Your personal operating system for career transition.

Implement an Evaluation Configuration Loader

Eval pipelines need configuration: which metrics to run, what thresholds to enforce, which model to use as the judge, and which categories to include. Hard-coding these values makes the pipeline difficult to adapt. A configuration loader centralises them in one place.

What to build

Implement EvalConfig (a Pydantic-style dataclass):

Fields: metrics: list[str] (default ["faithfulness", "relevance"]), pass_threshold: float (default 0.7), regression_threshold: float (default 0.10), judge_model: str (default "gpt-4o-mini"), categories: list[str] (default empty = all), max_cases: int | None (default None = unlimited), sample_rate: float (default 1.0, clamped to 0.0-1.0).

Implement load_eval_config(source: dict | None = None) -> EvalConfig:

• Accepts a dict (from JSON file, env vars, CLI args).
• Unknown keys are silently ignored.
• Invalid float ranges raise ValueError with a descriptive message.
• Missing keys use defaults.

Implement EvalConfig.to_dict() -> dict returning all fields.

Constraints

• Standard library only. No Pydantic. Use dataclasses and manual validation.