rag.ask_contexts() cannot succeed: backend returns UNIMPLEMENTED on v1 and a generic INVALID_ARGUMENT on v1beta1. Is AskContexts allowlisted?

Bắt đầu bằng cách đọc phần chuyển đổi yêu cầu AskContexts trong agentplatform._genai.rag và so sánh với đường dẫn retrieve_contexts; tái hiện các phản hồi v1beta1 và v1 được mô tả trong issue. Công việc được xem là hoàn tất khi trạng thái allowlisting của dịch vụ và hành vi SDK dự kiến đã được xác định, đồng thời mọi thay đổi được đề xuất đều có một kiểm thử tập trung hoặc một phản hồi được ghi chép làm cơ sở.

Summary

agentplatform (shipped in google-cloud-aiplatform) exposes
client.rag.ask_contexts(). Every call returns 400 INVALID_ARGUMENT with no
field named.

This does not appear to be a client defect. I checked the request the SDK builds
against the discovery schema and it matches. The same request sent by hand fails
identically, and the v1 form of the method answers:

501 UNIMPLEMENTED
"Operation is not implemented, or supported, or enabled."

So the question is whether AskContexts requires allowlisting, and if so whether
the SDK should surface the method (or raise something more useful) until it is
generally served.

Environment

• google-cloud-aiplatform 1.165.1 (also reproduced on 1.163.0)
• google-genai 2.16.0
• Python 3.13
• Regions tested: us-west1, us-central1, us-east4, europe-west4

What I measured

Live RAG corpus in us-west1, 4 documents, state ACTIVE. Same credentials and
same client for every row.

probe	result
retrieve_contexts, same corpus, creds, region	OK, 4 chunks
ask_contexts() via the SDK	400 INVALID_ARGUMENT, no field named
raw REST, .../locations/L:askContexts	400, identical
body {"query": {...}} only (minimal valid per schema)	400
body {}	400
body naming a corpus that does not exist	400, identical to the valid one
body with an unknown field	400 Unknown name "zzz": Cannot find field
us-central1, us-east4, europe-west4	400 in all three
same method on v1	501 UNIMPLEMENTED
control: retrieveContexts with {}	400 "request must set datasource."

Two rows are the interesting ones:

1. The endpoint parses the body, because it names an unknown field precisely.
2. It never reaches corpus validation, because a nonexistent corpus and a
   valid corpus produce byte-identical errors.

And the control shows the RAG service returns specific, actionable errors in this
region when it is actually serving a method.

The request the SDK builds is schema-correct

GoogleCloudAiplatformV1beta1AskContextsRequest declares exactly {query, tools}.
That is what the SDK sends, and the corpus does reach the server inside tools:

from agentplatform._genai import rag, types as T
from google.genai import types as G

CORPUS = "projects/PROJECT/locations/us-west1/ragCorpora/CORPUS_ID"
tools = [G.Tool(retrieval=G.Retrieval(vertex_rag_store=G.VertexRagStore(
    rag_resources=[G.VertexRagStoreRagResource(rag_corpus=CORPUS)])))]

params = T._AskContextsRequestParameters(
    query=T.RagQuery(text="example question"),
    config=T.AskContextsConfig(tools=tools))

d = rag._AskContextsRequestParameters_to_vertex(params)
d.pop("config", None)          # the call site does this
print(sorted(d))               # ['query', 'tools']
print(CORPUS in str(d))        # True

Reproduction

import agentplatform
from agentplatform._genai import types as T
from google.genai import types as G

c = agentplatform.Client(project="PROJECT", location="us-west1")
CORPUS = "projects/PROJECT/locations/us-west1/ragCorpora/CORPUS_ID"

# works
c.rag.retrieve_contexts(
    vertex_rag_store=G.VertexRagStore(
        rag_resources=[G.VertexRagStoreRagResource(rag_corpus=CORPUS)]),
    query=T.RagQuery(text="example question"))

# 400 INVALID_ARGUMENT, no field named
c.rag.ask_contexts(
    query=T.RagQuery(text="example question"),
    config=T.AskContextsConfig(tools=[G.Tool(retrieval=G.Retrieval(
        vertex_rag_store=G.VertexRagStore(
            rag_resources=[G.VertexRagStoreRagResource(rag_corpus=CORPUS)])))]))

And the v1 form, which is the clearest signal:

POST https://us-west1-aiplatform.googleapis.com/v1/projects/PROJECT/locations/us-west1:askContexts
{"query": {"text": "example question"}}

501 {"error": {"code": 501,
     "message": "Operation is not implemented, or supported, or enabled.",
     "status": "UNIMPLEMENTED"}}

Questions

1. Is AskContexts allowlisted or otherwise gated? If so, what is the path to
   enablement?
2. If it is not generally available, could the v1beta1 endpoint return the same
   UNIMPLEMENTED that v1 returns, instead of a generic INVALID_ARGUMENT that
   sends callers looking for a bad field?

Note on a non-issue

While investigating I noticed the built path is
projects/P/locations/L/:askContexts, with a separator before the colon. That is
not the cause: retrieve_contexts builds projects/P/locations/L/:retrieveContexts
in exactly the same way and works. Mentioning it only so nobody else spends time
on it.