MLE-5322: docs(audio): clarify prompt is Whisper-only#270
MLE-5322: docs(audio): clarify prompt is Whisper-only#270rishabh-bhargava merged 1 commit intomainfrom
Conversation
Both /audio/transcriptions and /audio/translations expose `prompt`, but it only takes effect on Whisper-family models. Parakeet silently ignores it. Update the field description so users don't copy-paste Whisper code with prompt and expect biasing on Parakeet.
✱ Stainless preview buildsThis PR will update the go openapi python terraform typescript
|
sbeurnier
left a comment
There was a problem hiding this comment.
theres a limit to the tokens for the promot, anything past that and we get silent 200 empty transcript. on our end we should fix to truncate user-passed prompt to fit within the limit but we should mention in docs the existence of such a limit (we can also rebuild engine to increase the limit beyond what it currently is)
prompt is not working correctly anyway. I think we need to do that separately, here I am simply trying to clarify that it does not apply to Parakeet |
|
can we remove the prompt field entirely? It doesnt work well with whisper, we dont receomment people using it. I don't feel the point of having this field |
|
It's supposed to work. Would like to keep this around in case Sebastien is able to figure out a fix.
Maybe we time box this and if we can't, then we get rid.
…On Thu, 7 May 2026 at 15:13, Sahil Yadav < ***@***.*** > wrote:
*yadavsahil197* left a comment (togethercomputer/openapi#270) (
#270 (comment)
)
can we remove the prompt field entirely? It doesnt work well with whisper,
we dont receomment people using it. I don't feel the point of having this
field
—
Reply to this email directly, view it on GitHub (
#270 (comment)
) , or unsubscribe (
https://github.com/notifications/unsubscribe-auth/AAL63CONX4RMVT52BFTVBED4ZUDALAVCNFSM6AAAAACYVGJYWOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DIMBRGU3DIOJRHE
).
You are receiving this because you authored the thread. Message ID: <togethercomputer/openapi/pull/270/c4401564919
@ github. com>
|
Summary
Updates the
promptfield description on bothAudioTranscriptionRequestandAudioTranslationRequestto call out that it is supported only on Whisper-family models. Other STT models (Parakeet) accept the field for API compatibility but ignore it.Fixes MLE-5322.
Why
promptis a Whisper-style decoder-priming hint. Parakeet is a CTC/Transducer model with no slot for it —tvllm/inference_worker/parakeet_engine.pyreadsparams.request_databut never accessesprompt. Behavioral test confirms: same audio +temperature=0, with vs withoutprompt, returns byte-identical text on Parakeet and dramatically different text on Whisper.Test plan
mintlify-docs/openapi.yaml).🤖 Generated with Claude Code