Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use to deploy the vss-behavior-analytics service standalone (entrypoint, config-source, optional calibration). Not for the full warehouse deploy.
.claude/skills/nvidia-vss-setup-behavior-analytics/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 57% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 484% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 101% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 4% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 13% | 0% |
Deploy the behavior-analytics service standalone with the user's chosen entrypoint, config, and calibration.
Follow the routing tables and step-by-step workflows below. Each section that ends in workflow, quick start, or flow is intended to be executed top-to-bottom. Detailed reference material lives in references/.
Worked end-to-end examples are kept under evals/ (each *.json manifest contains a runnable scenario). Run a Tier-3 evaluation to replay them:
bashnv-base validate skills/vss-setup-behavior-analytics --agent-eval
A minimal standalone bring-up looks like:
bashcd $REPO/deploy/docker export VSS_APPS_DIR=$(pwd) docker compose -f services/analytics/behavior-analytics/compose.yml up -d vss-behavior-analytics-base
Follow references/deploy-behavior-analytics-service.md for the full workflow (entrypoint pick, config source, dynamic updates).
/docs or /health; redeploy via vss-deploy-profile or the matching vss-deploy-* skill.NGC_CLI_API_KEY. Solution: docker login nvcr.io and re-export the key before retrying.docker compose down.Deploy just the vss-behavior-analytics container (the spatial-AI analytics pipeline from the upstream behavior-analytics repo), not as part of the full warehouse blueprint stack.
The full operational walkthrough — entrypoint table, config-source options, calibration types, dynamic-update wire contract, troubleshooting — is references/deploy-behavior-analytics-service.md. This SKILL.md only handles routing and prerequisites.
$VSS_APPS_DIR pointing at <repo>/deploy/docker/. Required by the service compose's volume binds.$NGC_CLI_API_KEY set so docker can pull the image. See references/ngc-api-key-registry-login.md.docker --version and docker compose version.restart: always cycles the container. Status will show Restarting (N) in docker ps until a broker is reachable. With a broker, dynamic config / dynamic calibration over mdx-notification become available.If any required prerequisite fails, surface the gap before going further.
Hand the user references/deploy-behavior-analytics-service.md and walk them through its steps in order:
The compose-file edits, YAML diffs, deploy + verify commands, and troubleshooting table all live in that reference — don't duplicate them here.
Once the container is up and a broker is reachable, two runtime-update flows are available — neither requires redeploying:
Publish an upsert (per-key patch) or upsert-all (full snapshot) message to the mdx-notification topic with Kafka key behavior-analytics-config and headers:
event.type: upsert | upsert-all | request-config | ackreference-id: video-analytics-api-<uuid> (web-api originated), behavior-analytics-<uuid> (bootstrap reply), or the source-type literal (kafka / redis / mqtt) for direct-publisher upserts.Body: {"status": ..., "config": <patch>, "error": ...}.
The listener validates each message at the envelope layer (rejects unknown keys, missing config, malformed status/error) and at the per-payload layer (rejects forbidden sections, bad item shapes). Successful upserts are persisted to disk, applied to every worker, and ACK'd back over the topic.
Full wire contract + ack semantics: references/dynamic-config.md.
Publish to the same topic with Kafka key calibration and headers:
event.type: upsert-all (full snapshot) | upsert (per-sensor merge) | delete (per-sensor removal)timestamp: ISO-8601 UTC (YYYY-MM-DDTHH:MM:SS.fffZ).Body: JSON sensor list (and ROIs / tripwires / homographies for upsert-all).
The listener validates against the vendored AJV schema before persisting. Schema violations log a calibration schema violation warning and are dropped — the previously-good calibration stays loaded.
Full wire contract + per-action validation policy: references/dynamic-calibration.md.
Both flows live entirely on the broker — the producer can be video-analytics-api, your own script, or any Kafka client that mirrors the wire shape. They're the recommended way to change configuration after the container is running, so the operator doesn't have to redeploy.
vss-deploy-profile with profile warehouse (or alerts). Don't run this skill in parallel.references/configuration.md, references/dynamic-config.md, or references/dynamic-calibration.md before editing the JSON.bump:1
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 27,166 | 11,210 | -59% | 1 | 1 | 0% | 2,703 | 4,256 | +57% | 0 | 0 | — |
case-02 | fail→pass | 18,924 | 6,902 | -64% | 1 | 1 | 0% | 567 | 3,309 | +484% | 0 | 0 | — |
case-03 | fail→pass | 18,963 | 5,625 | -70% | 1 | 1 | 0% | 1,543 | 3,107 | +101% | 0 | 0 | — |
case-04 | fail→fail | 12,320 | 7,190 | -42% | 1 | 1 | 0% | 2,356 | 3,182 | +35% | 0 | 0 | — |
case-05 | fail→pass | 12,639 | 2,917 | -77% | 1 | 1 | 0% | 2,359 | 2,456 | +4% | 0 | 0 | — |
case-06 | fail→pass | 12,355 | 5,306 | -57% | 1 | 1 | 0% | 2,501 | 2,814 | +13% | 0 | 0 | — |
case-07 | fail→pass | 13,477 | 2,984 | -78% | 1 | 1 | 0% | 2,706 | 2,346 | -13% | 0 | 0 | — |
case-08 | fail→pass | 7,450 | 3,550 | -52% | 1 | 1 | 0% | 1,391 | 2,435 | +75% | 0 | 0 | — |
case-09 | fail→pass | 11,597 | 3,018 | -74% | 1 | 1 | 0% | 1,971 | 2,412 | +22% | 0 | 0 | — |
case-10 | fail→pass | 8,532 | 3,822 | -55% | 1 | 1 | 0% | 1,609 | 2,404 | +49% | 0 | 0 | — |
case-11 | pass→pass | 10,313 | 2,071 | -80% | 1 | 1 | 0% | 1,755 | 2,235 | +27% | 0 | 0 | — |
case-12 | pass→pass | 9,855 | 2,053 | -79% | 1 | 1 | 0% | 1,647 | 2,206 | +34% | 0 | 0 | — |
case-13 | fail→pass | 11,660 | 3,918 | -66% | 1 | 1 | 0% | 1,933 | 2,466 | +28% | 0 | 0 | — |
case-14 | fail→pass | 13,930 | 5,048 | -64% | 1 | 1 | 0% | 2,401 | 2,865 | +19% | 0 | 0 | — |
case-15 | pass→pass | 7,384 | 2,480 | -66% | 1 | 1 | 0% | 1,294 | 2,301 | +78% | 0 | 0 | — |
case-16 | fail→pass | 11,255 | 3,690 | -67% | 1 | 1 | 0% | 1,803 | 2,473 | +37% | 0 | 0 | — |
case-17 | pass→pass | 8,835 | 5,256 | -41% | 1 | 1 | 0% | 1,674 | 2,849 | +70% | 0 | 0 | — |
case-18 | fail→pass | 10,099 | 7,221 | -28% | 1 | 1 | 0% | 2,037 | 3,470 | +70% | 0 | 0 | — |
case-19 | fail→fail | 17,720 | 5,934 | -67% | 1 | 1 | 0% | 1,836 | 2,967 | +62% | 0 | 0 | — |
case-20 | fail→pass | 11,470 | 4,692 | -59% | 1 | 1 | 0% | 2,077 | 2,782 | +34% | 0 | 0 | — |
case-21 | fail→pass | 9,101 | 4,194 | -54% | 1 | 1 | 0% | 1,473 | 2,547 | +73% | 0 | 0 | — |
case-22 | fail→pass | 18,549 | 2,972 | -84% | 1 | 1 | 0% | 1,775 | 2,386 | +34% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +73 percentage points is the difference between those two pass rates over the 21 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.