Remove evaluator defs

[akshaylive]

Sorry for the massive PR.

The main change is that evaluators are currently not serializable. Additionally, evaluators cannot be model_dump'd and reconstructed. Addressing these issues by ensuring that all evaluator models are naturally pydantic.

This addresses #1246
This also fixes #1220

What changed?

1. Renamed BaseEvaluator to GenericBaseEvaluator.
2. BaseLegacyEvaluator now extends GenericBaseEvaluator. All legacy evaluators extends from BaseLegacyEvaluator.
3. BaseEvaluator now extends GenericBaseEvaluator. All coded agent evaluators extends from BaseEvaluator. The name BaseEvaluator is NOT named BaseCodedEvaluator in order to keep in sync with the custom evaluator contract.
4. name and description are now pydantic fields.
5. config field removed from the object as it is redundant.
6. *_type fields in GenericBaseEvaluator are marked as exclude=True for serialization safety.
7. Evaluator factory is now vastly simplified considering they can be instantiated using TypeAdapter(Evaluator).validate_python(...).
8. Standardizing evaluator naming scheme.

Development Package

• Use uipath pack --nolock to get the latest dev build from this PR (requires version range).
• Add this package as a dependency in your pyproject.toml:

[project]
dependencies = [
  # Exact version:
  "uipath==2.7.1.dev1012554529",

  # Any version from PR
  "uipath>=2.7.1.dev1012550000,<2.7.1.dev1012560000"
]

[[tool.uv.index]]
name = "testpypi"
url = "https://test.pypi.org/simple/"
publish-url = "https://test.pypi.org/legacy/"
explicit = true

[tool.uv.sources]
uipath = { index = "testpypi" }

[tool.uv]
override-dependencies = [
    "uipath>=2.7.1.dev1012550000,<2.7.1.dev1012560000",
]

[mjnovice]

Meta question: What is the pressing need for evaluators to be serializable ?

[akshaylive]

Meta question: What is the pressing need for evaluators to be serializable ?

During conversion to coded agents, we need to change the implementation from evals/ to evaluators/. The coded agents frontend does not understand legacy evaluators so we need to convert it. This change would simplify the conversion of evaluators.

[mjnovice]

dumb question: where did this go ? how are we doing this conversion now ?

[akshaylive]

We don't need this. TypeAdapter(CodedEvaluator).validate_python(...) uses the discriminator to resolve it correctly.

[mjnovice]

you mean, dynamic class conversion ? Am curious, why couldn't we do that in the previous form - because of the naming inconsistency ?

[akshaylive]

In the previous implementation, the coded evaluators were discriminating on the configs and not the evaluators. Ideally the original implementation should've been this implementation. 🙃

[Chibionos]

I believe this was done to be too explicit in the previous implementation thanks for fixing it @akshaylive. @mjnovice we will need a few more rounds of refactoring to get it to GA level in the progress reporting side too :D

[akshaylive]

Manually tested this part. Screenshot: