Azure/azure-sdk-for-java com.azure+azure-ai-projects_2.7.0

on GitHub

2.7.0 (2026-10-06)

Features Added

• Added EvaluatorsClient and EvaluatorsAsyncClient, built directly through
  AIProjectClientBuilder.buildEvaluatorsClient() and buildEvaluatorsAsyncClient(), for generally available
  evaluator-version management and rubric-generation jobs. Preview evaluator definitions such as
  EndpointBasedEvaluatorDefinition require allowPreview(true). BetaEvaluatorsClient and
  BetaEvaluatorsAsyncClient retain preview pending-upload and credential operations.
• Evaluation data generation graduated to general availability on DatasetsClient and DatasetsAsyncClient.
  Added scenario-specific input and result models for evaluation, supervised fine-tuning, and reinforcement
  fine-tuning. Fine-tuning scenarios remain in preview and require allowPreview(true); their output configurations
  support merge-file IDs.
• Added ConnectionType.OPEN_API (OpenAPI) and ConnectionType.REMOTE_A2A (RemoteA2A) for listing and identifying
  OpenAPI and remote agent-to-agent connections.
• Added ApiError.getMisalignment() to expose structured misalignment details using the OpenAI
  ErrorObject.Misalignment model.

Breaking Changes

• Removed BetaDatasetsClient, BetaDatasetsAsyncClient, and their beta-builder methods. Use DatasetsClient /
  DatasetsAsyncClient, built directly through AIProjectClientBuilder, for data generation job operations.
  Default requests do not send a preview feature header; set allowPreview(true) for preview data generation.
• Moved evaluator-version CRUD, listing, and evaluator-generation job operations from BetaEvaluatorsClient /
  BetaEvaluatorsAsyncClient to EvaluatorsClient / EvaluatorsAsyncClient. Build the new clients directly through
  AIProjectClientBuilder; generally available operations no longer require the Evaluations=V1Preview header.
• Data generation creation methods now accept DataGenerationJobInputs instead of DataGenerationJob. Construct
  EvaluationDataGenerationJobInputs, SupervisedFineTuningDataGenerationJobInputs, or
  ReinforcementFineTuningDataGenerationJobInputs with a name, sources, and generation configuration.
  DataGenerationJob is now an immutable response model with top-level getName(), getSources(), and
  getGenerationConfiguration() properties; its no-argument constructor and getInputs() / setInputs(...) were
  removed.
• Renamed DataGenerationJobOptions and its SimpleQnA, SimulationSeed, ToolUseFineTuning, and Traces subtypes to
  the corresponding *DataGenerationJobConfiguration types. DataGenerationJobInputs.getOptions() was replaced by
  getGenerationConfiguration(). DataGenerationJobOutputOptions and the getOutputOptions() /
  setOutputOptions(...) accessors were replaced by scenario-specific *DataGenerationJobOutputConfiguration types
  and getOutputConfiguration() / setOutputConfiguration(...) accessors on concrete input models.
• Renamed DataGenerationJobScenario.SUPERVISED_FINETUNING and REINFORCEMENT_FINETUNING to
  SUPERVISED_FINETUNING_PREVIEW and REINFORCEMENT_FINETUNING_PREVIEW. Their wire values now use the
  supervised_finetuning_preview and reinforcement_finetuning_preview discriminator values.
• Renamed DataGenerationJobResult.getGeneratedSamples() to getGeneratedSampleCount().
• Evaluator generation creation methods now accept EvaluatorGenerationInputs instead of EvaluatorGenerationJob.
  EvaluatorGenerationJob is now an immutable response model with top-level source, model, and evaluator-name
  properties; its no-argument constructor and getInputs() / setInputs(...) were removed.
• Changed public constructors on polymorphic base models to protected: DataGenerationJobInputs,
  DataGenerationJobSource, EvaluationTaxonomyInput, EvaluatorDefinition, EvaluatorGenerationJobSource,
  InsightRequest, RecurrenceSchedule, RoutineAction, RoutineDispatchPayload, RoutineTrigger, ScheduleTask,
  TargetConfig, and Trigger. The renamed DataGenerationJobConfiguration base also has a protected constructor.
  Construct concrete subtypes instead; deserialization of unknown discriminator values remains supported.

Bugs Fixed

• Fixed OpenAI clients built from AIProjectClientBuilder to acquire bearer tokens through the Azure HTTP pipeline.
  Asynchronous clients now use asynchronous token acquisition instead of calling TokenCredential.getTokenSync(...);
  synchronous clients continue to use synchronous token acquisition.
• Fixed skill file uploads to omit an unset default flag instead of sending the literal text null.

Other Changes

• Organized Java samples into feature-specific folders, with synchronous and asynchronous samples together.
• Added synchronous and asynchronous evaluation samples for inline data, uploaded JSONL datasets, and native OpenAI graders.
• Added synchronous and asynchronous evaluator catalog and rubric generation samples, including metadata updates, version review, evaluation runs, and cleanup.
• Updated data generation samples for the scenario-specific input, configuration, and output models.
• Regenerated the client from the updated TypeSpec specification.

Dependency Updates

• Updated azure-ai-agents from 2.6.0 to 2.7.0.
• Updated openai-java from 4.45.0 to 4.69.0.