What's new
PLACEHOLDER
Commits details since the last release
- [0351793] by @thiswillbeyourgithub, 33 minutes ago:
bump version 5.1.3 -> 5.2.0
bumpver.toml
docs/source/conf.py
setup.py
wdoc/wdoc.py
- [41d06a6] by @thiswillbeyourgithub, 83 seconds ago:
ignore serena folder
Signed-off-by: thiswillbeyourgithub 26625900+thiswillbeyourgithub@users.noreply.github.com
.gitignore
- [44b085b] by @thiswillbeyourgithub, 43 minutes ago:
fix(local_audio): encode unsilenced audio as mp3 instead of ogg
When audio_unsilence is on (the default), the silence-removal path re-encoded
the audio to .ogg and uploaded that to the whisper endpoint. Some
OpenAI-compatible whisper servers reject audio/ogg with a 415 Unsupported
Media Type. mp3 is the most widely supported format for these endpoints, and
it is already what local_video.py and online_media.py produce, so this also
makes the three audio-producing loaders consistent.
Adds a regression test asserting the file handed to transcribe_audio_whisper
is .mp3 (never .ogg). It mocks torchaudio/ffmpeg/file_hasher and the
transcriber, and uses a unique file_hash to bypass the joblib doc-loaders
cache so it stays reproducible.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
tests/test_wdoc.py
wdoc/utils/loaders/local_audio.py
- [0267b9d] by @thiswillbeyourgithub, 55 minutes ago:
fix(whisper): rewind audio file and preserve real error in requests fallback
When litellm.transcription fails and wdoc falls back to a direct requests
call, it reused the same file handle that litellm had already read to EOF, so
the fallback uploaded an empty body. Endpoints answer an empty upload with a
misleading status (e.g. 415 Unsupported Media Type) that has nothing to do
with the actual failure, which is likely what produced the confusing 415 seen
behind an auth proxy. Rewind the handle with seek(0) before the fallback POST.
Also, the original litellm error was only logged as a warning and then lost
when the fallback failed. Wrap the fallback request so its failure re-raises
with BOTH the endpoint error and the original litellm error, so the true root
cause (often an auth/proxy misconfiguration) survives.
Adds two regression tests (call the undecorated function to bypass the joblib
doc-loaders cache): one asserts the fallback uploads the full file, one
asserts the raised error mentions both the endpoint and litellm errors.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
tests/test_wdoc.py
wdoc/utils/loaders/shared_audio.py
- [c0a62c3] by @thiswillbeyourgithub, 60 minutes ago:
fix(loaders): stop stacking three tracebacks on a failed doc load
When load_one_doc failed (e.g. a single audio file whose whisper endpoint
returns "415 Unsupported Media Type"), the error wrapper embedded the full
traceback in the log message, then re-emitted it via logger.exception, then
raised (excepthook printing it a third time). The real one-line error was
buried under three near-identical tracebacks, which was misleading.
Now the message stays concise and the traceback is shown exactly once:
attached via exc_info on the warn path, or printed by the re-raise on the
crash paths. Also drop the now-unused traceback import.
Adds a regression test asserting the logged message is concise (no embedded
"Full traceback:" dump) and that crash mode does not double-attach the
traceback to the log record.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
tests/test_wdoc.py
wdoc/utils/loaders/init.py
