feat(deepeval): support custom metric instances in evaluator

[ClaireXi99]

Fixes #3838

What

DeepEvalEvaluator now accepts an already initialized DeepEval metric, so any subclass
of deepeval.metrics.BaseMetric can be used for evaluation:

class ResponseLengthMetric(BaseMetric):
    _required_params = [SingleTurnParams.INPUT, SingleTurnParams.ACTUAL_OUTPUT]

    def measure(self, test_case: LLMTestCase) -> float:
        words = len((test_case.actual_output or "").split())
        self.score = min(1.0, words / self.max_words)
        self.reason = f"The response contains {words} words"
        return self.score


evaluator = DeepEvalEvaluator(metric=ResponseLengthMetric(max_words=20))

The inputs the component exposes are derived from the test case parameters the metric
declares, and results come back in the same name/score/explanation shape as the
built-ins.

Backwards compatibility

Purely additive. metric widens from str | DeepEvalMetric to
str | DeepEvalMetric | BaseMetric; the built-in branch, its init validation, its input
validation and its serde are untouched. Existing tests still pass unchanged.

Two design points worth your call

1. Parameter discovery reads _required_params. DeepEval exposes no public way to
ask a metric which test case params it needs, so this reads the private attribute. Two
details:

• Any collection of SingleTurnParams is accepted, not just list — DeepEval does not
  promise a concrete container type, and silently discarding a tuple/set declaration
  would map the component's inputs to the RAG defaults without telling the user.
• BaseMetric only annotates _required_params without assigning it, so on a metric
  that never declares it getattr returns the typing alias
  typing.List[SingleTurnParams] rather than None. That case — and anything else that
  cannot be read as params — falls back to the built-in RAG inputs.

If you would rather not depend on a private attribute, the alternative is an explicit
constructor argument for the expected inputs. Happy to switch.

2. A metric instance cannot be serialized. to_dict raises for that case, since an
instance carries runtime state that cannot be reliably reconstructed. The class docstring
points at the built-in metrics when a pipeline has to survive a to_dict/from_dict
round trip. Tell me if you would prefer a warning plus a degraded dict instead of a hard
error.

Also note this supports a single metric instance, matching the current DeepEval
evaluator shape. RagasEvaluator takes a list of metrics — say the word and I will align
this one to that API.

Tests

integrations/deepeval/tests/test_evaluator.py gains coverage for input derivation
(declared as list, declared as tuple, undeclared, unreadable, unsupported params), the
measure() call path, output shape, LLMTestCase conversion, and passing a built-in
metric as an instance. Locally: 32 passed, 5 skipped (the skips need OPENAI_API_KEY).
ruff check and ruff format --check are clean.

Two notes so they don't surprise you in review:

• I did not touch CHANGELOG.md — it looks like HaystackBot generates it via
  cliff.toml during the release workflow. Let me know if contributors are expected to
  add an entry.
• mypy src reports 8 errors in this integration, but they are identical before and
  after this change (only one line number shifts, because the docstring grew). They come
  from deepeval.metrics attribute resolution against the installed deepeval version, so
  I left them alone rather than widening the diff.

[github-actions]

Heads-up for maintainers

This PR is from a fork and touches integrations whose integration tests require API keys.
Those tests are skipped in CI because fork PRs don't have access to repo secrets for security reasons.

Affected integrations:

• deepeval

Please run the integration tests locally (hatch run test:integration inside each folder) before approving.

[CLAassistant]

All committers have signed the CLA.

[github-actions]

Coverage report (deepeval)

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
integrations/deepeval/src/haystack_integrations/components/evaluators/deepeval
evaluator.py
metrics.py
Project Total

This report was generated by python-coverage-comment-action