Structured output does not work with Grounding

[samshipengs]

I'm trying to use structured output (python pydantic model schema) with grounding = true, and I got the following error:

ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Unable to submit request because controlled generation is not supported with google_search tool. Learn more: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/gemini', 'status': 'INVALID_ARGUMENT'}}

Based on the error, my guess is ground=true cannot be used with structured output?

my simple code in python:
google-genai==1.10.0

google_search_tool = (
                [Tool(google_search=GoogleSearch())] if grounding else None
            )

            response = self.gemini_client.models.generate_content(
                model=self.model_name,
                contents=user_prompt,
                config=GenerateContentConfig(
                    tools=google_search_tool,
                    system_instruction=prompt,
                    temperature=0.2,
                    # top_k=100,
                    # top_p=0.2,
                    response_mime_type="application/json",
                    response_schema=schema,
                    seed=42,
                ),
            )

[sasha-gitg]

The exception returned is correct. Can only use either grouding or structured output but not both.

[mrdziuban]

@sasha-gitg if possible, could you share more info on the reasoning behind this limitation? Are there any plans for the Gemini 2.0/2.5 models to support structured output in combination with grounding?

[VladimirArustamian]

It worked perfectly for both gemini 1.5 models. Just for that combination I had both outdated models live in production. But now they are deprecated.

[hacky-code]

Any update here? It's a massive gap with so much potential value, plus a key competitive differentiator for Gemini. Hope the PM is listening :)

[ege90]

issue still persists, any update here ?

[Emilie2029]

The generate_content function does work when you call it without passing the proper structured output config (response_schema and response_mime_type) but asking in the prompt for json-like output.
However, the grounding chunks and supports are empty lists. Even though the grounding tool was passed correctly.

Is there a reason why no grounding information is provided the moment you hint towards structured output in the prompt?

[greyraven]

I'm having to do a 2-step process where I do the first api call with grounding and then I pass the results to a Gemini 2.0 flash-lite to output it to structured format. For google, if native structured output is infeasible in the short term, consider integrating a lightweight post-processing layer within Gemini 2.5 Pro to handle structuring tasks, eliminating the need for a separate model

[as10896]

I was able to use both structured output and grounding simultaneously with Gemini 2.0.
However, this doesn't seem to be possible with the latest Gemini 2.5, which is a bit disappointing.
Is there any timeline for when this might be addressed?

[jtollefson]

Unfortunately you can't do both, I've worked around the problem by grounding with search and then sending it back to obtain the structured output. It's double the calls, but, solves the problem.