rt-vlm

Generate dense video captions from RTSP live streams by registering the source via /v1/streams/add, then calling /v1/generate_captions_alerts with streaming enabled to receive timestamped outputs and anomaly incident alerts.

Kafka integration for video captioning works by publishing caption, incident, and error events to dedicated topics during stream processing, allowing downstream consumers to monitor and validate outputs from the /v1/generate_captions_alerts endpoint.

You can use OpenAI-compatible chat completions with VLM video captioning via the /v1/chat/completions endpoint, which supports NIM-compatible requests alongside the core /v1/generate_captions_alerts and /v1/files endpoints.

Uploading stored video files for timestamped captions requires using the /v1/files endpoint to upload media and obtain a file ID, then passing that ID to /v1/generate_captions_alerts to retrieve dense chunked caption outputs.

The VSS 3.1 VLM API supports Server-Sent Events for streaming caption outputs by enabling the stream parameter when calling /v1/generate_captions_alerts, returning chunked dense captions and alerts incrementally as they are processed.

Manage and delete active RTSP stream sources for video captioning by registering new feeds through /v1/streams/add and removing existing sources using the /v1/streams/delete/{stream_id} endpoint within the VSS 3.1 microservice.