Implement native JSON mode with schema for Gemini

[ipriyaaanshu]

Implement native JSON mode with schema for Gemini

Closes #472

This pull request enhances the Google Gemini integration within Mirascope to support native, schema-constrained JSON outputs, aligning with the latest capabilities of the Gemini API.

Key Changes & Motivations:

1. Native Gemini JSON Mode (response_schema):

   • When using @llm.call(provider="google", ...) (or the older @google_call) with json_mode=True and a Pydantic response_model, Mirascope now directly passes the Pydantic model to the generation_config.response_schema parameter of the Google Gemini API.
   • This leverages Gemini's built-in mechanism for enforcing that the LLM's output strictly adheres to the provided schema, offering a more direct and potentially more reliable method for structured JSON generation compared to relying solely on prompting or tool calling as a fallback for this specific mode.
   • If json_mode=True is used without a response_model (and no tools are specified), a generic prompt instructing the model to output JSON is still appended to the messages, maintaining existing behavior for that scenario.

2. Updated Warning for Strict Outputs:

   • The warning message in BaseTool (triggered when model_config={"strict": True} is used on a response_model) has been updated to include "google" in its list of providers that support such strict structured outputs when json_mode=True. This reflects the new capability.

3. Unit Tests:

   • New unit tests have been added to tests/core/google/_utils/test_setup_call.py to specifically verify the new logic:
     • Ensures response_schema is correctly set with the Pydantic model.
     • Ensures response_mime_type is set to "application/json".
     • Confirms that the generic JSON prompting mechanism (via _utils.json_mode_content) is not invoked when response_schema is active.
   • The new tests align with the existing testing structure and conventions in the file.

4. Minor Docstring Fix:

   • Corrected a minor typo from json_modem to json_mode in the docstring of google_call in mirascope/core/google/_call.py.

How to Test:

Reviewers can test this by:

1. Checking out this branch.
2. Ensuring Mirascope is installed in editable mode with the google extra (uv pip install -e ".[google]").
3. Setting the GOOGLE_API_KEY environment variable.
4. Running a test code that uses @llm.call(provider="google", response_model=MyModel, json_mode=True).
5. Observing that the Gemini API call successfully returns structured data conforming to MyModel.
6. Running pytest to ensure the new and existing unit tests pass, particularly tests/core/google/_utils/test_setup_call.py.

This change aims to provide a more robust and idiomatic integration with Gemini's evolving features for structured data extraction.

[ipriyaaanshu]

@willbakst can you please confirm if my recent commits look good?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 100.00%. Comparing base (9be4325) to head (5d90862).

Additional details and impacted files

@@            Coverage Diff             @@
##              main      #983    +/-   ##
==========================================
  Coverage   100.00%   100.00%            
==========================================
  Files          511       511            
  Lines        21072     21172   +100     
==========================================
+ Hits         21072     21172   +100     

Flag	Coverage Δ
tests	100.00% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[willbakst]

the null check on response_model should guarantee the presence of model_config, so you can do something like this:

if json_mode or (response_model and response_model.model_config.get("strict", False)): ...

[willbakst]

this was previously inside of the if not tools: ... check because it's possible to set both response_model and tools. We need to make sure that if both are set we don't force the response into application/json since the model may want to respond with a tool call.

[willbakst]

why move this inside of this check for not tools? Previously we've always included this json mode content message.