2026-09-26

Added device-specific visual context support for VLM workflow stages, improving scene understanding and workflow accuracy

Hub Workflows

Added device-specific visual context support for VLM workflow stages, improving scene understanding and workflow accuracy
Improved VLM stage behavior by supplying stable device state context during processing
Standardized workflow output digest generation to ensure consistent result tracking and comparison across services

Improved video analysis to distinguish stable scene context from changing activity, producing more focused and actionable observations
Added support for operator-defined “stable states” so recurring camera scenes can be learned and future alerts emphasize deviations and new activity
Enhanced generated markers with clearer subject/action separation, more accurate tagging, and improved event timelines
Recording metadata now includes structured VLM scene and observation data for richer downstream workflows and integrations
Improved handling of recordings with no significant activity while still preserving useful scene context
Increased reliability of observation generation by filtering incomplete detections and validating scene summaries before metadata updates