Freshservice attachment Procedures prepare bounded image observations and voicemail transcripts for
triage while retaining video without analysis. New audio processing uses
openai/gpt-transcribe
through the existing approved OpenAI connection, with provider duration accounting before transcript
release. Explicit legacy Whisper pins keep their model and pricing.Usable attachment observations can inform Priority even when the captured ticket message has no
assessable text. A deterministic transcript-availability output supports voicemail routing without
rewriting ticket facts. Unprocessed files and incomplete capture remain visible as incomplete
evidence. Deployment supplies these versioned Procedures; workflow activation and live media
qualification remain separate operator steps.Provider-truncated image text stays explicitly incomplete, and model refusals do not become
successful observations. Paid usage remains accounted for before either result is released.Rejected OpenAI audio requests now log their HTTP status, selected model, validated request ID
and recognized error classifications. Diagnostics exclude provider messages, audio content and
credentials. Error-body inspection has strict size and time limits. Existing attachment failure
handling and usage accounting remain unchanged.

