Vertex structured output silently returns empty dict-typed fields (additionalProperties)

[CarSanoja]

Summary

With a Pydantic response schema containing a plain dict field:

class AuditResult(BaseModel):
    mappings: dict[str, list[str]]

used as response_schema on Vertex (gemini-3.5-flash-lite, location global), the model is explicitly instructed to return e.g.
{"mappings": {"alg-factor-1": ["A-SSE.2"]}} — and the API response contains "mappings": {} every time. No error, no warning: dict[str, ...] compiles to JSON-Schema additionalProperties, which Vertex structured output does not
honor, so the field arrives empty.

Why this is worse than a rejection

The call succeeds and validates. In our grading pipeline the curriculum-audit agent silently reported zero curriculum coverage for weeks of development — the feature was fully dead in production while every test on the SDK surface passed. A hard error on unsupported schema constructs (or dropping the field from the generated schema with a warning) would have surfaced this instantly.

Reproduction

1. Define the model above; call generate_content on Vertex with response_mime_type="application/json", response_schema=AuditResult.
2. Prompt: "Return mappings = {'alg-factor-1': ['A-SSE.2']}".
3. Observe mappings == {} in the parsed response.

Workaround we shipped: a list-shaped wire model ([{"criterion_id": ..., "competency_codes": [...]}]) converted back to the
dict-shaped domain model after parsing.

Environment

google-genai 2.18.1 (Vertex mode), gemini-3.5-flash-lite @ global, Python 3.13. Found while building a K-12 exam-grading fleet for the All Things Agentic hackathon.

[Venkaiahbabuneelam]

Hi @CarSanoja, Thanks for reaching out us!

Thank you for bringing this to our attention. We truly appreciate you flagging this issue, we will file a bug internally.

Thanks

[CarSanoja]

Thanks for filing it internally.

One detail worth carrying into the internal bug, since it is what made this expensive for us: the call succeeds. There is no error, no warning, and the response validates against the schema, the field simply arrives empty. Our curriculum-audit agent reported zero coverage while every test on the SDK surface passed, because "empty mappings" is a structurally valid AuditResult. A hard error (or dropping the unsupported construct from the generated schema with a warning) would have surfaced it on the first call.

For anyone who lands here before the fix, the workaround we shipped is to avoid dict[...] in the response schema entirely and use a list-shaped wire model, converting back after parsing:

class CompetencyMapping(BaseModel):
    criterion_id: str
    competency_codes: list[str]

class AuditResponse(BaseModel):           # what Vertex actually returns
    mappings: list[CompetencyMapping]

    def to_domain(self) -> AuditResult:   # dict-shaped model kept internal
        return AuditResult(mappings={m.criterion_id: m.competency_codes for m in self.mappings})

That restored full coverage with no prompt changes beyond describing the list shape. Happy to test a fix against our real workload whenever there is a branch, we hit this path on every batch.