ -5250,7 +5259,7 @@ async def streaming_chat_response_handler(response, ctx):
                                                 # solution or other _tag_type message
                                                 msg_parts = last_item.get('content', [])
                                                 if msg_parts and msg_parts[-1].get('type') == 'output_text':
-                                                    msg_parts[-1]['text'] += value
+                                                    append_text(msg_parts[-1], 'text', value)
                                                 else:
                                                     last_item['content'] = [
                                                         {
@@ -5278,7 +5287,7 @@ async def streaming_chat_response_handler(response, ctx):
                                             # Append value to last message item's text
                                             msg_parts = output[-1].get('content', [])
                                             if msg_parts and msg_parts[-1].get('type') == 'output_text':
-                                                msg_parts[-1]['text'] += value
+                                                append_text(msg_parts[-1], 'text', value)
                                             else:
                                                 output[-1]['content'] = [
                                                     {