Vertex AI provider
A model provider for Vertex AI. Named in the architecture notes; no code exists.
PlannedNot implemented. Nothing described here exists in the current code.
The architecture notes describe four providers — Anthropic, OpenAI-compatible, Vertex, and vLLM. Three of the four ship; Vertex is the one that does not.
What exists today: src/beam_agents/model/ contains anthropic.py,
openai_compat.py, vllm.py, and fake.py, plus the client, facade, and
replay cache. vLLM is covered both as a served endpoint
(VllmEndpointProvider) and as a GPU-worker sidecar (VllmSidecarProvider,
engine installed from the vllm extra) — see
the install page's extras table. There is no vertex.py.
Until it exists, a Vertex-hosted model is reachable only where Vertex exposes
an OpenAI-compatible surface the existing openai_compat provider can speak
to — a workaround using an existing provider, not the described integration,
and it comes with none of Vertex's native auth or endpoint handling.