Version 1.31.0
about 4 hours ago by Tech admin
Agents
- Monitoring Templates: To streamline the process for users who want to quickly govern their agents, we’ve introduced a new set of monitoring templates. These provide instant access to common, pre-configured alerting use cases—such as low throughput, guardrail violations, and high latency—allowing you to move from deployment to active monitoring in seconds.
- Unpublish Agent: Users can now undeploy their agents without the need to archive or delete them. This allows you to take an agent "offline" while maintaining full access to its settings and configurations for further iteration.
- UX Improvements: To prevent data loss, we’ve added a notification system for unsaved changes. Since agent configurations are only finalized upon versioning and deployment, users will now be alerted if they attempt to navigate away with unpublished modifications. Claude Integration Fix: Resolved the top_p parameter issue reported by the community last week, ensuring stable performance for Claude-based agents.
Observability
- Native Image Support in Datasets: To support the growing variety of enterprise AI use cases, our datasets now officially support the Image data type. Users can now log images alongside structured or semi-structured (JSON) properties, providing a complete multimodal view of their model data.
Guardrails
- Automated Observability Instrumentation: Similar to our agent instrumentation, guardrails are now automatically integrated into our observability services. By default, every guardrail interaction is logged into a dedicated dataset containing detailed, verbose logs of each check and intermediate calculation. Users can opt-out of this via settings. Out-of-the-Box Dashboards: As part of this instrumentation, a pre-configured dashboard is automatically generated in the Overview section to monitor guardrail performance and violation trends in real-time.
- New Image-Based Checks: We have added four sophisticated new rules to our guardrail library to ensure visual data integrity: Image Resolution: Ensures visual inputs or outputs meet minimum height and width requirements to prevent low-quality processing. Sharpness Level: Uses the Variance of the Laplacian method to quantify and filter out blurry or out-of-focus images.
- Semantic Image: Categorizes visual content (e.g., Photograph, Document Scan, Meme) to ensure only contextually relevant images are processed. Image Tampering Detection: Identifies digital manipulations or "deepfakes" by analyzing pixel-level signals for authenticity.