feat: Declarative eval

[CakeCrusher]

Description

Adds declarative_eval and transform_field_mappings_for_explanation to arize-phoenix-evals

Offline evaluator function designed after llm_classify for declarative evaluations enabling the user to describe evaluations through a Pydantic BaseModel subclass.

This enables free-form extensive evaluations according to the user's needs. Useful when llm_classify or other specialized evaluator functions are insufficient or the user may need multiple evaluations but wants to minimize overhead.

Note: This PR is 45% me learning the repo and dev flow, 45% I find it useful, 10% speculate others may find it useful.

Expample use

(copied from docstring)

# Define a schema with nested models
class Conciseness(BaseModel):
    is_concise: bool = Field(..., description="Whether the output is concise")

class Formatting(BaseModel):
    language: Literal["High", "Average", "Low"] = Field(
        ..., description="The complexity of the formatting used in the output"
    )

class Schema(BaseModel):
    conciseness: Conciseness = Field(..., description="A custom evaluation of the output")
    formatting: Formatting = Field(..., description="A custom evaluation of the output")

# Prepare sample data
data = pd.DataFrame({
    "attributes.llm.input_messages": [
        [{"role": "user", "content": "What is 2+2?"}],
        [{"role": "user", "content": "Who was the first president?"}],
    ],
    "attributes.llm.output_messages": [
        [{"role": "assistant", "content": "Whenever you add those two numbers, you get 4"}],
        [{"role": "assistant", "content": "George Washington"}],
    ],
})

# Define field mappings
field_mappings = {
    "conciseness.label": "conciseness.is_concise",
    "formatting.label": "formatting.language",
}

# Run the evaluation
result = await declarative_eval(
    data=data,
    model=openai_client,
    schema=Schema,
    field_mappings=field_mappings,
)

# Result will be a DataFrame with columns:
# - conciseness.label (containing boolean values)
# - formatting.label (containing "High", "Average", or "Low")
# - execution_seconds (execution time)
# - exceptions (any errors encountered)

Test coverage

• Success with base fields
• Success with provide_explanation
• Failure with incorrect field_mappings

TODO

• Enable synchronous evaluation of rows (could use generator but will require updating OpenAIModel)
• Integrate it for use in online "Tasks"
• Abstraction to upsert eval into project (potentially using Client(...).log_evaluations_sync(...))

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[CakeCrusher]

I see im getting errors with importing the new functions, not sure How concerned I should be:

=================================== ERRORS ====================================
_________ ERROR collecting tests/unit/evals/test_declarative_eval.py __________
ImportError while importing test module 'D:\a\phoenix\phoenix\tests\unit\evals\test_declarative_eval.py'.
Hint: make sure your test modules/packages have valid Python names.
Traceback:
unit\evals\test_declarative_eval.py:12: in <module>
    from phoenix.evals import declarative_eval, transform_field_mappings_for_explanation
E   ImportError: cannot import name 'declarative_eval' from 'phoenix.evals' (D:\a\phoenix\phoenix\.tox\unit_tests\Lib\site-packages\phoenix\evals\__init__.py)

As discussed with @cephalization in https://arize-ai.slack.com/archives/C018252LE1E/p1745724863399289?thread_ts=1745694183.495699&cid=C018252LE1E the import was working as long as arize-phoenix-evals was downloaded from local.
The tests are working with tox run -e unit_tests_local_evals -- PATH_TO/test_declarative_eval.py or tox run -e unit_tests_local_evals