fix(logging): keep a streamed transcript out of callbacks that turned off message logging

A streamed transcription's assembled response reached success callbacks with its text intact under turn_off_message_logging or a callback's message_logging=False, because the shared streaming redactor had no transcription branch. It now redacts the text, as the unstreamed call already does
This commit is contained in:
mateo-berri 2026-11-02 17:46:37 -08:00
parent f589986f98
commit 336231ed75
2 changed files with 43 additions and 0 deletions

View file

@ -171,6 +171,8 @@ def _redact_streaming_response(streaming_response):
_redact_bedrock_guardrail_trace(streaming_response)
elif isinstance(streaming_response, litellm.ResponsesAPIResponse):
return perform_redaction({}, streaming_response)
elif isinstance(streaming_response, litellm.TranscriptionResponse):
streaming_response.text = REDACTED_BY_LITELLM
elif hasattr(streaming_response, "output"):
_redact_responses_api_output(streaming_response.output)
if hasattr(streaming_response, "reasoning") and streaming_response.reasoning is not None:

View file

@ -19,6 +19,7 @@ from litellm.litellm_core_utils.audio_utils.transcription_streaming import (
litellm_exception_mapper,
wrap_transcription_stream,
)
from litellm.litellm_core_utils.redact_messages import REDACTED_BY_LITELLM
from litellm.llms.azure.audio_transcriptions import AzureAudioTranscription
from litellm.llms.custom_httpx.http_handler import HTTPHandler
from litellm.llms.openai.transcriptions.gpt_transformation import (
@ -724,6 +725,46 @@ async def test_atranscription_stream_under_a_post_call_guardrail_streams_the_gua
]
class _StreamedTranscriptRecorder(CustomLogger):
def __init__(self, message_logging: bool) -> None:
super().__init__(message_logging=message_logging)
self.logged: Final[asyncio.Queue[frozenset[object]]] = asyncio.Queue()
async def async_log_success_event(self, kwargs, response_obj, start_time, end_time) -> None:
logged_responses: Final = (
response_obj,
*(kwargs.get(key) for key in ("complete_streaming_response", "async_complete_streaming_response")),
)
await self.logged.put(frozenset(response.text for response in logged_responses if response is not None))
@pytest.mark.asyncio
@pytest.mark.parametrize("opt_out", ("global", "callback"))
async def test_atranscription_stream_keeps_the_transcript_out_of_callbacks_that_turned_off_message_logging(
local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch, opt_out: str
) -> None:
async def send_response(request: httpx.Request) -> httpx.Response:
return _duration_priced_transcript(streamed=_asks_for_stream(await request.aread()))
recorder: Final = _StreamedTranscriptRecorder(message_logging=opt_out != "callback")
monkeypatch.setattr(litellm, "callbacks", [recorder])
monkeypatch.setattr(litellm, "turn_off_message_logging", opt_out == "global")
client: Final = AsyncOpenAI(
api_key="sk-test",
base_url="https://example.com/v1",
http_client=httpx.AsyncClient(transport=httpx.MockTransport(send_response)),
)
stream: Final = await litellm.atranscription(
model="openai/gpt-transcribe", file=_one_second_wav(), stream=True, client=client
)
streamed_text: Final = [event.delta async for event in stream if event.type == "transcript.text.delta"]
logged_texts: Final = await asyncio.wait_for(recorder.logged.get(), timeout=5)
await client.close()
assert streamed_text == ["hello world"]
assert logged_texts == {REDACTED_BY_LITELLM}
def test_transcription_stream_bills_the_provider_duration_like_the_unstreamed_call(
local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch
) -> None: