perf: optional orjson JSON codec behind ENABLE_ORJSON (#27583)

Swap the JSON encoder/decoder used across the backend from stdlib json to
orjson when ENABLE_ORJSON is set — HTTP request bodies, JSONResponse
bodies, upstream provider responses, SSE chunks, and socket.io/Redis
payloads.

The flag defaults to off, in which case the app uses stdlib json and
engineio's codec verbatim, so default behaviour is unchanged.

- json_codec exports JSONCodec (stdlib json or the orjson codec) and
  SOCKETIO_JSON (engineio's codec or the orjson codec); call sites import
  JSONCodec and stay implementation-agnostic
- apply_orjson_http_json() is a no-op when the flag is off, leaving
  starlette's Request.json / JSONResponse.render untouched
- the orjson codec falls back to the stdlib for inputs orjson rejects
  (non-str dict keys, ints beyond 64 bits, NaN literals)
- orjson is imported only when the flag is on
- FastAPI(default_response_class=...) is deliberately not used: an
  explicit default disables the Pydantic direct-to-bytes fast path for
  response_model routes

backend/open_webui/env.py

@@ -150,6 +150,11 @@ INSTANCE_ID = os.getenv('INSTANCE_ID', str(uuid4()))
 
 ENABLE_DB_MIGRATIONS = os.getenv('ENABLE_DB_MIGRATIONS', 'True').lower() == 'true'
 
+# Swap the JSON encoder/decoder used across the app (HTTP request bodies, JSONResponse
+# bodies, upstream provider responses, socket.io payloads) from the stdlib `json` module
+# to orjson. Faster, but stricter: see open_webui/utils/json_codec.py for the differences.
+ENABLE_ORJSON = os.getenv('ENABLE_ORJSON', 'False').lower() == 'true'
+
 
 # Function to parse each section
 def parse_section(section):

backend/open_webui/main.py

@@ -229,6 +229,7 @@ from open_webui.utils.chat_variables import (
     normalize_chat_variables,
 )
 from open_webui.utils.embeddings import generate_embeddings
+from open_webui.utils.json_response import apply_orjson_http_json
 from open_webui.utils.logger import start_logger
 from open_webui.utils.middleware import (
     background_tasks_handler,
@@ -456,6 +457,10 @@ async def lifespan(app: FastAPI):
     await publish_event(app, EVENTS.SYSTEM_SHUTDOWN_COMPLETED, source='system')
 
 
+# Opt-in (ENABLE_ORJSON): orjson for request-body parsing and JSONResponse bodies;
+# response_model routes keep FastAPI's Pydantic fast path either way.
+apply_orjson_http_json()
+
 app = FastAPI(
     title='Open WebUI',
     docs_url='/docs' if ENV == 'dev' else None,

backend/open_webui/routers/ollama.py

@@ -42,6 +42,7 @@ from open_webui.utils.access_control import check_model_access
 from open_webui.utils.auth import get_admin_user, get_verified_user
 from open_webui.utils.headers import get_custom_headers, include_user_info_headers
 from open_webui.utils.model_ids import strip_provider_model_prefix
+from open_webui.utils.json_codec import JSONCodec
 from open_webui.utils.misc import calculate_sha256
 from open_webui.utils.payload import (
     apply_model_params_to_body_ollama,
@@ -86,7 +87,7 @@ async def send_get_request(
             ssl=AIOHTTP_CLIENT_SESSION_SSL,
             timeout=aiohttp.ClientTimeout(total=AIOHTTP_CLIENT_TIMEOUT_MODEL_LIST),
         ) as r:
-            return await r.json()
+            return await r.json(loads=JSONCodec.loads)
     except Exception as exc:
         log.error(f'Connection error: {exc}')
         return None
@@ -137,7 +138,7 @@ async def send_request(
 
         if not r.ok:
             try:
-                res = await r.json()
+                res = await r.json(loads=JSONCodec.loads)
                 await publish_model_provider_request_failed(
                     request,
                     actor=user,
@@ -179,7 +180,7 @@ async def send_request(
             )
         else:
             try:
-                return await r.json()
+                return await r.json(loads=JSONCodec.loads)
             except Exception:
                 return None
 
@@ -270,12 +271,12 @@ async def verify_connection(
         ) as r:
             if r.status != 200:
                 detail = f'HTTP Error: {r.status}'
-                res = await r.json()
+                res = await r.json(loads=JSONCodec.loads)
                 if 'error' in res:
                     detail = f'External Error: {res["error"]}'
                 raise Exception(detail)
 
-            return await r.json()
+            return await r.json(loads=JSONCodec.loads)
     except aiohttp.ClientError as exc:
         log.exception(f'Client error: {exc}')
         raise HTTPException(status_code=500, detail=ERROR_MESSAGES.SERVER_CONNECTION_ERROR)

backend/open_webui/routers/openai.py

@@ -44,6 +44,7 @@ from open_webui.utils.access_control import check_model_access, has_connection_a
 from open_webui.utils.anthropic import get_anthropic_models, is_anthropic_url
 from open_webui.utils.auth import get_admin_user, get_verified_user
 from open_webui.utils.headers import get_custom_headers, include_user_info_headers
+from open_webui.utils.json_codec import JSONCodec
 from open_webui.utils.model_ids import strip_provider_model_prefix
 from open_webui.utils.misc import (
     convert_logit_bias_input_to_json,
@@ -112,7 +113,7 @@ async def send_get_request(
                 cookies=cookies,
                 ssl=AIOHTTP_CLIENT_SESSION_SSL,
             ) as response:
-                return await response.json()
+                return await response.json(loads=JSONCodec.loads)
     except Exception as e:
         # Handle connection error here
         log.error(f'Connection error: {e}')
@@ -357,7 +358,7 @@ async def count_anthropic_tokens(request: Request, form_data: dict, user: UserMo
         )
 
         try:
-            response_data = await response.json()
+            response_data = await response.json(loads=JSONCodec.loads)
         except Exception:
             response_data = await response.text()
 
@@ -509,7 +510,7 @@ async def speech(request: Request, user=Depends(get_verified_user)):
             detail = None
             if r is not None:
                 try:
-                    res = await r.json()
+                    res = await r.json(loads=JSONCodec.loads)
                     if 'error' in res:
                         detail = f'External: {res["error"]}'
                 except Exception:
@@ -763,14 +764,14 @@ async def get_models(request: Request, url_idx: int | None = None, user=Depends(
                         if r.status != 200:
                             error_detail = f'HTTP Error: {r.status}'
                             try:
-                                res = await r.json()
+                                res = await r.json(loads=JSONCodec.loads)
                                 if 'error' in res:
                                     error_detail = f'External Error: {res["error"]}'
                             except Exception:
                                 pass
                             raise Exception(error_detail)
 
-                        response_data = await r.json()
+                        response_data = await r.json(loads=JSONCodec.loads)
 
                         if 'api.openai.com' in url:
                             response_data['data'] = [
@@ -853,7 +854,7 @@ async def verify_connection(
                     ssl=AIOHTTP_CLIENT_SESSION_SSL,
                 ) as r:
                     try:
-                        response_data = await r.json()
+                        response_data = await r.json(loads=JSONCodec.loads)
                     except Exception:
                         response_data = await r.text()
 
@@ -879,7 +880,7 @@ async def verify_connection(
                     ssl=AIOHTTP_CLIENT_SESSION_SSL,
                 ) as r:
                     try:
-                        response_data = await r.json()
+                        response_data = await r.json(loads=JSONCodec.loads)
                     except Exception:
                         response_data = await r.text()
 
@@ -1417,7 +1418,7 @@ async def generate_chat_completion(
             )
         else:
             try:
-                response = await r.json()
+                response = await r.json(loads=JSONCodec.loads)
             except Exception as e:
                 log.error(e)
                 response = await r.text()
@@ -1529,7 +1530,7 @@ async def embeddings(request: Request, form_data: dict, user):
             )
         else:
             try:
-                response_data = await r.json()
+                response_data = await r.json(loads=JSONCodec.loads)
             except Exception:
                 response_data = await r.text()
 
@@ -1656,7 +1657,7 @@ async def responses(
             )
         else:
             try:
-                response_data = await r.json()
+                response_data = await r.json(loads=JSONCodec.loads)
             except Exception:
                 response_data = await r.text()
 
@@ -1778,7 +1779,7 @@ async def proxy(path: str, request: Request, user=Depends(get_verified_user)):
             )
         else:
             try:
-                response_data = await r.json()
+                response_data = await r.json(loads=JSONCodec.loads)
             except Exception:
                 response_data = await r.text()
 

backend/open_webui/socket/main.py

@@ -39,6 +39,7 @@ from open_webui.tasks import create_task, stop_item_tasks
 from open_webui.utils.access_control import has_permission
 from open_webui.utils.auth import get_verified_user_by_token
 from open_webui.utils.chat_id import is_saved_chat_id
+from open_webui.utils.json_codec import SOCKETIO_JSON
 from open_webui.utils.misc import get_output_text
 from open_webui.utils.redis import (
     build_sentinel_url,
@@ -70,10 +71,11 @@ if WEBSOCKET_MANAGER == 'redis':
         if sentinel_hosts
         else WEBSOCKET_REDIS_URL
     )
-    redis_manager = socketio.AsyncRedisManager(ws_redis_url, redis_options=WEBSOCKET_REDIS_OPTIONS)
+    redis_manager = socketio.AsyncRedisManager(ws_redis_url, redis_options=WEBSOCKET_REDIS_OPTIONS, json=SOCKETIO_JSON)
     sio = socketio.AsyncServer(
         cors_allowed_origins=SOCKETIO_CORS_ORIGINS,
         async_mode='asgi',
+        json=SOCKETIO_JSON,
         transports=(['websocket'] if ENABLE_WEBSOCKET_SUPPORT else ['polling']),
         allow_upgrades=ENABLE_WEBSOCKET_SUPPORT,
         always_connect=True,
@@ -87,6 +89,7 @@ else:
     sio = socketio.AsyncServer(
         cors_allowed_origins=SOCKETIO_CORS_ORIGINS,
         async_mode='asgi',
+        json=SOCKETIO_JSON,
         transports=(['websocket'] if ENABLE_WEBSOCKET_SUPPORT else ['polling']),
         allow_upgrades=ENABLE_WEBSOCKET_SUPPORT,
         always_connect=True,

backend/open_webui/socket/utils.py

@@ -3,11 +3,11 @@
 from __future__ import annotations
 
 import hashlib
-import json
 import uuid
 
 import pycrdt as Y
 from open_webui.env import REDIS_KEY_PREFIX
+from open_webui.utils.json_codec import JSONCodec
 from open_webui.utils.redis import get_redis_connection
 
 YDOC_KEY_PREFIX = f'{REDIS_KEY_PREFIX}:ydoc:documents'
@@ -75,14 +75,14 @@ class RedisDict:
         )
 
     def __setitem__(self, key, value):
-        serialized_value = json.dumps(value)
+        serialized_value = JSONCodec.dumps(value)
         self.redis.hset(self.name, key, serialized_value)
 
     def __getitem__(self, key):
         value = self.redis.hget(self.name, key)
         if value is None:
             raise KeyError(key)
-        return json.loads(value)
+        return JSONCodec.loads(value)
 
     def __delitem__(self, key):
         result = self.redis.hdel(self.name, key)
@@ -99,10 +99,10 @@ class RedisDict:
         return self.redis.hkeys(self.name)
 
     def values(self):
-        return [json.loads(v) for v in self.redis.hvals(self.name)]
+        return [JSONCodec.loads(v) for v in self.redis.hvals(self.name)]
 
     def items(self):
-        return [(k, json.loads(v)) for k, v in self.redis.hgetall(self.name).items()]
+        return [(k, JSONCodec.loads(v)) for k, v in self.redis.hgetall(self.name).items()]
 
     def set(self, mapping: dict):
         if not mapping:
@@ -111,7 +111,7 @@ class RedisDict:
             return
 
         # Serialize values once — reused for both the fingerprint and the write.
-        serialized = {k: json.dumps(v) for k, v in mapping.items()}
+        serialized = {k: JSONCodec.dumps(v) for k, v in mapping.items()}
         digest = hashlib.sha256()
         for key in sorted(serialized):
             digest.update(key.encode())
@@ -182,7 +182,7 @@ class YdocManager:
         document_id = document_id.replace(':', '_')
         if self._redis:
             redis_key = f'{self._redis_key_prefix}:{document_id}:updates'
-            await self._redis.rpush(redis_key, json.dumps(list(update)))
+            await self._redis.rpush(redis_key, JSONCodec.dumps(list(update)))
             list_len = await self._redis.llen(redis_key)
             if list_len >= self.COMPACTION_THRESHOLD:
                 await self._compact_updates_redis(document_id)
@@ -202,8 +202,8 @@ class YdocManager:
         mid = len(all_updates) // 2
         ydoc = Y.Doc()
         for raw in all_updates[:mid]:
-            ydoc.apply_update(bytes(json.loads(raw)))
+            ydoc.apply_update(bytes(JSONCodec.loads(raw)))
-        snapshot = json.dumps(list(ydoc.get_update()))
+        snapshot = JSONCodec.dumps(list(ydoc.get_update()))
         pipe = self._redis.pipeline()
         pipe.delete(redis_key)
         pipe.rpush(redis_key, snapshot, *all_updates[mid:])
@@ -226,7 +226,7 @@ class YdocManager:
         if self._redis:
             redis_key = f'{self._redis_key_prefix}:{document_id}:updates'
             updates = await self._redis.lrange(redis_key, 0, -1)
-            return [bytes(json.loads(update)) for update in updates]
+            return [bytes(JSONCodec.loads(update)) for update in updates]
         else:
             return self._updates.get(document_id, [])
 

backend/open_webui/utils/json_codec.py

@@ -0,0 +1,50 @@
+"""The app-wide JSON codec, selected by the ``ENABLE_ORJSON`` env var.
+
+Every module that would otherwise reach for stdlib ``json`` imports ``JSONCodec``
+from here, so the whole app switches implementation from a single flag. With the
+flag off these are stdlib ``json`` and engineio's codec verbatim, so the default
+behaviour is exactly what it was before orjson entered the picture.
+"""
+
+from __future__ import annotations
+
+import json as stdlib_json
+
+from engineio import json as engineio_json
+
+from open_webui.env import ENABLE_ORJSON
+
+if ENABLE_ORJSON:
+    import orjson
+
+    class ORJSONCodec:
+        """stdlib-``json``-compatible codec backed by orjson.
+
+        Anything orjson rejects (non-str dict keys, ints beyond 64 bits, ``NaN``
+        literals) falls back to engineio's stdlib-based codec, which keeps its
+        oversized-integer guard for untrusted client payloads.
+        """
+
+        JSONDecodeError = engineio_json.JSONDecodeError
+
+        @staticmethod
+        def dumps(obj, *args, **kwargs):
+            try:
+                return orjson.dumps(obj).decode('utf-8')
+            except (TypeError, ValueError):
+                return engineio_json.dumps(obj, *args, **kwargs)
+
+        @staticmethod
+        def loads(s, *args, **kwargs):
+            try:
+                return orjson.loads(s)
+            except (TypeError, ValueError):
+                return engineio_json.loads(s, *args, **kwargs)
+
+    # Drop-in for stdlib ``json``: ``JSONCodec.dumps`` / ``JSONCodec.loads``.
+    JSONCodec = ORJSONCodec
+    # Codec handed to the socket.io/engineio managers, which default to their own.
+    SOCKETIO_JSON = ORJSONCodec
+else:
+    JSONCodec = stdlib_json
+    SOCKETIO_JSON = engineio_json

backend/open_webui/utils/json_response.py

@@ -0,0 +1,55 @@
+"""orjson-backed JSON parsing/rendering for starlette/FastAPI requests and responses."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from starlette.requests import Request
+from starlette.responses import JSONResponse
+
+from open_webui.env import ENABLE_ORJSON
+
+
+def apply_orjson_http_json() -> None:
+    """Parse request bodies and serialize ``JSONResponse`` with orjson.
+
+    A no-op unless ``ENABLE_ORJSON`` is set, leaving starlette's own
+    stdlib-``json`` implementations untouched.
+
+    Not ``FastAPI(default_response_class=...)`` on purpose: an explicit
+    default disables FastAPI's Pydantic direct-to-bytes fast path for
+    ``response_model`` routes.  NaN/Infinity floats serialize as ``null``
+    instead of raising.
+    """
+    if not ENABLE_ORJSON:
+        return
+
+    import orjson
+
+    def render(self, content: Any) -> bytes:
+        try:
+            return orjson.dumps(content)
+        except (TypeError, ValueError):
+            # Fallback matches starlette's JSONResponse.render exactly.
+            return json.dumps(
+                content,
+                ensure_ascii=False,
+                allow_nan=False,
+                indent=None,
+                separators=(',', ':'),
+            ).encode('utf-8')
+
+    async def request_json(self) -> Any:
+        if not hasattr(self, '_json'):
+            body = await self.body()
+            try:
+                self._json = orjson.loads(body)
+            except (TypeError, ValueError):
+                # Fallback matches starlette's Request.json exactly, including the
+                # json.JSONDecodeError that FastAPI turns into a 422.
+                self._json = json.loads(body)
+        return self._json
+
+    JSONResponse.render = render
+    Request.json = request_json

backend/open_webui/utils/middleware.py

@@ -94,6 +94,7 @@ from open_webui.utils.filter import (
     process_filter_functions,
 )
 
+from open_webui.utils.json_codec import JSONCodec
 from open_webui.utils.mcp.client import MCPClient
 from open_webui.utils.memory import add_memory_context, review_memory_after_turn
 from open_webui.utils.misc import (
@@ -299,7 +300,7 @@ def get_citation_source_from_tool_result(
 
     try:
         try:
-            tool_result = json.loads(tool_result)
+            tool_result = JSONCodec.loads(tool_result)
         except (json.JSONDecodeError, TypeError):
             pass  # keep tool_result as-is (e.g. fetch_url returns plain text)
         if isinstance(tool_result, dict) and 'error' in tool_result:
@@ -1000,7 +1001,7 @@ async def process_tool_result(
                         text = item.get('text', '')
                         if isinstance(text, str):
                             try:
-                                text = json.loads(text)
+                                text = JSONCodec.loads(text)
                             except json.JSONDecodeError:
                                 pass
                         tool_response.append(text)
@@ -1028,7 +1029,7 @@ async def process_tool_result(
                         text = resource.get('text', '')
                         if isinstance(text, str) and text:
                             try:
-                                text = json.loads(text)
+                                text = JSONCodec.loads(text)
                             except json.JSONDecodeError:
                                 pass
                             tool_response.append(text)
@@ -1103,7 +1104,7 @@ async def terminal_event_handler(
         parsed = tool_result
         if isinstance(parsed, str):
             try:
-                parsed = json.loads(parsed)
+                parsed = JSONCodec.loads(parsed)
             except (json.JSONDecodeError, TypeError):
                 pass
         if isinstance(parsed, dict) and parsed.get('exists') is False:
@@ -1141,7 +1142,7 @@ async def chat_completion_tools_handler(
         content = None
         if hasattr(response, 'body_iterator'):
             async for chunk in response.body_iterator:
-                data = json.loads(chunk.decode('utf-8', 'replace'))
+                data = JSONCodec.loads(chunk.decode('utf-8', 'replace'))
                 content = data['choices'][0]['message']['content']
 
             # Cleanup any remaining background tasks if necessary
@@ -1221,7 +1222,7 @@ async def chat_completion_tools_handler(
             if not content:
                 raise Exception('No JSON object found in the response')
 
-            result = json.loads(content)
+            result = JSONCodec.loads(content)
 
             async def tool_call_handler(tool_call):
                 nonlocal skip_files
@@ -1390,7 +1391,7 @@ async def chat_web_search_handler(request: Request, form_data: dict, extra_param
         # user message as the search query.
         if isinstance(res, JSONResponse):
             try:
-                error_body = json.loads(res.body)
+                error_body = JSONCodec.loads(res.body)
                 detail = error_body.get('detail', 'Query generation failed')
             except Exception:
                 detail = 'Query generation failed'
@@ -1406,7 +1407,7 @@ async def chat_web_search_handler(request: Request, form_data: dict, extra_param
                 raise Exception('No JSON object found in the response')
 
             response = response[bracket_start:bracket_end]
-            queries = json.loads(response)
+            queries = JSONCodec.loads(response)
             queries = queries.get('queries', [])
         except Exception as e:
             queries = [response]
@@ -1735,7 +1736,7 @@ async def chat_image_generation_handler(request: Request, form_data: dict, extra
                 # Handle JSONResponse from error paths
                 if isinstance(res, JSONResponse):
                     try:
-                        error_body = json.loads(res.body)
+                        error_body = JSONCodec.loads(res.body)
                         detail = error_body.get('detail', 'Image prompt generation failed')
                     except Exception:
                         detail = 'Image prompt generation failed'
@@ -1751,7 +1752,7 @@ async def chat_image_generation_handler(request: Request, form_data: dict, extra
                         raise Exception('No JSON object found in the response')
 
                     response = response[bracket_start:bracket_end]
-                    response = json.loads(response)
+                    response = JSONCodec.loads(response)
                     prompt = response.get('prompt', [])
                 except Exception as e:
                     prompt = user_message
@@ -1855,7 +1856,7 @@ async def chat_completion_files_handler(
                         raise Exception('No JSON object found in the response')
 
                     queries_response = queries_response[bracket_start:bracket_end]
-                    queries_response = json.loads(queries_response)
+                    queries_response = JSONCodec.loads(queries_response)
                 except Exception as e:
                     queries_response = {'queries': [queries_response]}
 
@@ -1966,7 +1967,7 @@ def apply_params_to_form_data(form_data, model):
             if isinstance(value, str):
                 try:
                     # Attempt to parse the string as JSON
-                    custom_params[key] = json.loads(value)
+                    custom_params[key] = JSONCodec.loads(value)
                 except json.JSONDecodeError:
                     # If it fails, keep the original string
                     pass
@@ -1988,7 +1989,7 @@ def apply_params_to_form_data(form_data, model):
                 logit_bias = convert_logit_bias_input_to_json(params['logit_bias'])
 
                 if logit_bias:
-                    form_data['logit_bias'] = json.loads(logit_bias)
+                    form_data['logit_bias'] = JSONCodec.loads(logit_bias)
             except Exception as e:
                 log.exception(f'Error parsing logit_bias: {e}')
 
@@ -3031,7 +3032,7 @@ def get_response_data(response):
     if isinstance(response, JSONResponse):
         if isinstance(response.body, bytes):
             try:
-                response_data = json.loads(response.body.decode('utf-8', 'replace'))
+                response_data = JSONCodec.loads(response.body.decode('utf-8', 'replace'))
             except json.JSONDecodeError:
                 response_data = {'error': {'detail': 'Invalid JSON response'}}
         else:
@@ -3090,7 +3091,7 @@ def update_assistant_message_from_stream(assistant_message, raw):
             continue
 
         try:
-            data = json.loads(part)
+            data = JSONCodec.loads(part)
         except Exception:
             continue
 
@@ -3278,7 +3279,7 @@ async def background_tasks_handler(ctx):
                     ]
 
                     try:
-                        follow_ups = json.loads(follow_ups_string).get('follow_ups', [])
+                        follow_ups = JSONCodec.loads(follow_ups_string).get('follow_ups', [])
                         await event_emitter(
                             {
                                 'type': 'chat:message:follow_ups',
@@ -3336,7 +3337,7 @@ async def background_tasks_handler(ctx):
                             title_string = title_string[title_string.find('{') : title_string.rfind('}') + 1]
 
                             try:
-                                title = json.loads(title_string).get('title', user_message)
+                                title = JSONCodec.loads(title_string).get('title', user_message)
                             except Exception as e:
                                 title = ''
 
@@ -3388,7 +3389,7 @@ async def background_tasks_handler(ctx):
                         tags_string = tags_string[tags_string.find('{') : tags_string.rfind('}') + 1]
 
                         try:
-                            tags = json.loads(tags_string).get('tags', [])
+                            tags = JSONCodec.loads(tags_string).get('tags', [])
                             await Chats.update_chat_tags_by_id(metadata['chat_id'], tags, user)
 
                             await event_emitter(
@@ -4212,7 +4213,7 @@ async def streaming_chat_response_handler(response, ctx):
                             # (without SSE `data:` prefix). Try to normalize these into standard
                             # error events so frontend and DB paths still receive them.
                             try:
-                                raw_obj = json.loads(data)
+                                raw_obj = JSONCodec.loads(data)
                                 raw_error = raw_obj.get('error') if isinstance(raw_obj, dict) else None
                                 if raw_error:
                                     if save_to_chat:
@@ -4235,7 +4236,7 @@ async def streaming_chat_response_handler(response, ctx):
                         data = data[5:].strip()
 
                         try:
-                            data = json.loads(data)
+                            data = JSONCodec.loads(data)
 
                             if filter_functions:
                                 data, _ = await process_filter_functions(
@@ -4937,7 +4938,7 @@ async def streaming_chat_response_handler(response, ctx):
                         params = {}
                         if tool_args and tool_args.strip():
                             try:
-                                params = json.loads(tool_args)
+                                params = JSONCodec.loads(tool_args)
                             except Exception:
                                 try:
                                     params = ast.literal_eval(tool_args)
@@ -5613,7 +5614,7 @@ async def streaming_chat_response_handler(response, ctx):
                 )
 
                 if event:
-                    yield wrap_item(json.dumps(event))
+                    yield wrap_item(JSONCodec.dumps(event))
 
             async for data in original_generator:
                 data, _ = await process_filter_functions(

backend/open_webui/utils/response.py

@@ -2,6 +2,7 @@ import json
 from numbers import Number
 from uuid import uuid4
 
+from open_webui.utils.json_codec import JSONCodec
 from open_webui.utils.misc import (
     openai_chat_chunk_message_template,
     openai_chat_completion_message_template,
@@ -237,7 +238,7 @@ async def convert_streaming_response_ollama_to_openai(ollama_streaming_response)
     completion_id = f'chatcmpl-{str(uuid4())}'
     first = True
     async for data in ollama_streaming_response.body_iterator:
-        data = json.loads(data)
+        data = JSONCodec.loads(data)
 
         model = data.get('model', 'ollama')
         message = data.get('message') or {}
@@ -268,7 +269,7 @@ async def convert_streaming_response_ollama_to_openai(ollama_streaming_response)
         if done and has_tool_calls:
             data['choices'][0]['finish_reason'] = 'tool_calls'
 
-        line = f'data: {json.dumps(data)}\n\n'
+        line = f'data: {JSONCodec.dumps(data)}\n\n'
         yield line
 
     yield 'data: [DONE]\n\n'

backend/requirements-min.txt

@@ -8,6 +8,7 @@ python-multipart==0.0.32
 itsdangerous==2.2.0
 
 python-socketio==5.16.2
+orjson==3.11.9
 cryptography
 bcrypt==5.0.0
 argon2-cffi==25.1.0

backend/requirements.txt

@@ -5,6 +5,7 @@ python-multipart==0.0.32
 itsdangerous==2.2.0
 
 python-socketio==5.16.2
+orjson==3.11.9
 cryptography==48.0.0
 bcrypt==5.0.0
 argon2-cffi==25.1.0

pyproject.toml

@@ -13,6 +13,7 @@ dependencies = [
     "itsdangerous==2.2.0",
 
     "python-socketio==5.16.2",
+    "orjson==3.11.9",
     "cryptography==48.0.0",
     "bcrypt==5.0.0",
     "argon2-cffi==25.1.0",