vss-deploy-video-embedding

This skill is an operational playbook for deploying, running, and integrating NVIDIA's VSS 3.2 GA RT-Embed Video Embedding microservice (Compose service `rtvi-embed`, container `vss-rtvi-embed`). It solves the problem of standing up the Cosmos-Embed1-448p embedding service correctly from a Docker Compose file, covering the GPU driver / NVIDIA Container Toolkit prerequisites, NGC registry login, environment-variable configuration, persistent model caches, and the long (20-minute) first-boot health grace period so that agents can bring the service up reliably instead of guessing at setup.

Once running, it documents the `/v1` REST API for generating text and video embeddings from uploaded files, HTTP/S3/file/data URLs, or live RTSP streams, plus health (`GET /v1/ready`) and metrics endpoints. It maps every host-side environment variable to its in-container name (image/tag, GPU device selection, RTSP reconnection tuning, OpenTelemetry, Kafka topics, Hugging Face and NGC cache volumes) and explains how to wire the service alongside Redis, Kafka, and OpenTelemetry. It also includes triage guidance for readiness, model-download, GPU, and stream-reconnection failures, plus clean teardown.

The target users are DevOps engineers, MLOps practitioners, and AI agents operating the NVIDIA Video Search and Summarization blueprint who need a repeatable standalone bring-up. Notably, the skill is security-conscious: it forbids interactive `sudo` prompts, prefers `sudo -n`, and explicitly refuses `chmod 777` / world-writable workarounds, stopping instead with the exact command a host owner must run.