Ben Thomas 35adfdb318 Python: Foundry Evals integration for Python (#4750) ...

* Foundry Evals integration for Python

Merged and refactored eval module per Eduard's PR review:

- Merge _eval.py + _local_eval.py into single _evaluation.py
- Convert EvalItem from dataclass to regular class
- Rename to_dict() to to_eval_data()
- Convert _AgentEvalData to TypedDict
- Simplify check system: unified async pattern with isawaitable
- Parallelize checks and evaluators with asyncio.gather
- Add all/any mode to tool_called_check
- Fix bool(passed) truthy bug in _coerce_result
- Remove deprecated function_evaluator/async_function_evaluator aliases
- Remove _MinimalAgent, tighten evaluate_agent signature
- Set self.name in __init__ (LocalEvaluator, FoundryEvals)
- Limit FoundryEvals to AsyncOpenAI only
- Type project_client as AIProjectClient
- Remove NotImplementedError continuous eval code
- Add evaluation samples in 02-agents/ and 03-workflows/
- Update all imports and tests (167 passing)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: resolve mypy redundant-cast errors while keeping pyright happy

Use cast(list[Any], x) with type: ignore[redundant-cast] comments to
satisfy both mypy (which considers casting Any redundant) and pyright
strict mode (which needs explicit casts to narrow Unknown types).

Also fix evaluator decorator check_name type annotation to be
explicitly str, resolving mypy str|Any|None mismatch.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: CI failures — pyupgrade, evaluator overloads, sample API, reset attr

- Apply pyupgrade: Sequence from collections.abc, remove forward-ref quotes
- Add @overload signatures to evaluator() for proper @evaluator usage
- Fix evaluate_workflow sample to use WorkflowBuilder(start_executor=) API
- Fix _workflow.py executor.reset() to use getattr pattern for pyright
- Remove unused EvalResults forward-ref string in default_factory lambda

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: skip gRPC-dependent observability test

The test_configure_otel_providers_with_env_file_and_vs_code_port test
triggers gRPC OTLP exporter creation, but the grpc dependency is
optional and not installed by default. Add skipif decorator matching
the pattern used by all other gRPC exporter tests in the same file.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: add nosec B101 for bandit assert check

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* style: align eval samples with repo conventions

- Move module docstrings before imports (after copyright header)
- Add -> None return type to all main() and helper functions
- Fix line-too-long in multiturn sample conversation data
- Add Workflow import for typed return in all_patterns_sample

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review feedback: async fixes, sample bugs, deprecation warnings

- Simplify _ensure_async_result to direct await (async-only clients)
- Replace get_event_loop() with get_running_loop()
- Narrow _fetch_output_items exception handling to specific types
- Add warning log when _filter_tool_evaluators falls back to defaults
- Add DeprecationWarning to options alias in Agent.__init__
- Add DeprecationWarning to evaluate_response()
- Rename raw key to _raw_arguments in convert_message fallback
- Fix evaluate_agent_sample.py: replace evals.select() with FoundryEvals()
- Fix evaluate_multiturn_sample.py: use Message/Content/FunctionTool types
- Fix evaluate_workflow_sample.py: replace evals.select() with FoundryEvals()
- Update test mocks to use AsyncMock for awaited API calls

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Add test coverage for review feedback items

- Add num_repetitions=2 positive test verifying 2×items and 4 agent calls
- Add _poll_eval_run tests: timeout, failed, and canceled paths
- Add evaluate_traces tests: validation error, response_ids path, trace_ids path
- Add evaluate_foundry_target happy-path test with target/query verification

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix ruff ISC004 lint error and apply formatter

- Wrap implicit string concatenation in parens in evaluate_multiturn_sample.py
- Apply ruff formatter to 6 other files with minor formatting drift

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove core type changes (extracted to fix/workflow-stale-session branch)

Reverts changes to _agents.py, _agent_executor.py, and _workflow.py
back to upstream/main. These fixes are now in a separate PR.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review round 2: bugs, tests, and architecture

Code fixes:
- Fix _normalize_queries inverted condition (single query now replicates
  to match expected_count)
- Fix substring match bug: 'end' in 'backend' matched; use exact set
  lookup for executor ID filtering
- Fix used_available_tools sample: tool_definitions→tools param, use
  FunctionTool attribute access instead of dict .get()
- Add None-check in _resolve_openai_client for misconfigured project
- Add Returns section to evaluate_workflow docstring
- Cache inspect.signature in @evaluator wrapper (avoid per-item reflection)

Architecture:
- Extract _evaluate_via_responses as module-level helper; evaluate_traces
  now calls it directly instead of creating a FoundryEvals instance
- Move Foundry-specific typed-content conversion out of core to_eval_data;
  core now returns plain role/content dicts, FoundryEvals applies
  AgentEvalConverter in _evaluate_via_dataset

Tests:
- evaluate_response() deprecation warning emission and delegation
- num_repetitions > 1 with expected_output and expected_tool_calls
- Mock output_items.list in test_evaluate_calls_evals_api
- Update to_eval_data assertions for plain-dict format
- Unknown param error now raised at @evaluator decoration time

Skipped (separate PR): executor reset loop, xfail removal, options alias

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix CI: revert test_full_conversation, fix pyright errors

- Revert test_full_conversation.py to upstream/main (the session
  preservation test was incorrectly changed to assert clearing)
- Fix pyright reportUnnecessaryComparison on get_openai_client() None
  check by adding ignore comment
- Fix pyright reportPrivateUsage: add public EvalItem.split_messages()
  method and use it in FoundryEvals._evaluate_via_dataset instead of
  accessing private _split_conversation

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review round 3: reliability, test gaps, cleanup

- Add try/except guard for non-numeric score in _coerce_result
- Add poll_interval minimum bound (0.1s) to prevent tight loops
- Add runtime async client check in _resolve_openai_client
- Remove _ensure_async_result wrapper (10 call sites → direct await)
- Better error message when queries provided without agent
- Import-time asserts for evaluator set consistency
- Remove 28 redundant @pytest.mark.asyncio decorators
- Add doc note about _raw_arguments sensitive data
- Tests: tool_called_check mode=any, _normalize_queries branches,
  _extract_result_counts paths, _extract_per_evaluator, bare check
  via evaluate_agent, output_items assertion, modulo wrapping,
  async client check, queries-without-agent error

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix CI: ruff S101 assert, pyright and mypy arg-type errors

- Replace module-level assert with if/raise for evaluator set
  consistency checks (ruff S101 disallows bare assert)
- Add type: ignore[arg-type] and pyright: ignore[reportArgumentType]
  on OpenAI SDK evals API calls that pass dicts where typed params
  are expected (SDK accepts dicts at runtime)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review round 4: bugs, reliability, test fixes

- Fix all_passed ignoring parent result_counts when sub_results present
- Fix _extract_tool_calls: parse string arguments via json.loads before
  falling back to None (real LLM responses use string arguments)
- Sanitize _raw_arguments to '[unparseable]' to avoid leaking sensitive
  tool-call data to external evaluation services
- Add NOTE comment on to_eval_data message serialization dropping
  non-text content (tool calls, results)
- Eliminate double conversation split in _evaluate_via_dataset: build
  JSONL dicts directly from split_messages + AgentEvalConverter
- Raise poll_interval floor from 0.1s to 1.0s to prevent rate-limit
  exhaustion
- Fix MagicMock(name=...) bug in test: sets display name not .name attr
- Fix mock_output_item.sample: use MagicMock object instead of dict so
  _fetch_output_items exercises error/usage/input/output extraction

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review round 5: reliability, docs, test coverage

Code fixes:
- Move import-time RuntimeError checks to unit tests (avoids breaking
  imports for all users on developer set-drift mistake)
- _filter_tool_evaluators now raises ValueError when all evaluators
  require tools but no items have tools (was silently substituting)
- Add poll_interval upper bound (60s) to prevent single-iteration sleep
- Log exc_info=True in _fetch_output_items for debugging API changes
- Fix evaluate() docstring: remove claim about Responses API optimization
- Validate target dict has 'type' key in evaluate_foundry_target
- Document to_eval_data() limitation: non-text content is omitted

Tests:
- TestEvaluatorSetConsistency: verify _AGENT/_TOOL subsets of _BUILTIN
- TestEvaluateTracesAgentId: agent_id-only path with lookback_hours
- TestFilterToolEvaluatorsRaises: ValueError on all-tool no-items
- TestEvaluateFoundryTargetValidation: target without 'type' key
- Assert items==[] on failed/canceled poll results
- Mock output_items.list in response_ids test for full flow
- TestAllPassedSubResults: result_counts=None + sub_results delegation
  and parent failures override sub_results
- TestBuildOverallItemEmpty: empty workflow outputs returns None

Skipped r5-07 (_raw_arguments length hint): marginal debugging value,
could leak content size information.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix error message: evaluate_responses() → evaluate_traces(response_ids=...)

The referenced function doesn't exist; the correct API is
evaluate_traces(response_ids=...) from the azure-ai package.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove dead to_eval_data() method, fix docstring claims

- Remove to_eval_data() from EvalItem (dead code after r4-05 JSONL refactor)
- Migrate 15 tests from to_eval_data() to split_messages()
- Update sample to use split_messages() + Message properties
- Remove unimplemented Responses API optimization docstring claim
- Update split_messages() docstring to not reference removed method

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Reduce default eval timeout from 600s to 180s (3 minutes)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove dead _evaluate_via_responses method from FoundryEvals

The method was never called — evaluate() uses _evaluate_via_dataset,
and evaluate_traces() calls _evaluate_via_responses_impl directly.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Revert unrelated formatting changes to get-started samples

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix pyright: remove phantom FoundryMemoryProvider import, apply ruff format

- Remove import of non-existent _foundry_memory_provider module
  (incorrectly kept during rebase conflict resolution)
- Apply ruff formatter to test_local_eval.py and get-started samples

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix eval samples: use FoundryChatClient for Agent()

The upstream provider-leading client refactor (#4818) made client=
a required parameter on Agent(). Update the three getting-started
eval samples to use FoundryChatClient with FOUNDRY_PROJECT_ENDPOINT,
matching the standard pattern from 01-get-started samples.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Simplify self-reflection sample using FoundryEvals

Replace ~80 lines of manual OpenAI evals API code (create_eval,
run_eval, manual polling, raw JSONL params) with FoundryEvals:

- evaluate_groundedness() uses FoundryEvals.evaluate() with EvalItem
- Remove create_openai_client(), create_eval(), run_eval() functions
- Remove openai SDK type imports (DataSourceConfigCustom, etc.)
- run_self_reflection_batch creates FoundryEvals instance once,
  reuses it for all iterations across all prompts

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Update eval samples to FoundryChatClient and FOUNDRY_PROJECT_ENDPOINT

- Migrate all foundry_evals samples from AzureOpenAIResponsesClient to FoundryChatClient
- Update env var from AZURE_AI_PROJECT_ENDPOINT to FOUNDRY_PROJECT_ENDPOINT
- Use AzureCliCredential consistently across all samples
- Fix README.md: correct function names (evaluate_dataset -> FoundryEvals.evaluate, evaluate_responses -> evaluate_traces)
- Update self_reflection .env.example and README.md

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix lint errors in eval samples (E501, ASYNC240, formatting)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove evaluate_all_patterns_sample.py (redundant with focused samples)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix async credential mismatch: use azure.identity.aio for async AIProjectClient

AIProjectClient from azure.ai.projects.aio requires an async credential.
Switch all foundry_evals samples from azure.identity.AzureCliCredential
to azure.identity.aio.AzureCliCredential. Also pass project_client to
FoundryChatClient instead of duplicating endpoint+credential.

Close credential in self_reflection sample to avoid resource leak.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Revert test_observability.py to upstream/main (not our test)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address moonbox3 review: sphinx docstrings, pagination, isinstance check

- Convert all Example:: / Typical usage:: code blocks to .. code-block:: python
  format matching codebase convention (both _evaluation.py and _foundry_evals.py)
- Add async pagination in _fetch_output_items via async for (handles large result sets)
- Replace hasattr(__aenter__) with isinstance(client, AsyncOpenAI) in _resolve_openai_client
- Move AsyncOpenAI import from TYPE_CHECKING to runtime (needed for isinstance)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix test failures and address remaining moonbox3 review comments

- Fix tests: use MagicMock(spec=AsyncOpenAI) for project_client mocks
  (isinstance check now requires proper type, not duck-typing)
- Fix tests: replace mock_page.__iter__ with _AsyncPage helper for async for
- Fix evaluate_response: auto-extract queries from response messages when
  query is not provided (previously always raised ValueError)
- Add debug logging when skipping internal _-prefixed executor IDs

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address Tao's PR review comments on Foundry Evals

- T1: Add comment explaining builtin.* pass-through in _resolve_evaluator
- T2: Add comment referencing OpenAI evals API for testing_criteria dict
- T3: Document Mustache-style {{item.*}} template placeholders
- T4: Document poll loop 60s sleep upper bound rationale
- T5: Narrow run type to RunRetrieveResponse, use typed field access
  instead of vars()/getattr dance in _extract_result_counts and
  _extract_per_evaluator; use run.error and run.report_url directly
- T6: Clarify openai_client docstring re: Azure Foundry endpoint
- T8: Remove misleading empty expected_tool_calls from sample
- Update tests to match real SDK PerTestingCriteriaResult shape

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove unnecessary Any union from run type annotations

RunRetrieveResponse is the correct type — no backward compat needed
for a brand new feature.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Accept FoundryChatClient instead of raw AsyncOpenAI

FoundryEvals now takes client: FoundryChatClient as its primary
parameter instead of openai_client: AsyncOpenAI.  The builtin.*
evaluators require a Foundry endpoint, so the type should reflect that.

- FoundryEvals.__init__: client: FoundryChatClient replaces openai_client
- evaluate_traces / evaluate_foundry_target: same change
- _resolve_openai_client: extracts .client from FoundryChatClient
- project_client fallback retained for standalone functions
- All samples updated to construct FoundryChatClient and pass as client=
- Tests updated (openai_client= → client=)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove implicit 60s upper bound on poll interval

If a developer sets a higher poll_interval, respect it. Only clamp
to remaining time and enforce a 1s minimum for rate-limit protection.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Remove 1s floor on poll interval — let the developer control it

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Update python/samples/05-end-to-end/evaluation/foundry_evals/.env.example

Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>

* Update python/samples/02-agents/evaluation/evaluate_agent.py

Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>

* Address eavanvalkenburg review (round 2) on Python eval PR

- Rename model_deployment -> model across FoundryEvals and all samples
- Make model param optional, resolves from client.model
- Convert EvalResults from dataclass to regular class
- Remove deprecated evaluate_response() function
- Refactor splitters: BUILT_IN_SPLITTERS dict + standalone functions
- Change per_turn_items from classmethod to staticmethod
- Simplify EvalCheck type alias to use Awaitable[CheckResult]
- Remove errored property from EvalResults
- Remove default value from Evaluator protocol eval_name
- Rename assert_passed -> raise_for_status, add EvalNotPassedError
- Type agent param as SupportsAgentRun | None
- Fix Arguments docstring
- Update __init__.py exports
- Update all tests and samples

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Move FoundryEvals to foundry package, split tool eval sample

- Move _foundry_evals.py from azure-ai to foundry package
- Move test_foundry_evals.py to foundry/tests/
- Update lazy re-exports in agent_framework.foundry namespace
- Update .pyi type stubs
- All samples now import from agent_framework.foundry
- Split tool-call evaluation into evaluate_tool_calls_sample.py
- Fix all_passed to check errored count from result_counts
- Fix raise_for_status to include errored item details

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Auto-create FoundryChatClient from env vars when no client provided

FoundryEvals() now works zero-config when FOUNDRY_PROJECT_ENDPOINT and
FOUNDRY_MODEL environment variables are set. Auto-creates a FoundryChatClient
under the hood, matching the established env var pattern.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix pyright errors: remove dead _normalize_queries, suppress EvalAPIError check

- Remove unused _normalize_queries function and its tests
- Add pyright ignore for EvalAPIError None check (defensive guard)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Support multimodal image content in eval pipeline

Add image (data/uri) content handling to AgentEvalConverter.convert_message()
so that Content.from_data() and Content.from_uri() image payloads are
preserved as input_image parts in the Foundry evaluator format.

- Handle Content type='data' and type='uri' → emit input_image parts
- Add 6 unit tests for image content through convert_message/convert_messages
- Add integration test verifying images flow through EvalItem → JSONL path
- Add evaluate_multimodal.py sample demonstrating local image eval

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address remaining review comments

- Fix project_client docstring to say async-only (not sync/async)
- Add builtin evaluator name validation warning in _resolve_evaluator
- Replace getattr with typed attribute access in _poll_eval_run,
  _extract_result_counts, _extract_per_evaluator, _fetch_output_items
- Remove cast import from _foundry_evals (no longer needed)
- Tighten _coerce_result: honour explicit 'passed' when both 'score'
  and 'passed' are present; remove performative cast
- Fix self_reflection sample: add env file existence check
- Fix traces sample: correct Pattern 2 section label
- Update all Foundry eval samples to FoundryChatClient + FOUNDRY_MODEL
  (remove AIProjectClient + AZURE_AI_MODEL_DEPLOYMENT_NAME pattern)
- Add eval_name and OpenAI client docs to FoundryEvals docstring
- Update test mocks to match typed SDK objects (_MockResultCounts)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix ruff lint errors (E501, SIM108, SIM102)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Fix pyright errors: type-narrow dict to dict[str, Any], add ignore comments

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Replace ConversationSplitter type alias with Protocol

ConversationSplitter is now a runtime-checkable Protocol with a named
'conversation' parameter, making the expected signature self-documenting.

ConversationSplit enum members gain a __call__ method so they satisfy
the protocol directly -- ConversationSplit.LAST_TURN(conversation) works.

This simplifies _split_conversation from an isinstance dispatch to a
single split(conversation) call.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Standardize on AZURE_AI_MODEL_DEPLOYMENT_NAME and fix Unicode in samples

- Replace FOUNDRY_MODEL with AZURE_AI_MODEL_DEPLOYMENT_NAME in all
  eval samples to match repo convention
- Replace Unicode symbols with ASCII equivalents in all eval sample
  print statements to avoid cp1252 encoding errors on Windows

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Update python/samples/03-workflows/evaluation/evaluate_workflow.py

Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>

* Apply suggestions from code review

Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>

* Rename ADR 0020 to 0023 (foundry evals integration)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: alliscode <bentho@microsoft.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>

35adfdb318 · 2026-03-31 15:53:06 +00:00

History

..

evaluate_workflow.py

Python: Foundry Evals integration for Python (#4750)

2026-03-31 15:53:06 +00:00