[Evaluation] Align shared evaluator utils/validators/base with azureml-assets builtin

[m7md7sien]

Description

Backports the changes where the azureml-assets builtin evaluators are ahead of / diverged from the shared azure-ai-evaluation code, so the builtin evaluators can eventually import these symbols directly from the SDK instead of inlining private copies. Grouped as buckets C (reconcile diverged shared code), D (PII-safe logging), and E (additive helpers) from the parity analysis.

Bucket C — reconcile diverged shared code

• _normalize_function_call_types (_evaluators/_common/_base_prompty_eval.py): also normalize openapi_call / openapi_call_output → tool_call / tool_result (kept the existing function_call key-rename, so this is a strict superset). This also unblocks importing the widely-used _preprocess_messages.
• _get_agent_response (_common/utils.py): render tool results via the new _stringify_tool_result, so list/dict grounding-tool outputs (Azure AI Search, SharePoint, Fabric) are readable JSON instead of a Python repr. Plain string results are unchanged → backward compatible.
• _get_conversation_history / reformat_conversation_history: new opt-in include_tool_calls path that renders tool calls/results inline. Default behavior is unchanged.

Bucket D — PII-safe fallback logging

• reformat_agent_response, reformat_conversation_history, and reformat_tool_definitions now emit richer, structural-only fallback summaries via the new _log_safe_summary (never raw payloads).

Bucket E — additive shared helpers

• New in _common/utils.py: _stringify_tool_result, _log_safe_summary, _coerce_bool, _coerce_number, _collect_failed_tool_calls, _get_tool_calls_results, _reformat_tool_calls_results.
• EvaluationLevel already exists in _common/constants.py — not re-added.
• Instead of vendoring _create_extended_error_target, added the missing ErrorTarget members: QUALITY_GRADER_EVALUATOR, CUSTOMER_SATISFACTION_EVALUATOR, DEFLECTION_RATE_EVALUATOR, REGEX_MATCH_EVALUATOR (the other extended names already existed).

Testing

• Added unit tests: TestToolResultAndLoggingHelpers in test_utils.py and an OpenAPI preprocessing test in test_multi_turn_utilities.py. Pre-existing OpenAPI-normalization tests now pass.
• test_multi_turn_utilities.py, test_common_validators.py, test_utils.py: 201 passed.
• Affected evaluator suites (agent / groundedness / task / tool): 92 passed.
• black (24.4.0, repo config) clean.

Notes

• All shared-function changes are backward-compatible supersets; no existing evaluator behavior changes except the two intentional ones above (OpenAPI normalization + mid-execution guard).

[Copilot]

Pull request overview

This PR backports shared evaluator utilities in azure-ai-evaluation so they align with the divergent azureml-assets builtin evaluators, enabling the builtin evaluators to eventually import these symbols directly from the SDK. It reconciles diverged shared code (OpenAPI normalization/expansion, JSON-rendered tool results), makes fallback/debug logging PII-safe, and adds new additive helpers. All shared-function changes are described as backward-compatible supersets.

Changes:

• Bucket C: _normalize_function_call_types now also maps openapi_call/openapi_call_output → tool_call/tool_result; _extract_needed_tool_definitions expands OpenAPI tools into nested functions when present; _get_agent_response/conversation-history helpers render tool results as JSON via the new _stringify_tool_result and add an opt-in include_tool_calls path.
• Bucket D: fallback/debug logging in reformat helpers and Groundedness now emits structural-only summaries via _log_safe_summary, never raw payloads.
• Bucket E: new shared helpers (_stringify_tool_result, _log_safe_summary, _coerce_bool, _coerce_number, _collect_failed_tool_calls, _get_tool_calls_results, _reformat_tool_calls_results) plus 4 new ErrorTarget members.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
_common/utils.py	Adds the new shared helpers, include_tool_calls conversation path, JSON tool-result rendering, and PII-safe fallback logging.
_evaluators/_common/_base_prompty_eval.py	Extends _normalize_function_call_types for OpenAPI types and rewrites OpenAPI tool-definition expansion; removes now-unused chain import.
_evaluators/_tool_call_success/_tool_call_success.py	Uses _stringify_tool_result and switches fallback logs to _log_safe_summary.
_evaluators/_groundedness/_groundedness.py	Replaces payload-logging debug line with a count-only summary.
_exceptions.py	Adds four new ErrorTarget enum members for parity.
CHANGELOG.md	Documents the OpenAPI fix and the "Other Changes" (normalization, JSON tool results, PII-safe logging).
tests/unittests/test_utils.py	Adds TestToolResultAndLoggingHelpers covering the new helpers and ErrorTarget members.
tests/unittests/test_tool_call_success_logging.py	New tests asserting fallback paths don't leak raw payloads.
tests/unittests/test_tool_call_accuracy_evaluator.py	Adds OpenAPI tool-definition expansion/validation tests.
tests/unittests/test_multi_turn_utilities.py	Adds an OpenAPI normalization preprocessing test.

Notes on findings: The description's Bucket C MessagesOrQueryResponseInputValidator mid-execution guard (and the associated test_last_message_tool_role_accepted update) is not present in the diff or the current code, and the two _get_tool_calls_results/_reformat_tool_calls_results copies now diverge in argument formatting.