* .NET: feat(evals): add ground_truth/expected_output support for workflow eval
Brings .NET to parity with Python PR #5234 for issue #5135:
- Add expectedOutput parameter to Run.EvaluateAsync (workflow) and stamp on the overall EvalItem.ExpectedOutput.
- Map EvalItem.ExpectedOutput -> ground_truth in the Foundry JSONL payload, item_schema, and data_mapping for similarity.
- Add GroundTruthEvaluators set (currently builtin.similarity) and a FindMissingGroundTruthEvaluators helper.
- Fail fast with InvalidOperationException when a ground-truth evaluator is selected but no item provides an ExpectedOutput, instead of surfacing a remote provider error.
- Add tests in FoundryEvalConverterTests and WorkflowEvaluationTests.
- Add Evaluation_WorkflowExpectedOutputs sample (workflow + Foundry similarity).
Fixesmicrosoft/agent-framework#5135 (.NET side).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address review: relax BuildOverallItem events to IReadOnlyList<WorkflowEvent>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Sample: disable per-agent breakdown when using reference-based evaluator
Per-agent EvalItems are intentionally left without ExpectedOutput, so the new fail-fast validation in FoundryEvals would throw when Similarity is invoked for per-agent items. Pass includePerAgent: false in the workflow + similarity sample, and document this gotcha in the EvaluateAsync XML doc.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix BuildOverallItem: fall back to last ExecutorCompletedEvent
AgentResponseEvent is only emitted when AIAgentHostOptions.EmitAgentResponseEvents is enabled, which is not the default for WorkflowBuilder(agent).AddEdge(...). When it is absent, fall back to the last non-internal ExecutorCompletedEvent whose Data is an AgentResponse / ChatMessage / string so the overall EvalItem (and any expectedOutput) is produced. Without this, samples wired up the standard way returned 0 evaluation items.
Update test to cover the fallback path.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Sample: enable EmitAgentResponseEvents; eval throws clear error when no overall response found
Root cause of '0 results': AIAgentHostExecutor only emits AgentResponseEvent when AIAgentHostOptions.EmitAgentResponseEvents is true (default false). For ordinary AIAgent executors the runtime's ExecutorCompletedEvent.Data is null, so the prior fallback couldn't find a final response either.
Sample now builds executors with EmitAgentResponseEvents=true via BindAsExecutor(hostOptions). EvaluateAsync now throws InvalidOperationException with a remediation hint when the user supplies expectedOutput but no overall final response can be located, instead of silently returning 0/0.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Guard against null sample/error/usage/datasource_item in ParseDetailedItem
Foundry eval responses can have these properties present with JSON null
or non-object values, which caused JsonElement.TryGetProperty to throw
'requires Object, has Null'. Check ValueKind == Object before drilling in.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address PR review: reorder expectedOutput, tighten ground-truth check, add fail-fast test
* WorkflowEvaluationExtensions.EvaluateAsync: move 'expectedOutput' to
after 'splitter' so the original positional contract of (splitter,
cancellationToken) is preserved for existing callers.
* FoundryEvals: require ALL items to carry ExpectedOutput when a
ground-truth evaluator is selected (e.g. similarity), not just any.
Reference-based evaluators score per-item, so a single missing GT
would still surface as a provider-side validation error. Updated
fail-fast message accordingly.
* WorkflowEvaluationTests: add EvaluateAsync_WithExpectedOutputButNoFinalResponse_ThrowsAsync
to verify the InvalidOperationException is thrown (and that the
message mentions EmitAgentResponseEvents).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fail-fast on missing overall item regardless of expectedOutput; harden BuildOverallItem default
* EvaluateAsync now throws InvalidOperationException whenever 'includeOverall'
is requested but BuildOverallItem cannot produce an item, instead of only
when 'expectedOutput' is supplied. Same misconfiguration (agents not bound
with EmitAgentResponseEvents) used to silently return empty results — now
it surfaces a clear, actionable error in both cases.
* BuildOverallItem switch default now throws instead of returning null. The
preceding for-loop already constrains Data to AgentResponse/ChatMessage/
string, so reaching default would indicate a contract drift; throw to make
the bug visible.
* Test renamed and broadened to verify the throw fires without expectedOutput.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: alliscode <25218250+alliscode@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test: Split out Handoff Orchestration tests
* fix: Synthesized Handoff FunctionResult is never sent to agent
When we receive a handoff request from the agent, we need to service it outside of the Agent Loop to terminate the loop. What this means is that we take ownership of terminating the call by feeding the result back into the agent on a subsequent invocation.
When we refactored Handoff to support HITL and make use of AgentSession, we inadvertantly removed this step, causing subsequent invocations to the Handoff agent to fail (first works, but breaks the state).
The fix is to be more precise about the agent's bookmark when concatenating the result of agent invocation to the shared conversation history.
* test: Add unit tests for Handoff FunctionCall/Result matching fix
When MultiPartyConversation gets saved during checkpointing, the data for the chat history is not persisted, resulting in failures to deserialize after. The fix is to make the history visible to the source generated serialization code.
* refactor: remove dead code
* refactor: remove ignore YieldsMessageAttribute
- the correct one to use is YieldsOutputAttribute
- fixes a comment that mistakenly refers to `.YieldsMessage()` which does not exist.
* fix: ChatForwardingExecutor does not use correct role for string messages
- make ChatForwardingExecutor use its configured role for string messages rather than always use ChatRole.User
- add ChatForwardingExecutor tests
* fixup: remove unused attribute
* test: Add tests for failure when .AsAgent used on a non-ChatProtocol workflow
* test: Add FunctionExecutor tests
- also fixes Send and YieldOutput type registration for synchronous output-returning delegates
* test: Suppress CodeCoverage for obsolete names
* fix: Re-add Obsolete attributes
- avoid hard-breaking change
- properly notify users that these attributes get ignored
Some providers, e.g. Gemini, do not use the CallId mechanism to disambiguate simultaneous function calls. This can result in message lists containing multiple turn to fail to filter properly.
The fix is to take advantage of the expectation that Handoff Orchestration is a "single-speaker" flow, which only has a single active AIAgent per "turn" and an agent's turn is not finished until all outstanding function calls are finished.
This allows us to expect that any ambiguous-CallId FunctionCallContent are either in separate turns or will have had a response before the next issued call with the same Id.
* fix: Add session support for Handoff-hosted Agents
In order to better support using `Workflows` hosted as `AIAgents` inside of Handoff workflows, we need to make proper use of AgentSession. This causes potential issues around checkpointing and making sure that we properly compute only the new incoming messages for each agent invocation.
* fix: AgentSession checkpointing using AIAgent's Serialize/Deserialize methods
We cannot rely on implicit serialization through `HandoffHostState` because we are missing type information.
* fix: Thread safety issue in `MultiPartyConversation.AllMessages`
* fix: Enable unwrapping of FunctionResultContent when ExternalRequest was wrapped into FunctionCallContent
* fix: Foundry Agents without description in Handoff
Foundry Agents without a description set will return an empty string (rather than null) for the description. This was breaking the fallback logic for `handoffReason`.
* test: Add unit tests
* Foundry Evals integration for .NET
- Core evaluation framework: EvalItem, LocalEvaluator, FunctionEvaluator, EvalChecks
- IAgentEvaluator interface with MeaiEvaluatorAdapter bridge
- AgentEvaluationExtensions for agent.EvaluateAsync() overloads
- FoundryEvals wrapping MEAI quality/safety evaluators
- ConversationSplitters (LastTurn, Full) and IConversationSplitter
- EvalItem.PerTurnItems() for multi-turn decomposition
- HasImageContent for multimodal content detection
- WorkflowEvaluationExtensions for per-agent workflow evaluation
- 7 eval samples mirroring Python parity:
02-agents/Evaluation: SimpleEval, ExpectedOutputs, Multimodal
03-workflows/Evaluation: WorkflowEval
05-end-to-end/Evaluation: FoundryQuality, MixedProviders, ConversationSplits
- Comprehensive unit tests (1958 passing)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Rewrite FoundryEvals to use real Foundry Evals API
Replace MEAI evaluator shim with actual OpenAI EvaluationClient protocol
methods. FoundryEvals now creates eval definitions, submits runs, polls
for completion, and fetches per-item results server-side.
- New constructor: FoundryEvals(AIProjectClient, model, evaluators)
- Add FoundryEvalConverter for MEAI ChatMessage -> Foundry JSON format
- Add EvalId, RunId, ReportUrl to AgentEvaluationResults
- All 20 built-in evaluator constants now work (agent, tool, quality, safety)
- Remove Microsoft.Extensions.AI.Evaluation.Quality/Safety dependencies
- Update all samples for new constructor (no more ChatConfiguration)
- Replace BuildEvaluators tests with ResolveEvaluator tests
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add response output to CustomEvals and ExpectedOutputs samples
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address review: pagination, validation, error handling, tests
FoundryEvals fixes:
- Add pagination for output items (has_more/after cursor)
- Add guard clauses for pollIntervalSeconds/timeoutSeconds <= 0
- Fix double TryGetProperty for passed field parsing
- Throw on all-tool-evaluators with no tool definitions
- Fix XML doc (default 300s, not 180s)
New tests (30 added, 1989 total):
- EvalChecks: NonEmpty, ContainsExpected (pass/fail/skip/case),
HasImageContent, ToolCallsPresent
- FoundryEvalConverter: ConvertMessage (text, image, function call,
function results fan-out, empty fallback, mixed content),
ConvertEvalItem, BuildTestingCriteria (quality/agent/tool/groundedness
data mappings), BuildItemSchema
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix review: null-refs, Data.ToString() bug, ContainsExpected, add tests
- Fix NullReferenceException in sample Response display (pattern matching)
- Fix WorkflowEvaluationExtensions Data?.ToString() producing type names
instead of message text (pattern-match ChatMessage/AgentResponse/list)
- Change EvalChecks.ContainsExpected to return Passed=false when no
ExpectedOutput (was silently passing, masking misconfiguration)
- Add EvalItem constructor tests with LastTurn/Full/null splitters
- Add FoundryEvalConverter.ConvertMessage DataContent (base64 image) test
- Add ExtractAgentData tests with ChatMessage, list, and AgentResponse data
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix review: conversation fidelity, eval caching, fallback tests
- WorkflowEvaluationExtensions: preserve full response messages (tool calls,
intermediate) instead of synthetic 2-message conversation. Cast completed
Data to AgentResponse and use Messages when available, fallback to text.
- FoundryEvals: cache evalId per schema shape (hasContext, hasTools) so
subsequent EvaluateAsync calls create runs under the same eval definition.
- MeaiEvaluatorAdapter: code already correctly passes queryMessages (not full
conversation) to IEvaluator — no change needed, verified by inspection.
- Add tests: AgentResponse full messages preservation, unknown object
ToString() fallback for ExtractAgentData.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Rename AzureAI→Foundry: move eval files, update references
- Move FoundryEvals.cs and FoundryEvalConverter.cs from
Microsoft.Agents.AI.AzureAI to Microsoft.Agents.AI.Foundry
- Update namespace from AzureAI to Foundry in both files
- Add explicit usings required by Foundry project (no implicit usings)
- Move FoundryEvalConverter tests to Foundry.UnitTests project
(avoids ReplacingRedactor type conflict from dual project refs)
- Update all sample csproj references and using statements
- Remove Foundry project reference from AI UnitTests
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* PR review round 4: wire up tool extraction, remove eval cache, fix null safety
- BuildEvalItem: extract tools from agent via GetService<ChatOptions>() into EvalItem.Tools (Python parity)
- FoundryEvals: remove eval ID cache - each call creates fresh definition (matches Python behavior)
- FoundryEvals: replace null-forgiving operators with descriptive InvalidOperationException
- MixedProviders sample: remove unnecessary explicit PackageReferences (transitively provided)
- FoundryEvalConverter: document that tool results take precedence over text content
- Add LocalEvaluator zero-checks test documenting 0 metrics = failed behavior
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python-dotnet parity: 9 feature gaps filled
New checks:
- ToolCallArgsMatch() — verify tool call names + argument subset match
- ToolCalledCheck(ToolCalledMode.Any, ...) — match any of the specified tools
- ToolCalledMode enum (All/Any)
FoundryEvals enhancements:
- Default evaluators now [Relevance, Coherence, TaskAdherence] (was Relevance, Coherence)
- Auto-add ToolCallAccuracy when items have tool definitions
- EvaluateTracesAsync — evaluate by response_ids, trace_ids, or agent_id
- EvaluateFoundryTargetAsync — evaluate deployed Foundry targets
Result type enrichment:
- AgentEvaluationResults: added Status, Error, PerEvaluator, DetailedItems
- New EvalItemResult/EvalScoreResult/PerEvaluatorResult types
- FoundryEvals populates all new fields from API responses
Workflow fix:
- Skip internal executors (_*, input-conversation, end-conversation, end)
Tests: 8 new tests covering ToolCallArgsMatch, ToolCalledMode.Any, internal executor filtering
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add MeaiEvaluatorAdapter and PerTurnItems edge case tests
- 3 tests for MeaiEvaluatorAdapter: query message forwarding, synthetic
response fallback, multiple items aggregation
- 3 tests for EvalItem.PerTurnItems: empty conversation, no user messages,
system+assistant only
- StubEvaluator and StubChatClient test helpers
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Blocking link check for outdated package in DevUI.
* Replace Dictionary<string, object> payloads with typed wire models
Introduce internal FoundryEvalWireModels.cs with compile-time-safe types
for the OpenAI Evals API wire format. The OpenAI .NET SDK (2.9.1) only
provides protocol-level methods with BinaryContent/ClientResult — no
typed request models. These internal models replace scattered dictionary
literals with [JsonPropertyName]-annotated classes, giving:
- Compile-time safety (typos become build errors)
- Single point of change when the API evolves
- IntelliSense discoverability
- Cleaner serialization via JsonPolymorphic for content items
Models: WireContentItem hierarchy (text, image, tool_call, tool_result),
WireMessage, WireEvalItemPayload, WireTestingCriterion, WireItemSchema,
WireCreateEvalRequest, WireCreateRunRequest, and data source variants.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Skip metric when Foundry returns neither score nor passed
When an evaluator returns no score and no passed value, the previous
code created BooleanMetric(name, false), which falsely failed items
via ItemPassed. Now we skip the MEAI metric entirely for indeterminate
results — the raw data remains available in DetailedItems for diagnostics.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address PR #4914 review comments: fix tool evaluator bug and add tests
- Fix duplicate ToolCallAccuracy: resolve evaluator names before checking
against ToolEvaluators set (Comment 2)
- Make FilterToolEvaluators internal for testability; add tests for the
ArgumentException edge case when all evaluators are tool-type (Comment 3)
- Add CancellationToken test for LocalEvaluator (Comment 4)
- Add EvaluateAsync integration test on Run with sequential workflow and
per-agent SubResults verification (Comment 5)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address Peter's review comments on PR #4914
- Add trailing newline to Evaluation_FoundryQuality.csproj (Comment 6)
- Make evaluator name lookups case-insensitive: switch BuiltinEvaluators,
ToolEvaluators, AgentEvaluators, and ResolveEvaluator's StartsWith check
from Ordinal to OrdinalIgnoreCase (Comment 7)
- Add Trace.TraceWarning when Foundry returns fewer results than submitted
items, indicating expected vs actual count before padding (Comment 8)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add Microsoft.Extensions.AI.Evaluation packages to Directory.Packages.props
These were removed in #5269 as unused, but are needed by the Foundry
and core evaluation integration added in this PR.
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>
