GenAI Proxy transcription returns only text even when verbose_json is requested

Products

• VMware Tanzu Platform Core

Issue/Introduction

When calling the OpenAI-compatible transcription API through the GenAI proxy, the response only contains the text field, regardless of the response_format specified in the request.

Expected fields such as segments[], start, end, and other verbose metadata are missing from the response, even when response_format=verbose_json is used. This behavior occurs even if the upstream vLLM backend is verified to be returning the full verbose JSON payload.

Environment

Product: GenAI on Tanzu Platform / AI Services

Cause

The GenAI proxy currently only supports the text response format for transcription services. The mapping logic required to parse and pass through the additional fields provided by verbose_json (such as segment timestamps and speaker metadata) is not yet implemented in the proxy layer.

Resolution

This is a known product limitation at this time. Support for verbose_json and advanced transcription features like speaker diarization are currently under development for a future release of Tanzu AI Services.

Additional Information

There is no configuration change or workaround to force the current version of the GenAI proxy to return verbose metadata.

Users requiring segment-level timestamps or diarization data must wait for a future update to the Tanzu AI Services tile. Please monitor the official Broadcom release notes for Tanzu Platform for the announcement of this feature support.