[NA] [Optimizer] Improved structured output support

[jverre]

Details

Not all models support structured outputs but it's a key strategy we use to ensure the optimizations go smoothly. In this PR, we introduce 2 new strategies to support all models for optimization:

1. Native support
2. Tool calling support
3. Prompting support

Change checklist

• User facing
• Documentation update

Issues

• Resolves #
• OPIK-000

Testing

Added tests

Documentation

N/A

[Copilot]

Pull Request Overview

This PR introduces improved structured output support for the optimizer by implementing three strategies to handle models with varying capabilities: native response format support, tool calling support, and prompt injection. This ensures the optimizer can work with all models, even those without native JSON schema support.

Key Changes

• Added three structured output strategies (native, tool_call, prompt_injection) to accommodate different model capabilities
• Refactored parameter handling to extract and validate response_format_type from model parameters
• Enhanced error messages to guide users toward alternative strategies when parsing fails

Reviewed Changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
sdks/opik_optimizer/src/opik_optimizer/_llm_calls.py	Core implementation of three structured output strategies with helper functions for tool definitions, prompt injection, and response extraction
sdks/opik_optimizer/tests/unit/test_llm_calls.py	Unit tests for all three strategies using mocked litellm responses
sdks/opik_optimizer/tests/e2e/test_llm_calls_integration.py	End-to-end integration tests verifying each strategy with real model calls and trace validation
sdks/opik_optimizer/tests/conftest.py	Test fixtures recreated from Python SDK for fake backend and trace capture
sdks/opik_optimizer/pyproject.toml	Added mypy ignore configuration for testlib module

[dsblank]

LGTM!