Feature Request: Allow custom `base_url` in `chat_google_vertex()` for API gateway proxies

[Tazinho]

Note: Steps so far were run locally and final issue was created with assistance of Opus 4.6

Problem

When using Google Vertex AI models through a corporate API gateway (Azure API Management),
the gateway proxies requests using Google's native API format
(/v1beta1/publishers/google/models/{model}:generateContent) with API key auth
(x-goog-api-key header).

Currently, chat_google_vertex() does not accept a base_url argument, so there's
no way to point it at a custom endpoint. The gateway is NOT OpenAI-compatible, so
chat_openai_compatible() doesn't work either.

Workaround

Using httr2 directly works:

url <- "https://my-gateway.example.com/vertex-ai-express/v1beta1/publishers/google/models/gemini-2.5-flash-lite:generateContent"

resp <- request(url) |>
  req_headers(`x-goog-api-key` = api_key, `Content-Type` = "application/json") |>
  req_body_json(list(
    contents = list(list(role = "user", parts = list(list(text = "Hello!"))))
  )) |>
  req_perform()

The Python google.genai SDK supports this via HttpOptions(base_url=...):

from google.genai import Client
from google.genai.types import HttpOptions

client = Client(vertexai=True, api_key=api_key, http_options=HttpOptions(base_url=ENDPOINT))

Proposed Solution

Add a base_url parameter to chat_google_vertex() that overrides the default Google endpoint URL, similar to how the Python SDK's HttpOptions(base_url=...) works.

This would also need to support API key authentication (via x-goog-api-key header) as an alternative to OAuth/service account credentials.

Environment

• ellmer version: 0.4.1
• R version: [your R version]
• Gateway path format: /v1beta1/publishers/google/models/{model}:{method}
• Auth: x-goog-api-key header with API key

[thisisnic]

Thanks for reporting this @Tazinho, will have a think about how we can look to support this.

In the meantime, Claude suggested this temporary workaround which uses the Google Gemini rather than Vertex code path, but with the caveat that while basic chatting could work, some functionality like model listing and pricing might be incorrect, so I'm not recommending it for anything beyond very basic experimentation.

chat <- chat_google_gemini(
    base_url = "https://my-gateway.example.com/vertex-ai-express/v1beta1/publishers/google/",                                                                                                 
    credentials = \() Sys.getenv("MY_API_KEY"),
    model = "gemini-2.5-flash-lite"                                                                                                                                                           
  )