VSS Video Embedding Deployment
Deploy and operate NVIDIA’s video embedding service for files, text, and live streams.
The documentation specifies least-privilege practices, non-interactive sudo, avoidance of chmod 777, and discloses NGC, Hugging Face, Redis, Kafka, RTSP, file:// data flows, and authentication boundaries. It also documents teardown, upgrade, and rollback. However, deployment starts Docker services and accesses external registries/model sources without explicit per-action confirmation; file:// exposure still depends on operator configuration; and the supplied license metadata is NOASSERTION, so points are deducted.
SKILL.md and its references cover startup, readiness, model download, permissions, streaming constraints, and common HTTP failures with reasonably coherent key paths. However, this is static review only; the committed evaluation has two tasks and reports 50% correctness, with no independently verifiable key-path execution or comprehensive automated test suite. The score therefore stays below the static ceiling and is reduced.
Trigger phrases, the boundary between standalone RT-Embed and the full VSS profile, inputs, outputs, and environment variables are fairly clear. However, Chinese-language support is not documented, first boot depends on overseas nvcr.io and huggingface.co services that may be difficult to reach from mainland China, and hardware/GPU resource limits remain partly unspecified, so points are deducted.
The material is layered across overview, deployment, verification, error handling, troubleshooting, upgrade/rollback, and references, and includes version 3.2.1, Apache-2.0, a product-repository pointer, and an environment matrix. However, the benchmark identifies missing recommended Instructions and Examples sections, author metadata is absent, changelog and maintenance/update responsibility are not clearly documented, and license metadata remains uncertain, so points are deducted.
The skill directly covers Docker Compose deployment, Cosmos-Embed1 text/video/RTSP embedding, readiness verification, and integration, with curl examples that are directly usable. The committed evaluation reports 100% security and approximately 84%–88% effectiveness, but only 50% correctness, no negative routing cases, and no static proof of real service outputs. Points are therefore deducted.
Versioned reference material, a benchmark report, routing evaluations, and a standalone deployment checklist provide limited auditability. However, there are no independently reproducible execution results, complete CI workflow, or committed tests covering the key paths available in the supplied material, and the evaluation sample has only two tasks, so evidence coverage is limited.
- First boot requires valid NGC credentials, potentially a Hugging Face token, and network reachability to nvcr.io and huggingface.co; mainland-China deployments should validate reachability and mirror/cache options first.
- Do not expose the unauthenticated API beyond loopback; restrict file:// access to the smallest media allow-list to avoid reading sensitive files visible inside the container.
- Obtain user confirmation before docker compose up, sudo chown, or down -v, including the target host, data paths, and cache-destruction consequences.
- The evaluation evidence is very small and does not replace deployment validation on real compatible hardware.
What does this skill do, and when should you use it?
This skill covers standalone deployment and operation of the VSS 3.2 GA RT-Embed video embedding microservice with Docker Compose. It documents the Cosmos-Embed1-448p model, file and text embeddings, HTTP/S3/file/data URLs, and live RTSP streams. It also explains optional Redis, Kafka, and OpenTelemetry integration, plus health checks, metrics, error handling, and troubleshooting. First boot can take about 20 minutes while model weights download and the Triton model repository is built.
Guides the agent through preparing NVIDIA GPU, Docker, NGC credentials, cache directories, and environment variables; starting the rtvi-embed Compose service; calling /v1/ready, /v1/models, /v1/files, /v1/generate_text_embeddings, /v1/generate_video_embeddings, and live-stream control endpoints; handling synchronous responses or SSE video-embedding output; inspecting logs, Prometheus metrics, asset statistics, and container status; diagnosing model-download, permission, GPU, Redis/Kafka connectivity, busy-service, and stream-reconnection failures; and performing upgrades, rollbacks, and teardown.
- A developer or operator needs to bring up the VSS video embedding microservice on an NVIDIA GPU host.
- An application team needs to chunk an uploaded video and generate Cosmos-Embed1-448p video vectors.
- A VSS integration developer needs text embeddings for text-to-video search.
- A systems integrator needs continuous embedding from a camera’s RTSP stream with chunked SSE output.
- A platform engineer needs to connect the service to Redis, Kafka, or OpenTelemetry.
- An operator needs to investigate /v1/ready 503 responses, model-download throttling, cache permissions, or CUDA memory failures.
What are this skill's strengths and limitations?
- Covers the operational path from Docker Compose deployment through teardown.
- Supports text embeddings, video-file embeddings, URL inputs, and live RTSP streams.
- Provides concrete health checks, error codes, logs, metrics, and troubleshooting guidance.
- Documents integration switches and interfaces for Redis, Kafka, and OpenTelemetry.
- Requires an NVIDIA GPU stack, Docker, and a valid NGC API key.
- First boot may take about 20 minutes and requires multi-GB persistent cache storage.
- Synchronous embedding jobs are constrained by service busy-state behavior and may return 503.
- The source provides no test suite, cost information, or platform compatibility matrix.
How do you install this skill?
Install it through the NVIDIA skills CLI: npx skills add nvidia/skills --skill vss-deploy-video-embedding --yes. The skill becomes available when the agent loads skills and encounters a relevant task. The source does not document the complete process for independently obtaining the VSS service repository itself.
How do you use this skill?
Ask the agent to “deploy vss-deploy-video-embedding RT-Embed” or “generate video embeddings from a live RTSP stream.” For standalone deployment, change to deploy/docker/services/rtvi/rtvi-embed, set RTVI_EMBED_PORT, VSS_DATA_DIR, NGC_API_KEY, and related variables, then run docker compose -f rtvi-embed-docker-compose.yml --profile bp_developer_search_2d up -d rtvi-embed. Verify readiness with curl against /v1/ready; use /v1/files and /v1/generate_video_embeddings for files, or /v1/generate_text_embeddings for text. Live streams require stream=true and chunk_duration>0, with results returned over SSE.