NVIDIA VSS (Video Search & Summarization)
NVIDIA VSS is a blueprint, not a single container: a video-ingestion engine orchestrating four model services. Cordatus launches all of them as one deployment, and lets each service be a new container, a container you already have running, or a remote endpoint.
LLM-Launcher/vss-launch.mp4 — Applications → NVIDIA VSS → device and version → the numbered
service rail, setting each service's source → Review & start → the resulting container group.
The five services
| # | Service | What it does |
|---|---|---|
| 1 | VSS | The blueprint engine itself — ingests video, orchestrates the other four services and serves the UI. |
| 2 | VLM | Vision-language model that captions video segments. Needs a model and a GPU. |
| 3 | LLM | Summarises captions and answers questions. A remote endpoint is fine here. |
| 4 | Embed | Turns captions into vectors for retrieval. Small model, small GPU footprint. |
| 5 | Rerank | Re-orders retrieved passages before the LLM answers. Optional, but it improves question answering. |
Launching
Steps 1 and 2 are the same as any application — see the Launch Guide:
- Select Device
- Select Version — the version you pick here pins the blueprint engine's image.
- Advanced Configuration
The service rail
The third step's rail is numbered, one entry per service, then Review & start under a Finish heading. The order matters and is part of how the pipeline is described, so each entry carries its number rather than an icon.
Every entry shows:
- what the service currently resolves to — a new container, an existing one, or a URL;
- a pill naming its mode;
- a tick when it is complete, or a warning when it is not.
At the foot of the rail, the readiness card names the first thing each service is missing (for example "VLM · model missing"), and its heading becomes Fix 2 services when something is outstanding.
Configuring one service
Each service pane starts with Where this service comes from:
| Mode | Meaning |
|---|---|
| Create new | A fresh container on the selected device. |
| Assign existing | Reuse a container already running there. Only containers that actually serve a model can stand in. |
| Remote endpoint | An OpenAI-compatible URL, with Test connection and an optional Access Token. |
Not every mode is available for every service — the VSS engine itself is always created.
When a service is set to Create new, the pane below it is the ordinary configuration rail for that container: Model, Compute, Overrides, exactly as described in the Launch Guide. The engine image is pinned by Select Version and is labelled as such.
If nothing on the device can fill a role:
Nothing on {device} serves this role yet. Create a new one, or point the service at a remote endpoint.
Event Reviewer
The VSS service has an extra Event Reviewer switch, which adds the event-review component to the deployment.
Resolved services
Above the rail, Resolved services shows what each of the five will actually be once the screen is done — new containers, containers already running on the device, and remote endpoints — so the whole pipeline can be read at a glance before you commit.
Review & start
The review pane for VSS shows the resolved service map rather than a single command, because the deployment is several containers rather than one. Underneath, Containers to be created lists every container the launch will create, with its image and model; services set to Assign existing or Remote endpoint are not listed, because nothing new is created for them.
Blockers appear first and each is a button that jumps to the service that needs attention.
After launching
All five services form one container group on the Containers page. Open the group to reach each service's logs, parameters and ports.
The VSS container is the one that serves the UI — use its Ports tab for the local address, or generate a public URL from there.