diff --git a/backend/open_webui/config.py b/backend/open_webui/config.py
index 116ac77ade..6dfec3c42e 100644
--- a/backend/open_webui/config.py
+++ b/backend/open_webui/config.py
@@ -2143,6 +2143,10 @@ CONTEXT_COMPACTION_TOKEN_THRESHOLD = int(os.getenv('CONTEXT_COMPACTION_TOKEN_THR
_CONTEXT_COMPACTION_TOKEN_CAP = os.getenv('CONTEXT_COMPACTION_TOKEN_CAP')
CONTEXT_COMPACTION_TOKEN_CAP = int(_CONTEXT_COMPACTION_TOKEN_CAP) if _CONTEXT_COMPACTION_TOKEN_CAP else None
+CONTEXT_COMPACTION_RETENTION_PERCENTAGE = min(
+ 50, max(10, int(os.getenv('CONTEXT_COMPACTION_RETENTION_PERCENTAGE', '40')))
+)
+
CONTEXT_COMPACTION_PROMPT_TEMPLATE = os.getenv('CONTEXT_COMPACTION_PROMPT_TEMPLATE', '')
TITLE_GENERATION_PROMPT_TEMPLATE = os.getenv('TITLE_GENERATION_PROMPT_TEMPLATE', '')
@@ -3054,6 +3058,7 @@ DEFAULT_CONFIG = {
'chat.context_compaction.enable': ENABLE_CONTEXT_COMPACTION,
'chat.context_compaction.token_threshold': CONTEXT_COMPACTION_TOKEN_THRESHOLD,
'chat.context_compaction.token_cap': CONTEXT_COMPACTION_TOKEN_CAP,
+ 'chat.context_compaction.retention_percentage': CONTEXT_COMPACTION_RETENTION_PERCENTAGE,
'chat.context_compaction.prompt_template': CONTEXT_COMPACTION_PROMPT_TEMPLATE,
'task.title.prompt_template': TITLE_GENERATION_PROMPT_TEMPLATE,
'task.tags.prompt_template': TAGS_GENERATION_PROMPT_TEMPLATE,
diff --git a/backend/open_webui/models/chats.py b/backend/open_webui/models/chats.py
index ac2844149b..b2e252f588 100644
--- a/backend/open_webui/models/chats.py
+++ b/backend/open_webui/models/chats.py
@@ -13,7 +13,7 @@ from open_webui.models.automations import AutomationRun
from open_webui.models.chat_messages import ChatMessage, ChatMessages
from open_webui.models.folders import Folders
from open_webui.models.tags import Tag, TagModel, Tags
-from open_webui.utils.misc import sanitize_data_for_db, sanitize_text_for_db
+from open_webui.utils.misc import get_output_text, sanitize_data_for_db, sanitize_text_for_db
from pydantic import BaseModel, ConfigDict, field_validator
from sqlalchemy import (
JSON,
@@ -840,6 +840,11 @@ class ChatTable:
if chat is None:
return None
+ if not message.get('content'):
+ output_text = get_output_text(message.get('output'))
+ if output_text:
+ message['content'] = output_text
+
# Sanitize message content for null characters before upserting
if isinstance(message.get('content'), str):
message['content'] = sanitize_text_for_db(message['content'])
diff --git a/backend/open_webui/retrieval/utils.py b/backend/open_webui/retrieval/utils.py
index 2a6ac5c3e1..b402c12e03 100644
--- a/backend/open_webui/retrieval/utils.py
+++ b/backend/open_webui/retrieval/utils.py
@@ -51,7 +51,7 @@ from open_webui.retrieval.web.utils import get_web_loader
from open_webui.utils.access_control.files import has_access_to_file
from open_webui.utils.access_control.folders import has_folder_access
from open_webui.utils.headers import include_user_info_headers
-from open_webui.utils.misc import get_message_list
+from open_webui.utils.misc import get_content_from_message, get_message_list
log = logging.getLogger(__name__)
@@ -1409,7 +1409,10 @@ async def get_sources_from_items(
# Reconstruct the message list in order
message_list = get_message_list(messages_map, message_id)
message_history = '\n'.join(
- [f'#### {m.get("role", "user").capitalize()}\n{m.get("content")}\n' for m in message_list]
+ [
+ f'#### {m.get("role", "user").capitalize()}\n{get_content_from_message(m) or ""}\n'
+ for m in message_list
+ ]
)
# User has access to the chat
diff --git a/backend/open_webui/routers/chats.py b/backend/open_webui/routers/chats.py
index a136d3679d..1a707607e0 100644
--- a/backend/open_webui/routers/chats.py
+++ b/backend/open_webui/routers/chats.py
@@ -55,6 +55,7 @@ CHAT_CONFIG_KEYS = {
'ENABLE_CONTEXT_COMPACTION': 'chat.context_compaction.enable',
'CONTEXT_COMPACTION_TOKEN_THRESHOLD': 'chat.context_compaction.token_threshold',
'CONTEXT_COMPACTION_TOKEN_CAP': 'chat.context_compaction.token_cap',
+ 'CONTEXT_COMPACTION_RETENTION_PERCENTAGE': 'chat.context_compaction.retention_percentage',
'CONTEXT_COMPACTION_PROMPT_TEMPLATE': 'chat.context_compaction.prompt_template',
}
@@ -76,6 +77,7 @@ class ChatConfigForm(BaseModel):
ENABLE_CONTEXT_COMPACTION: bool
CONTEXT_COMPACTION_TOKEN_THRESHOLD: int
CONTEXT_COMPACTION_TOKEN_CAP: int | None = None
+ CONTEXT_COMPACTION_RETENTION_PERCENTAGE: int = 40
CONTEXT_COMPACTION_PROMPT_TEMPLATE: str
@@ -125,6 +127,8 @@ async def get_chat_config_values() -> dict:
config = {field: values[storage_key] for field, storage_key in CHAT_CONFIG_KEYS.items() if storage_key in values}
if config.get('CONTEXT_COMPACTION_TOKEN_CAP') is None:
config['CONTEXT_COMPACTION_TOKEN_CAP'] = config.get('CONTEXT_COMPACTION_TOKEN_THRESHOLD', 80000)
+ if config.get('CONTEXT_COMPACTION_RETENTION_PERCENTAGE') is None:
+ config['CONTEXT_COMPACTION_RETENTION_PERCENTAGE'] = 40
return config
@@ -738,12 +742,14 @@ async def get_chat_config(user=Depends(get_admin_user)):
async def set_chat_config(form_data: ChatConfigForm, user=Depends(get_admin_user)):
threshold = max(1, int(form_data.CONTEXT_COMPACTION_TOKEN_THRESHOLD))
token_cap = max(1, int(form_data.CONTEXT_COMPACTION_TOKEN_CAP or threshold))
+ retention_percentage = min(50, max(10, int(form_data.CONTEXT_COMPACTION_RETENTION_PERCENTAGE)))
await Config.upsert(
chat_config_updates(
{
**form_data.model_dump(),
'CONTEXT_COMPACTION_TOKEN_THRESHOLD': threshold,
'CONTEXT_COMPACTION_TOKEN_CAP': token_cap,
+ 'CONTEXT_COMPACTION_RETENTION_PERCENTAGE': retention_percentage,
}
)
)
diff --git a/backend/open_webui/utils/context_compaction.py b/backend/open_webui/utils/context_compaction.py
index 910fa5b7b6..1a9b5e9f96 100644
--- a/backend/open_webui/utils/context_compaction.py
+++ b/backend/open_webui/utils/context_compaction.py
@@ -56,7 +56,7 @@ async def compact_messages_for_request(
if not _exceeds_token_threshold(messages, system_prompt, previous_summary, token_threshold) or len(messages) <= 3:
return messages, previous_summary, False
- boundary = _find_compaction_boundary(messages)
+ boundary = _find_compaction_boundary(messages, config['retention_percentage'])
compacted_messages = messages[:boundary]
recent_messages = messages[boundary:]
if not compacted_messages or not recent_messages:
@@ -196,6 +196,7 @@ async def _load_config() -> dict:
'chat.context_compaction.enable',
'chat.context_compaction.token_threshold',
'chat.context_compaction.token_cap',
+ 'chat.context_compaction.retention_percentage',
'chat.context_compaction.prompt_template',
)
token_threshold = _parse_positive_int(values.get('chat.context_compaction.token_threshold')) or 80000
@@ -203,6 +204,9 @@ async def _load_config() -> dict:
'enable': bool(values.get('chat.context_compaction.enable', False)),
'token_threshold': token_threshold,
'token_cap': _parse_positive_int(values.get('chat.context_compaction.token_cap')) or token_threshold,
+ 'retention_percentage': _clamp_retention_percentage(
+ values.get('chat.context_compaction.retention_percentage')
+ ),
'prompt_template': values.get('chat.context_compaction.prompt_template', '') or '',
}
@@ -215,6 +219,14 @@ def _parse_positive_int(value: Any) -> int | None:
return parsed if parsed > 0 else None
+def _clamp_retention_percentage(value: Any) -> int:
+ try:
+ parsed = int(value)
+ except (TypeError, ValueError):
+ parsed = 40
+ return min(50, max(10, parsed))
+
+
def _resolve_token_threshold(global_threshold: int, global_cap: int, metadata: dict) -> int:
configured_threshold = _parse_positive_int((metadata.get('params') or {}).get('compact_token_threshold'))
return min(configured_threshold or global_threshold, global_cap)
@@ -297,8 +309,9 @@ def _exceeds_token_threshold(messages: list[dict], system_prompt: str, summary:
return estimated > threshold
-def _find_compaction_boundary(messages: list[dict]) -> int:
- keep_count = max(2, len(messages) * 2 // 5)
+def _find_compaction_boundary(messages: list[dict], retention_percentage: int = 40) -> int:
+ retention_percentage = _clamp_retention_percentage(retention_percentage)
+ keep_count = max(2, len(messages) * retention_percentage // 100)
target = max(1, len(messages) - keep_count)
boundaries = [idx for idx, message in enumerate(messages) if message.get('role') == 'user'][1:]
return next((idx for idx in reversed(boundaries) if idx <= target), 0)
diff --git a/backend/open_webui/utils/middleware.py b/backend/open_webui/utils/middleware.py
index d7f00959a5..72e32d68b8 100644
--- a/backend/open_webui/utils/middleware.py
+++ b/backend/open_webui/utils/middleware.py
@@ -107,6 +107,7 @@ from open_webui.utils.misc import (
get_last_user_message,
get_last_user_message_item,
get_message_list,
+ get_output_text,
get_system_message,
is_string_allowed,
merge_system_messages,
@@ -137,17 +138,6 @@ logging.basicConfig(stream=sys.stdout, level=GLOBAL_LOG_LEVEL)
log = logging.getLogger(__name__)
-def _last_output_text(output: list | None) -> str:
- for item in reversed(output or []):
- if item.get('type') != 'message':
- continue
- parts = item.get('content') or []
- text = ''.join(str(part.get('text') or '') for part in parts if part.get('type') == 'output_text')
- if text:
- return text
- return ''
-
-
async def publish_chat_finished_event(
request: Request, user: UserModel, metadata: dict, title: str, content: str, output: list | None = None
):
@@ -155,7 +145,7 @@ async def publish_chat_finished_event(
if getattr(request.state, 'internal', False) is True or not chat_id or chat_id.startswith(('channel:', 'local:')):
return
- content = content or _last_output_text(output)
+ content = content or get_output_text(output)
webui_url = await Config.get('webui.url')
await publish_event(
request,
@@ -3398,7 +3388,7 @@ async def outlet_filter_handler(ctx):
message_list = [
{
'role': m.get('role'),
- 'content': m.get('content', ''),
+ 'content': m.get('content') or get_output_text(m.get('output')),
}
for m in form_messages
]
@@ -3431,7 +3421,7 @@ async def outlet_filter_handler(ctx):
{
'id': m.get('id'),
'role': m.get('role'),
- 'content': m.get('content', ''),
+ 'content': m.get('content') or get_output_text(m.get('output')),
'info': m.get('info'),
'timestamp': m.get('timestamp'),
**({'output': m['output']} if m.get('output') else {}),
@@ -3483,17 +3473,21 @@ async def outlet_filter_handler(ctx):
outlet_message_id = message.get('id')
if outlet_message_id and outlet_message_id in messages_map:
original_message = messages_map[outlet_message_id]
- content_changed = original_message.get('content') != message.get('content')
+ original_content = original_message.get('content') or get_output_text(
+ original_message.get('output')
+ )
+ message_content = message.get('content') or get_output_text(message.get('output'))
+ content_changed = original_content != message_content
output_changed = message.get('output') and message.get('output') != original_message.get(
'output'
)
if content_changed or output_changed:
message_update = {
- 'originalContent': original_message.get('content'),
+ 'originalContent': original_content,
**({'output': message['output']} if output_changed else {}),
}
if content_changed:
- message_update['content'] = message.get('content', '')
+ message_update['content'] = message_content or ''
await Chats.upsert_message_to_chat_by_id_and_message_id(
chat_id,
outlet_message_id,
diff --git a/backend/open_webui/utils/misc.py b/backend/open_webui/utils/misc.py
index ead329b278..074f7c859b 100644
--- a/backend/open_webui/utils/misc.py
+++ b/backend/open_webui/utils/misc.py
@@ -160,13 +160,40 @@ def get_last_user_message_item(messages: list[dict]) -> dict | None:
def get_content_from_message(message: dict) -> str | None:
- if isinstance(message.get('content'), list):
- for item in message['content']:
- if item['type'] == 'text':
- return item['text']
- else:
- return message.get('content')
- return None
+ content = message.get('content')
+ if isinstance(content, list):
+ for item in content:
+ if isinstance(item, dict) and item.get('type') == 'text':
+ return item.get('text')
+ elif content:
+ return content
+
+ output_text = get_output_text(message.get('output'))
+ return output_text or (content if isinstance(content, str) else None)
+
+
+def get_output_text(output: list | None) -> str:
+ if not isinstance(output, list):
+ return ''
+
+ texts = []
+ for item in output:
+ if not isinstance(item, dict) or item.get('type') != 'message':
+ continue
+
+ parts = item.get('content') or []
+ if not isinstance(parts, list):
+ continue
+
+ text = ''.join(
+ str(part.get('text'))
+ for part in parts
+ if isinstance(part, dict) and part.get('text') is not None
+ )
+ if text.strip():
+ texts.append(text)
+
+ return '\n'.join(texts)
def reconcile_tool_pairs(messages: list[dict]) -> list[dict]:
@@ -213,7 +240,7 @@ def reconcile_tool_pairs(messages: list[dict]) -> list[dict]:
# All tool_calls were orphans — keep the message only if it
# carries meaningful text or reasoning content.
- content = message.get('content', '')
+ content = get_content_from_message(message) or ''
has_meaningful_content = content.strip() if isinstance(content, str) else content
if has_meaningful_content or message.get('reasoning_content'):
reconciled_messages.append({key: value for key, value in message.items() if key != 'tool_calls'})
diff --git a/src/lib/components/admin/Settings/Interface.svelte b/src/lib/components/admin/Settings/Interface.svelte
index 5d56aee356..197491e637 100644
--- a/src/lib/components/admin/Settings/Interface.svelte
+++ b/src/lib/components/admin/Settings/Interface.svelte
@@ -44,6 +44,7 @@
ENABLE_CONTEXT_COMPACTION: false,
CONTEXT_COMPACTION_TOKEN_THRESHOLD: 80000,
CONTEXT_COMPACTION_TOKEN_CAP: 80000,
+ CONTEXT_COMPACTION_RETENTION_PERCENTAGE: 40,
CONTEXT_COMPACTION_PROMPT_TEMPLATE: ''
};
@@ -273,6 +274,22 @@
/>
+
+
+
+