How It Works#
The operator uses the VSG agent console in a browser to interact with Argus, the VSG agent. Argus runs on a Hermes agent server, reasons with a local model, and calls VSS tools through the VSS Model Context Protocol (MCP) server. The console renders Argus’s results and lets operators review them.
Architecture#
Component |
Role |
|---|---|
vsg-ui |
Serves the React application and proxies an allowlist of Hermes and VSS paths. |
Hermes agent server |
Runs the VSG agent, keeps one session per investigation, and streams run events over Server-Sent Events (SSE). |
VSG Hermes skills |
Provide the |
VSS MCP server |
Provides the agent with VSS tools. |
VSS gateway |
Stores and indexes footage, runs searches, and serves clips and frames. |
Model server |
Provides the LLM used by the VSG agent, such as OpenVINO™ model server. |
Investigation Flow#
Search: The UI uploads the photo to VSS and starts a VSG agent run. The VSG agent calls
vss_search_video.Review: The UI displays VSS results in a grid. The operator confirms the sightings that match. If at least one is confirmed, the other shown matches are sent to Argus as false positives; if none are confirmed, no feedback is sent.
Crop and search again (optional): The operator can crop a person from a clip and submit the crop as a new image search. The crop becomes the subject image.
Route: Confirmed sightings from all grids appear in chronological order as a route on the floor plan.
Live search (optional): When a live camera is connected, the operator can start live search from the route card by clicking the live-search control. The UI polls VSS every 10 seconds.
Report: The operator can generate a report from the route card. The VSG agent drafts the report, and the browser builds the PDF.