feat(openai): accept audio_url (and align multimodal parts) for vLLM chat completions

[groundsada]

Summary

When proxying OpenAI-compatible /v1/chat/completions to a vLLM backend that serves Gemma 4 (and similar) multimodal models, the gateway’s request schema validation rejects several audio content shapes that vLLM accepts on a direct connection. Only input_audio (OpenAI-style) is accepted through the proxy today.

Motivation

• vLLM (e.g. Gemma 4 E4B with audio) successfully handles:

  • {"type": "input_audio", "input_audio": {"data": "<base64>", "format": "wav"|"mp3"}} — works through Envoy AI Gateway.
  • {"type": "audio_url", "audio_url": {"url": "<https or data: URI>"}} — works direct to vLLM, but the gateway responds with 400 — malformed request: failed to parse JSON for /v1/chat/completions.
  • Hugging Face–style {"type": "audio", "url": "..."} — vLLM may return 501 / not implemented on the OpenAI path; lower priority unless upstream standardizes it.

• Clients and docs often use audio_url or remote URLs; aligning the gateway with vLLM reduces “works on localhost, fails behind ellm” surprises.

Desired behavior

1. Parse and forward OpenAI-compatible chat messages whose content array includes audio_url parts matching the structure vLLM expects (nested audio_url.url for HTTPS or data:audio/...;base64,...).
2. Optionally document which multimodal part types are supported per route/backend (OpenAI vs provider-specific).
3. Clear 400 messages that distinguish schema validation from true JSON parse errors (if feasible).

Current behavior (observed)

• input_audio + base64: 200 through gateway → vLLM.
• audio_url (public URL or data URI): 400 malformed request: failed to parse JSON at gateway.
• Same payloads 200 when posted directly to vLLM OpenAI server (same model).

Environment (reference)

• Backend: vLLM OpenAI API, Gemma 4 E4B–class model with --limit-mm-per-prompt including audio.
• Gateway: Envoy AI Gateway–style OpenAI route (e.g. NRP ellm).

Non-goals / follow-ups

• Full Hugging Face type: "audio" support is only worth it if the project wants parity with HF chat templates; vLLM’s OpenAI entrypoint may still reject it with 501.

Thank you for considering this — happy to help test or provide minimal JSON fixtures.