Skip to main content

3.25 Release Notes

3.25.3​

Improvements​

  • Allow configuring max retry, max backoff and max timeout for Custom API Endpoint AI System
  • Allow skipping custom endpoint AI system validation via a feature flag
  • Support setting max num input characters for DynamoEval Tests

Bug & Security Fixes​

  • Fix Dynamoguard monitoring logs ordering

3.25.2​

Improvements​

  • Bring back the dynamoguard monitoring dashboard in the UI

3.25.1​

Improvements​

  • Add support to allow users to set custom headers in Custom API Endpoint AI System

3.25.0​

New Features​

  • [Guard] Policy Augmentation: When adding training data to custom content policies, a new "augment" workflow provides a more targeted way to improve policies. Augmentation is recommended for smaller dataset additions (<50 datapoints), while regeneration is recommended for larger additions. See documentation for more details.
  • [Guard] Hallucination: New Response Faithfulness model optimized for high-volume usage scenarios. Response Faithfulness Standard offers the existing frontier-model faithfulness model, while Response Faithfulness High Volume is a custom-trained small language model guardrail. This guardrail outputs faithfulness subtypes to categorize faithfulness errors.
  • [Guard] Chinese input custom content policies — This is a Beta feature allowing users to create custom content policy guardrails in Chinese.
  • [Platform] Added Gemini AI System support
  • [Platform] API Key Management: clearer way to manage API Keys for your account.
  • [Platform] Add APIs for probes and detailed health of components

Improvements​

  • [Guard] Clearer status descriptions for when policy deployment fails.
  • [Eval] Improved Testing Tab UI

Bug & Security Fixes​

  • Add CORS policy in keycloak
  • Various minor stability and UI fixes
  • Address memory leak in data processing worker
  • Fix policy deployment status
  • Major robustness and stability fixes to Evaluation pipelines