Azure/azure-sdk-for-python azure-ai-evaluation

1.18.0 (2026-07-06)

Evaluators now raise EvaluationException (instead of bare ValueError/TypeError) for input and
configuration validation failures, and consistently set blame=ErrorBlame.USER_ERROR with an
appropriate category and target. Affected evaluators include ContentSafetyEvaluator,
QAEvaluator, RougeScoreEvaluator, DocumentRetrievalEvaluator, the task navigation efficiency
evaluator, and the shared evaluator base (conversation/tool-call input validation).
Fixed RedTeam.scan() storing decoded plaintext instead of the actual
encoded payload for converter-based attack strategies (Base64, Flip,
Morse, ROT13, etc.) in evaluation_results.json / results.json. The
persisted conversation[].content for user turns now reflects what the
target actually received (converted_value). The pre-converter
adversarial objective is additionally retained as an original_value
field on user messages in the intermediate per-strategy .jsonl
artifacts (it is not surfaced in the customer-facing results.json,
which continues to expose only role/content/name). Baseline
(non-encoded) strategies are unaffected.
Resolves #47228.