scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-07 08:26:10 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 36

feat: Redact messages in spend logs' proxy_server_request

Browse source 
This change ensures that sensitive message content within the `proxy_server_request` field of spend logs is redacted when message logging is turned off or when the redaction header is used. This enhances privacy and security by preventing the logging of potentially sensitive user inputs in the request details.

Co-authored-by: ishaan <ishaan@berri.ai>
This commit is contained in:
Cursor Agent 2026-01-09 22:06:34 +00:00
parent 0c7db97ad5
commit e317d20573
2 changed files with 276 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

52
litellm/proxy/spend_tracking/spend_tracking_utils.py
View file

@ -435,7 +435,7 @@ def get_logging_payload( # noqa: PLR0915
),
response=_get_response_for_spend_logs_payload(standard_logging_payload),
proxy_server_request=_get_proxy_server_request_for_spend_logs_payload(
metadata=metadata, litellm_params=litellm_params
metadata=metadata, litellm_params=litellm_params, kwargs=kwargs
),
session_id=_get_session_id_for_spend_log(
kwargs=kwargs,
@ -639,12 +639,47 @@ def _sanitize_request_body_for_spend_logs_payload(
return {k: _sanitize_value(v) for k, v in request_body.items()}
def _redact_messages_in_request_body(request_body: dict) -> dict:
"""
Redact messages in the request body.
This handles various request formats (chat completions, completions, etc.)
and redacts the user-provided content.
"""
import copy
# Create a shallow copy to avoid modifying the original
redacted_body = copy.copy(request_body)
# Redact messages in chat completion requests
if "messages" in redacted_body and isinstance(redacted_body["messages"], list):
redacted_body["messages"] = [
{"role": "user", "content": REDACTED_BY_LITELM_STRING}
]
# Redact prompt in completion requests
if "prompt" in redacted_body:
redacted_body["prompt"] = REDACTED_BY_LITELM_STRING
# Redact input in other request types (embeddings, etc.)
if "input" in redacted_body:
if isinstance(redacted_body["input"], str):
redacted_body["input"] = REDACTED_BY_LITELM_STRING
elif isinstance(redacted_body["input"], list):
redacted_body["input"] = [REDACTED_BY_LITELM_STRING]
return redacted_body
def _get_proxy_server_request_for_spend_logs_payload(
metadata: dict,
litellm_params: dict,
kwargs: Optional[dict] = None,
) -> str:
"""
Only store if _should_store_prompts_and_responses_in_spend_logs() is True
If turn_off_message_logging is enabled, redact messages in the request body.
"""
if _should_store_prompts_and_responses_in_spend_logs():
_proxy_server_request = cast(
@ -652,6 +687,21 @@ def _get_proxy_server_request_for_spend_logs_payload(
)
if _proxy_server_request is not None:
_request_body = _proxy_server_request.get("body", {}) or {}
# Check if message redaction should be applied
if kwargs is not None:
from litellm.litellm_core_utils.redact_messages import should_redact_message_logging
# Build model_call_details dict to check redaction settings
model_call_details = {
"litellm_params": litellm_params,
"standard_callback_dynamic_params": kwargs.get("standard_callback_dynamic_params"),
}
# If redaction is enabled, redact messages in the request body
if should_redact_message_logging(model_call_details):
_request_body = _redact_messages_in_request_body(_request_body)
_request_body = _sanitize_request_body_for_spend_logs_payload(_request_body)
_request_body_json_str = json.dumps(_request_body, default=str)
return _request_body_json_str

225
tests/logging_callback_tests/test_spend_logs.py
View file

@ -549,3 +549,228 @@ def test_truncation_preserves_beginning_and_end():
kept_chars = expected_start_chars + expected_end_chars
expected_skipped = total_chars - kept_chars
assert str(expected_skipped) in truncated_content
@pytest.mark.parametrize("turn_off_message_logging", [True, False])
def test_proxy_server_request_redaction(turn_off_message_logging: bool):
"""
Test that proxy_server_request messages are redacted when turn_off_message_logging is enabled.
This test verifies the fix for GitHub issue #16336:
https://github.com/BerriAI/litellm/issues/16336
When turn_off_message_logging is True, both the response AND the proxy_server_request
should be redacted in the spend logs.
"""
from litellm.constants import REDACTED_BY_LITELM_STRING
# Set up litellm global setting
original_setting = litellm.turn_off_message_logging
litellm.turn_off_message_logging = turn_off_message_logging
try:
# Mock request data
request_messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "What is the capital of France?"}
]
# Mock response data
response_obj = {
"id": "chatcmpl-test",
"object": "chat.completion",
"created": 1234567890,
"model": "gpt-4.1-mini",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "The capital of France is Paris."
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 20,
"completion_tokens": 10,
"total_tokens": 30
}
}
# Build kwargs with proxy_server_request
kwargs = {
"model": "gpt-4.1-mini",
"messages": request_messages,
"litellm_params": {
"metadata": {
"user_api_key": "sk-test-key",
"user_api_key_user_id": "test-user",
},
"proxy_server_request": {
"method": "POST",
"url": "http://localhost:4000/chat/completions",
"body": {
"model": "gpt-4.1-mini",
"messages": request_messages,
"temperature": 0.7
}
}
},
"call_type": "acompletion",
"response_cost": 0.0001,
}
# Add dynamic params for per-request override (if needed)
if turn_off_message_logging:
kwargs["standard_callback_dynamic_params"] = {
"turn_off_message_logging": turn_off_message_logging
}
# Generate spend logs payload
import datetime
start_time = datetime.datetime.now()
end_time = start_time + datetime.timedelta(seconds=1)
# Set store_prompts_in_spend_logs to True so proxy_server_request is stored
from litellm.proxy.proxy_server import general_settings
original_store_prompts = general_settings.get("store_prompts_in_spend_logs", False)
general_settings["store_prompts_in_spend_logs"] = True
try:
payload = get_logging_payload(
kwargs=kwargs,
response_obj=response_obj,
start_time=start_time,
end_time=end_time
)
# Parse the proxy_server_request JSON string
proxy_request_str = payload.get("proxy_server_request", "{}")
proxy_request = json.loads(proxy_request_str) if proxy_request_str != "{}" else {}
if turn_off_message_logging:
# When redaction is enabled, messages should be redacted
if proxy_request:
assert "messages" in proxy_request
assert len(proxy_request["messages"]) == 1
assert proxy_request["messages"][0]["role"] == "user"
assert proxy_request["messages"][0]["content"] == REDACTED_BY_LITELM_STRING
else:
# When redaction is disabled, messages should be preserved
if proxy_request:
assert "messages" in proxy_request
assert len(proxy_request["messages"]) == 2
assert proxy_request["messages"][0]["content"] == "You are a helpful assistant."
assert proxy_request["messages"][1]["content"] == "What is the capital of France?"
finally:
# Restore original setting
general_settings["store_prompts_in_spend_logs"] = original_store_prompts
finally:
# Restore original litellm setting
litellm.turn_off_message_logging = original_setting
def test_proxy_server_request_redaction_with_header():
"""
Test that proxy_server_request is redacted when x-litellm-enable-message-redaction header is set.
This test verifies that per-request redaction via header also works for proxy_server_request.
"""
from litellm.constants import REDACTED_BY_LITELM_STRING
# Set global setting to False
original_setting = litellm.turn_off_message_logging
litellm.turn_off_message_logging = False
try:
# Mock request data
request_messages = [
{"role": "user", "content": "Sensitive data that should be redacted"}
]
# Mock response data
response_obj = {
"id": "chatcmpl-test",
"object": "chat.completion",
"created": 1234567890,
"model": "gpt-4.1-mini",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Response content"
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 10,
"completion_tokens": 5,
"total_tokens": 15
}
}
# Build kwargs with proxy_server_request and redaction header
kwargs = {
"model": "gpt-4.1-mini",
"messages": request_messages,
"litellm_params": {
"metadata": {
"user_api_key": "sk-test-key",
"headers": {
"x-litellm-enable-message-redaction": "true"
}
},
"proxy_server_request": {
"method": "POST",
"url": "http://localhost:4000/chat/completions",
"body": {
"model": "gpt-4.1-mini",
"messages": request_messages,
}
}
},
"call_type": "acompletion",
"response_cost": 0.0001,
}
# Generate spend logs payload
import datetime
start_time = datetime.datetime.now()
end_time = start_time + datetime.timedelta(seconds=1)
# Set store_prompts_in_spend_logs to True
from litellm.proxy.proxy_server import general_settings
original_store_prompts = general_settings.get("store_prompts_in_spend_logs", False)
general_settings["store_prompts_in_spend_logs"] = True
try:
payload = get_logging_payload(
kwargs=kwargs,
response_obj=response_obj,
start_time=start_time,
end_time=end_time
)
# Parse the proxy_server_request JSON string
proxy_request_str = payload.get("proxy_server_request", "{}")
proxy_request = json.loads(proxy_request_str) if proxy_request_str != "{}" else {}
# Messages should be redacted due to the header
if proxy_request:
assert "messages" in proxy_request
assert len(proxy_request["messages"]) == 1
assert proxy_request["messages"][0]["role"] == "user"
assert proxy_request["messages"][0]["content"] == REDACTED_BY_LITELM_STRING
finally:
# Restore original setting
general_settings["store_prompts_in_spend_logs"] = original_store_prompts
finally:
# Restore original litellm setting
litellm.turn_off_message_logging = original_setting