feat: optionally penalize ambiguous answer relevancy verdicts

[Yashaswini1233]

Summary

• Adds a penalize_ambiguous_statements option to AnswerRelevancyMetric.
• When enabled, idk verdicts are treated as irrelevant when calculating the answer relevancy score.
• Preserves the existing behavior by default with penalize_ambiguous_statements=False.
• Includes ambiguous verdicts in the generated reason when penalization is enabled.

Motivation

Currently, AnswerRelevancyMetric counts every verdict other than no as relevant, which means an idk verdict contributes positively to the score.

This change provides an opt-in stricter scoring mode for users who want ambiguous statements to be penalized while remaining backward compatible with the existing behavior.

Addresses #1137.

Testing

Added focused tests covering:

• Existing/default idk scoring behavior
• Penalized idk scoring
• Ambiguous-only verdicts
• Strict mode behavior
• Existing empty-output regression behavior

Validation performed locally:

• 8 targeted tests passed
• black --check passed
• Python compilation passed
• git diff --check passed

[vercel]

@Yashaswini1233 is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.

[Yashaswini1233]

Note on CI: the lint check is currently failing on deepeval/test_case/llm_test_case.py, which is unchanged by this PR and identical to the version on upstream/main. I also reproduced the check locally with Black 25.12.0. The changes introduced in this PR are Black-formatted and the 8 targeted Answer Relevancy tests pass locally.

[rajeshl8]

Duplicate method definition.
This PR adds a new _calculate_score method at line 313 that only contains the empty-verdicts guard (return 1), but the original full _calculate_score is still defined below it at ~line 325. In Python, when a method is defined twice in the same class body, the second definition silently overrides the first — so this new one is unreachable dead code and will never execute.

[Yashaswini1233]

Thanks for catching this. I removed the duplicate _calculate_score() definition and retained the empty-verdict guard in the single complete method. All 8 targeted tests pass locally.

[Yashaswini1233]

The duplicate _calculate_score() definition has been removed, leaving one complete scoring method with the empty-verdict guard intact. All 8 targeted tests pass locally. The remaining Black CI failure is on .scripts/release.py, which is unrelated to this PR. The Vercel check requires repository authorization.