Quickstart

This guide provides a comprehensive step-by-step walkthrough to help you quickly set up and start using the Warehouse Blueprint.

The Warehouse Blueprint is a video analytics solution that supports real-time object detection, tracking, and analytics. It offers two message broker options to accommodate different deployment scenarios:

• Kafka: High-throughput message broker optimized for datacenter deployments with robust persistence and scalability.

• Redis Streams: Lightweight message broker ideal for edge deployments with minimal memory footprint and low-latency requirements.

Choose the message broker based on your deployment environment: Kafka for centralized datacenter installations, or Redis for distributed edge locations where resources are constrained.

Components Overview

The high-level diagram illustrates the components of the Warehouse Blueprint. The diagram features the 2D, 3D, or MV3DT blueprint. The Agent communicates with the Blueprint’s data services through APIs.

The following diagram shows the major components of the Warehouse Blueprint:

Camera streams are ingested by VST/VMS, processed by RT-CV (perception) and Behavior Analytics microservices, with CV metadata and alerts published to the Message Bus and persisted via Storage Services.

The following diagram shows how AI Agents interact with the Warehouse Blueprint:

The Agent orchestrates MCP servers for Video IO, Video Analytics, and Report Generation, leveraging Nemotron (LLM) and Cosmos-Reason (RTVI-VLM) inference microservices.

Prerequisites

Before you begin, ensure all of the prerequisites are met. See Prerequisites for more details.

Deployment Options

The Warehouse Blueprint offers flexible deployment options with the following profiles:

1. 2D Vision AI Profile - 2D detection and tracking (bp_wh_kafka, bp_wh_redis, bp_wh_auto_calib)

2. 2D Vision AI with Agents Profile - 2D detection and tracking with VSS agent integration (bp_wh)

3. 3D Vision AI Profile - 3D multi-camera detection, tracking, and analytics using Sparse4D (bp_wh_kafka, bp_wh_redis, bp_wh_auto_calib)

4. MV3DT Vision AI Profile - 3D multi-camera detection, tracking, and analytics using RT-DETR + MV3DT (bp_wh_kafka, bp_wh_redis, bp_wh_auto_calib)

Guidance Table for Choosing a Profile

Profile Type	Features	Constraints	Supported GPUs
2D Vision AI Profile	2D single-camera detection, tracking, and analytics	Model trained on real-world data	RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S, IGX-THOR, DGX-SPARK
2D Vision AI with Agents Profile	2D single-camera detection, tracking, analytics, and VSS agent integration	Model trained on real and synthetic data	RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S, IGX-THOR
3D Vision AI Profile	3D multi-camera detection, tracking, and analytics	Model trained on synthetic data, recommended for simulated environments	RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S, IGX-THOR, DGX-SPARK
MV3DT Vision AI Profile	3D multi-camera detection, tracking, and analytics using RT-DETR + MV3DT	RT-DETR model trained on real-world data, MV3DT distributed 3D tracking via camera projection matrices	RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S, IGX-THOR, DGX-SPARK

Supported GPU Hardware

This section provides an overview of the hardware profiles supporting the various deployment options described in the previous section. Each profile is designed to meet specific use case performance and resource requirements, ensuring optimal operation of the blueprint.

See supported deployment options per GPU type for 2D, 3D, and MV3DT Warehouse Blueprint further in this section.

Warehouse blueprints are optimized to work on a single supported GPU serving 4 real-time camera streams. When you choose to install the Agents, you have the following options:

• Download and locally run the required NVIDIA NIM services (requires additional GPU resources)

• Use a remote NVIDIA NIM endpoint or any OpenAI API compatible endpoint (no additional GPU required)

NVIDIA Certified GPU Servers and Workstations

2D Warehouse Blueprint Supported Deployment Options

GPU Type	Number of streams	2D Vision AI Profile (Number of GPUs)	2D Vision AI with Agents Profile (Number of GPUs)
RTX PRO 6000 BW	16	1	2 (remote LLM + local RTVI-VLM) 3 (local LLM + local RTVI-VLM (dedicated GPU each) (default))
H100 (NVL, SXM HBM3)	26	1	2 (remote LLM + local RTVI-VLM) 3 (local LLM + local RTVI-VLM (dedicated GPU each) (default))
L40S	12	1	2 (remote LLM + local RTVI-VLM) 3 (local LLM + local RTVI-VLM (dedicated GPU each) (default))
IGX-THOR	7	1	2 (remote LLM + local RTVI-VLM) Local RTVI-VLM is supported only on IGX-THOR systems with an iGPU and dGPU.
DGX-SPARK	7	1	Not supported

Note

Recommended GPUs:

• 2D Vision AI Profile: RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S, IGX-THOR, DGX-SPARK

• 2D Vision AI with Agents Profile: RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S

When you choose to install the agents, you have the following options:

• Download and locally run the required NVIDIA NIM services (requires additional GPU resources).

3D Vision AI Profile Supported Deployment Options

GPU Type	Number of Streams	FPS	3D Vision AI Profile (Number of GPUs)
RTX PRO 6000 Blackwell	19	30	1
H100 (NVL, SXM HBM3)	16	30	1
L40S	6	30	1
IGX THOR	7	15	1
DGX SPARK	7	15	1

Note

We recommend RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S, IGX-THOR, DGX-SPARK GPUs for 3D Vision AI Profile.

MV3DT Vision AI Profile Supported Deployment Options

GPU Type	Number of Streams	FPS	MV3DT Vision AI Profile (Number of GPUs)
RTX PRO 6000 Blackwell	18	30	1
H100 (NVL, SXM HBM3)	13	30	1
L40S	7	30	1
IGX THOR	4	30	1
DGX SPARK	4	30	1

Note

We recommend RTX PRO 6000 BW, H100 NVL, H100 SXM HBM3, L40S, IGX-THOR, DGX-SPARK GPUs for MV3DT Vision AI Profile.

Note

To use different hardware than the ones listed above, you can use the VSS Configurator. Add a new hardware profile in blueprint_config.yml with stream limits (and optional config overrides), then set HARDWARE_PROFILE to that profile name when deploying. For a step-by-step guide and examples, see Adding a New Hardware Profile in the VSS Configurator documentation.

Download Warehouse Artifacts

This section guides you through the steps to download the Warehouse Blueprint artifacts.

Setup NGC Access

# Setup NGC access
export NGC_CLI_API_KEY=<NGC_CLI_API_KEY>
export NGC_CLI_ORG='nvidia'

Download the Deployment Package

Perform this on the machine where you intend to deploy the blueprint.

Note

Git LFS required: The repository uses Git LFS for large files. Install it before cloning or pulling. For example, on Ubuntu/Debian: sudo apt-get install git-lfs. On other systems, see Git LFS installation.

git clone https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization.git
cd video-search-and-summarization
git lfs install
git lfs pull
cd deploy/docker
# This is the path to the deployments directory. It is set in the industry-profiles/warehouse-operations/.env file for env VSS_APPS_DIR.
#VSS_APPS_DIR="/path/to/deploy/docker"

Warehouse App Data

Warehouse App Data contains sample video datasets, pre-trained TensorRT models (RT-DETR for 2D and MV3DT, Sparse4D for 3D, NvDCF tracker, BodyPose3DNet for MV3DT), camera calibration data, and configuration templates for deploying 2D, 3D, and MV3DT warehouse profiles.

The package includes the following MP4 datasets at 1920x1080 and 30 FPS:

Dataset	Applies To	Cameras	Duration	Scene Details
nv-warehouse-4cams	2D with Agents, and Auto-Calibration profile	4	Approximately 10 minutes per camera (600 seconds)	Real-world high-angle aisle footage with racks, palletized inventory, workers in safety vests, boxes, shrink-wrapped pallets, forklift activity, and pallet movement.
warehouse-loading-dock-3cams-synthetic	2D Kafka, Redis, and Auto-Calibration profiles	3	4 minutes per camera (240 seconds)	Synthetic loading dock footage with overhead views, loading bays, yellow-and-black floor safety zones, boxes on pallets, and forklift equipment.
warehouse-4cams-20mx20m-synthetic	3D and MV3DT Kafka, Redis, and Auto-Calibration profiles	4	5 minutes per camera (300 seconds)	Synthetic 20mx20m 3D warehouse scene with overlapping corner-mounted camera views, numbered zones, palletized boxes, forklifts, workers, safety boundaries, and barriers. Used for multi-camera tracking, spatial analytics, and cross-camera object re-identification.
warehouse-loading-dock-3cams-synthetic-degraded	2D Kafka, Redis, and Auto-Calibration profiles	3	4 minutes per camera (240 seconds)	Degraded synthetic loading dock footage with low-light and lower-quality camera visuals, representing deployments where camera feeds vary in quality. Used for testing detection, tracking, and analytics under degraded visibility.

Note

After extracting the package, set VSS_DATA_DIR in industry-profiles/warehouse-operations/.env to point to the extracted data directory.

ngc \
   registry \
   resource \
   download-version \
   nvidia/vss-warehouse/vss-warehouse-app-data:3.2.0

# OR Manually download the tar file from NGC
# URL https://catalog.ngc.nvidia.com/orgs/nvidia/teams/vss-warehouse/resources/vss-warehouse-app-data?version=3.2.0

# Extract the package

cd vss-warehouse-app-data_v3.2.0
tar -xvf vss-warehouse-app-data.tar.gz

# Set permissions

sudo chmod -R 777 /path/to/vss-warehouse-app-data

# This is the path to the data directory. It is set in the industry-profiles/warehouse-operations/.env file for VSS_DATA_DIR.
#VSS_DATA_DIR="/path/to/vss-warehouse-app-data"

Deploy Warehouse Blueprint

Configure Environment settings

This section explains the most commonly edited variables in industry-profiles/warehouse-operations/.env.

Deployment Selection

Select the appropriate configuration based on your deployment profile:

Profile	MODE	BP_PROFILE	SAMPLE_VIDEO_DATASET	NUM_STREAMS
2D Vision AI with Agents Profile	2d	bp_wh or bp_wh_auto_calib	nv-warehouse-4cams	4
2D Vision AI Profile	2d	bp_wh_kafka or bp_wh_redis or bp_wh_auto_calib	warehouse-loading-dock-3cams-synthetic	3
3D Vision AI Profile	3d	bp_wh_kafka or bp_wh_redis or bp_wh_auto_calib	warehouse-4cams-20mx20m-synthetic	4
MV3DT Vision AI Profile	mv3dt	bp_wh_kafka or bp_wh_redis or bp_wh_auto_calib	warehouse-4cams-20mx20m-synthetic	4

Hardware Profile Configuration

Determine your HARDWARE_PROFILE by running nvidia-smi -L to identify your GPU model.

GPU	HARDWARE_PROFILE Value
RTX PRO 6000 BW	RTXPRO6000BW
H100 (NVL, SXM HBM3)	H100
L40S	L40S
IGX-THOR	IGX-THOR
DGX-SPARK	DGX-SPARK

Note

To update the existing configs of a hardware profile or to add a GPU which is not present in the above list, see Adding a New Hardware Profile in the VSS Configurator documentation.

Deployment Paths and Services

Variable	Description
VSS_APPS_DIR	(Required) Path to your deploy/docker directory
VSS_DATA_DIR	(Required) Path to the extracted warehouse app data directory
HOST_IP	(Required) Host IP address of the machine where the blueprint is deployed. For hosts that cannot be accessed from the browser via the HOST_IP, such as cloud VMs, set HOST_IP to the private IP address of the instance and set EXTERNAL_IP to the public IP address of the instance.
EXTERNAL_IP	(Required only if HOST_IP is not accessible from the browser) External IP address of the machine where the blueprint is deployed
STREAM_TYPE	kafka or redis. Default: kafka
PERCEPTION_TAG	Tag for RTVI CV container image. Use 3.2.0-<version> for x86_64/aarch64 (IGX-THOR), or 3.2.0-sbsa-<version> for aarch64 (DGX-SPARK).
BEV_FUSION_MV3DT_TAG	Tag for the BEV Fusion MV3DT container image (mv3dt mode). Use 3.2.0-<version>.
RTVI_VLM_IMAGE_TAG	Tag for the RTVI-VLM container image (VLM inference for the bp_wh 2D agent profile). Use 3.2.0-<version> for x86_64/aarch64 (IGX-THOR), or 3.2.0-<version>-sbsa for aarch64 (DGX-SPARK).
VST_STREAM_PROCESSOR_IMAGE_TAG VST_SENSOR_IMAGE_TAG NVSTREAMER_IMAGE_TAG VST_INGRESS_IMAGE_TAG	Tags for the VIOS / VST (Video IO and Storage) container images — stream processor, sensor manager, NvStreamer, and ingress. Common across all platforms.
NUM_STREAMS	Number of concurrent streams to process. Default: 4

Note

Switch between Kafka and Redis

With ENABLE_PROFILE_CONFIGURATOR=true, the VSS Configurator uses STREAM_TYPE to toggle Kafka or Redis settings for these services:

• DeepStream perception: updates broker connection string, broker config file, and protocol library.

• Behavior Analytics: updates source and sink type to kafka or redisStream.

• VST: updates notification message broker consumer settings.

• Video Analytics API: sets kafka.brokers to ["localhost:9092"] for Kafka or null for Redis.

For the common configuration pattern and examples of file updates automated by the VSS Configurator, see VSS Configurator commons and profiles.

Minimal Profile

The minimal profile applies to 2D Vision AI Profile, 3D Vision AI Profile, and MV3DT Vision AI Profile only. It is recommended for Thor deployment. The default is the minimal profile.

Set MINIMAL_PROFILE="true" in industry-profiles/warehouse-operations/.env for a minimal deployment. This excludes ELK (Elasticsearch, Logstash, Kibana), Video Analytics API, monitoring, and their dependent services.

For a full extended deployment (ELK, Video Analytics API, and monitoring included), set MINIMAL_PROFILE="" in industry-profiles/warehouse-operations/.env.

Note

In the VST UI, metadata for live streams is read from Kafka or Redis, while metadata for recorded streams is read from Elasticsearch. In the minimal profile, bounding boxes (bboxes) do not appear for recorded streams in the VST UI because Elasticsearch is not deployed.

Elasticsearch ILM policy retention period

Note

Elasticsearch data is cleaned up every 4 hours as part of the ILM policy provided as part of the blueprint.

To increase the retention period, adjust index lifecycle (ILM) time by setting ELASTICSEARCH_ILM_MIN_AGE in industry-profiles/warehouse-operations/.env (for example 12h or 1d). This value is passed to the Elasticsearch init container when ILM policies are applied.

Default: 4h

ELASTICSEARCH_ILM_MIN_AGE=4h

For how ILM policies are defined and other tuning options, see ELK (Elasticsearch, Logstash, Kibana).

NIM Deployment Modes

The following modes apply to the LLM_MODE environment variable.

NIM	Required Settings	Description
local	MODE=2d BP_PROFILE=bp_wh NGC_CLI_API_KEY	Download and run NIMs locally (requires additional GPU resources), Run LLM on a dedicated GPU device.
remote	MODE=2d BP_PROFILE=bp_wh LLM_BASE_URL NVIDIA_API_KEY	Use remote NIM endpoints (no additional GPU resources required). Obtain a valid NVIDIA_API_KEY via https://build.nvidia.com/explore/discover or via the NVIDIA Developer Program website.
none	MODE=2d/3d/mv3dt BP_PROFILE=bp_wh_kafka/bp_wh_redis/bp_wh_auto_calib	Deploy 2D, 3D, or MV3DT Vision AI profiles without agents. NIM is not required for bp_wh_kafka, bp_wh_redis, or bp_wh_auto_calib profiles.

Note

The Warehouse Blueprint uses rtvi-vlm for VLM inference. Configure the local rtvi-vlm service with RTVI_VLM_BASE_URL, RTVI_VLM_ENDPOINT, RTVI_VLM_MODEL_PATH, and RTVI_VLM_MODEL_TO_USE in industry-profiles/warehouse-operations/.env. RTVI-VLM is not supported on DGX-SPARK. For details on the RTVI-VLM microservice (configuration, supported models, and APIs), see Real-Time VLM.

Model Configuration

Variable	Description
LLM_NAME	Model selection for the agent when NIM is enabled

Note

For the full set of variables (ports, agent settings, VST adaptor, calibration, etc.), use the inline comments in the shipped industry-profiles/warehouse-operations/.env.

For DGX-SPARK (SBSA), RTVI CV(Perception) and VST separate container tags to be used ( Refer warehouse/.env for the latest tags commented out in the file, use the uncommented tags for the deployment and make sure to comment out the default multi-arch tags).

Deploy the Blueprint

Warning

Ensure industry-profiles/warehouse-operations/.env is configured for your deployment before proceeding. Review the Deployment Selection and environment settings sections above.

Deploy Warehouse Blueprint

source /path/to/deploy/docker/industry-profiles/warehouse-operations/.env
cd /path/to/deploy/docker

# Docker login to the NGC docker registry
docker login \
    --username '$oauthtoken' \
    --password "${NGC_CLI_API_KEY}" \
    nvcr.io

# Start the blueprint
docker compose \
    --env-file industry-profiles/warehouse-operations/.env \
    up \
    --detach \
    --pull always \
    --force-recreate \
    --build

Note

Initialization of some components might take a while, especially the first time as large containers will be pulled.

Verify Deployment

1. Verify if containers are in running state:

Verify deployment

docker ps
docker compose ls

2. Check to make sure streams were properly added to VST. To do so, navigate to the VST UI (see endpoint below) and check the Dashboard to confirm your streams are in a healthy state. If you do not see them there, check NVStreamer or your source to make sure they are active.

3. Check vss-rtvi-cv FPS to make sure DeepStream is running properly. View the vss-rtvi-cv logs by running the below command and looking for FPS lines in the logs. Ensure it is running at the desired FPS. If this is lower than expected, make sure your GPU is not oversaturated.

# For 2D/3D Vision AI Profile
docker logs -f vss-rtvi-cv

# For MV3DT Vision AI Profile
docker logs -f vss-rtvi-cv-mv3dt

4. Check Kibana for indexed analytics data and alerts. Go to the endpoint listed below and navigate to the dashboard.

5. Check the VSS UI (see endpoint below) and test a few prompts after the system is up for a few minutes and a few alerts are present.

For detailed testing and validation steps, refer to:

• 2D Vision AI Profile

• 2D Vision AI with Agents Profile

• 3D Vision AI Profile

• MV3DT Vision AI Profile

• Troubleshooting Guide

Service Access Points

Once deployed, the following service access points are available:

UI Service URLs

Service	URL
VSS-UI (2D Vision AI with Agents Profile)	http://<HOST_IP>:7777
Kibana-UI(2D/3D/MV3DT)	http://<HOST_IP>:7777/kibana
VST-UI(2D/3D/MV3DT)	http://<HOST_IP>:30888/vst
NvStreamer-UI(2D/3D/MV3DT)	http://<HOST_IP>:31000
Auto-Calibration UI (2D/3D/MV3DT auto-calibration profiles only)	http://<HOST_IP>:5000
Phoenix-UI (Telemetry)(2D Vision AI with Agents Profile)	http://<HOST_IP>:7777/phoenix

Note

• The Phoenix UI is only available when agent telemetry is enabled. For how to enable telemetry and inspect traces, see VSS Agents Observability.

• Observability stack access details (Grafana, Prometheus, and exporters) are documented in Observability.

API Service URLs

Service	Access Point
Video-Analytics-API	http://<HOST_IP>:8081
VST-MCP	http://<HOST_IP>:8001
VA-MCP	http://<HOST_IP>:5005
LLM-NIM	http://<HOST_IP>:30081

Teardown the Deployment

To stop and remove the Warehouse Blueprint deployment:

# Stop the running deployment

docker compose --env-file industry-profiles/warehouse-operations/.env down

# Alternatively to remove all the containers, images and volume

docker compose --env-file industry-profiles/warehouse-operations/.env down -v --rmi all

# Tear down all dangling volumes

docker volume ls -q -f "dangling=true" | xargs docker volume rm

# Cleanup all data (from deployments directory)

bash scripts/cleanup_all_datalog.sh -e industry-profiles/warehouse-operations/.env

Customization

The Blueprint supports several levels of customization, including but not limited to adding new cameras and updating models. For details, refer to the following sections under the Blueprint deep dive pages:

• 2D Vision AI Profile Customization

• 2D Vision AI with Agents Profile Customization

• 3D Vision AI Profile Customization

• MV3DT Vision AI Profile Customization

Warehouse Blueprint Skill

Agent Skills are reusable capabilities that coding agents such as Claude Code or Codex can load to deploy and operate the VSS Blueprint from natural-language requests. For the Warehouse Blueprint, the primary deployment Skill is vss-deploy-profile. It can select a warehouse profile, validate prerequisites, prepare environment overrides, run a compose dry-run, bring up the stack, and monitor readiness without requiring the operator to memorize every deployment command.

For the complete VSS Skills catalog and installation model, see Agent Skills.

Note

The screenshots in this section show the Claude UI. Use the same prompts in Codex or another supported coding-agent host when that host has access to the cloned VSS Blueprint repository.

Installing Skills

Install the Skills the same way as any other VSS Agent Skill. See Installing Skills for the full installation flow. Follow these steps to install and deploy the Warehouse Blueprint using the Skills:

1. Clone the VSS Blueprint repository and open it in your coding agent.

   git clone https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization.git
   cd video-search-and-summarization
   

2. Ask the coding agent to install the repository Skills:

   Read skills/README.md and every SKILL.md file under skills/. For each skill in the
   catalog, install it for this host so I can invoke it from a shell or chat session.
   Use the host's standard skills directory. Symlink each skill folder rather than
   copying it so a git pull here keeps every install up to date. Skip skills that are
   already installed and pointing at this checkout. When you're done, list the skills
   you registered and which directory you used.
   

   

   Install the VSS Skills from the cloned repository.

   Note

   After installation, each Skill is invokable by name in any new coding-agent session. It also works in your current session without restarting — the agent can read the SKILL.md files directly as a reference and follow the instructions.

3. Deploy the Warehouse Blueprint from your coding agent.

   Example warehouse deployment prompts

   Deploy the VSS warehouse blueprint on this host.
   Bring up the warehouse blueprint with the bp_wh agent profile and a remote LLM.
   Deploy warehouse in 3D mode with the kafka broker, extended profile.
   Deploy warehouse mv3dt against my own RTSP streams and generate calibration first.
   

   

   Deploy warehouse in 2D mode with agent, using remote LLM.

4. Run pre-flight checks and provide deployment inputs.

   The Skill first checks the host hardware profile, GPU and driver visibility, Docker and Compose access, NVIDIA Container Toolkit, writable data paths, and Brev or edge-specific prerequisites when they apply. If a check fails, the Skill fixes and re-verifies it when possible, and stops only for requirements that cannot be satisfied automatically.

   After the pre-flight checks pass, the Skill may ask for an NGC API key and the vss-warehouse app data location before it generates the Docker Compose configuration. When the deployment uses a remote LLM, the Skill also asks you to choose the remote LLM option and to provide an API key for accessing the remote endpoint.

   

   Provide the required NGC API key and vss warehouse app data inputs for the deployment.

5. Review the dry-run output before the stack starts.

   The Skill generates the Docker Compose configuration and shows the services that will start.

   

   Review the generated configuration and dry-run output.

6. Let the Skill bring up the docker compose stack and monitor readiness.

   The Skill starts the Docker Compose stack and polls deployment status until the expected services are ready.

   

   Confirm that the warehouse blueprint deployment completed successfully.

After the stack is running, use the Warehouse Blueprint VSS UI and analytics views to verify that the deployment is serving data.

Verify the running deployment in the VSS UI.

Verify analytics output in Kibana.

How the Skill Configures and Runs the Deployment

For the Warehouse Blueprint, vss-deploy-profile loads skills/vss-deploy-profile/references/warehouse.md in the VSS Blueprint repository as its per-profile playbook. That reference owns the warehouse-specific decisions the Skill makes on the operator’s behalf:

• Profile variants - picks one of bp_wh, bp_wh_kafka, bp_wh_redis, or bp_wh_auto_calib against MODE=2d|3d|mv3dt, including the recommended SAMPLE_VIDEO_DATASET and NUM_STREAMS for each combination. See Deployment Options.

• Minimal vs extended - for bp_wh_kafka and bp_wh_redis, chooses whether to deploy ELK, Video Analytics API, and monitoring on top of the core CV pipeline. See the Minimal Profile section above for the comparison.

• GPU layout - assigns RT_CV_DEVICE_ID, RT_VLM_DEVICE_ID, and, for bp_wh, LLM_DEVICE_ID or SHARED_LLM_VLM_DEVICE_ID against the host GPUs.

• LLM placement - for the bp_wh 2D agent profile, picks LLM_MODE=local or LLM_MODE=remote from the user’s hardware and request. For bp_wh_kafka, bp_wh_redis, and bp_wh_auto_calib in 2D, 3D, or MV3DT mode, sets LLM_MODE=none.

• App data source - chooses between <repo>/data, a custom local path, or the nvidia/vss-warehouse/vss-warehouse-app-data NGC resource.

• Calibration - for own-data deployments, routes to the standalone or warehouse auto-calibration profile to generate calibration JSON before the full warehouse stack comes up.

• Bring up and monitor - runs docker compose ... up against industry-profiles/warehouse-operations/.env, then watches logs and docker ps until every expected container reports Up and FPS is flowing through vss-rtvi-cv or vss-rtvi-cv-mv3dt for MV3DT.

Debugging and Updating Skills

For warehouse-specific debugging prompts such as “warehouse FPS low”, “BEV out of sync”, or “perception not starting”, the Skill loads the companion skills/vss-deploy-profile/references/warehouse-debug.md reference.

After a git pull of the VSS Blueprint repository, symlinked Skills automatically point at the updated content. If new Skills were added, ask the coding agent:

Re-read skills/README.md and add any new skills missing from this host's skills
directory.

Related Documentation

• Agent Skills - VSS Skills catalog, installation, usage, and update flow.

For the standalone RTVI-CV-3D / MV3DT stack without the full Warehouse agent, LLM, or VLM stack, use the vss-deploy-detection-tracking-3d skill instead. That skill covers bundled sample data, custom synchronized videos, live RTSP streams, MV3DT verification, and automatic handoff to vss-generate-video-calibration when calibration is missing. See the MV3DT Agent Skills walkthrough for sample prompts and expected agent steps.

Brev Launchable

Launchables are a part of NVIDIA Brev that deliver preconfigured, fully optimized compute and software environments.

This deployment uses the Docker Compose deployment method to set up the Warehouse Blueprint.

Link to GitHub Repository used in this launchable.

Launch Blueprint

Prerequisites

Host System Prerequisites

Brev Launchable takes care of all these requirements. The provided notebook adjusts the system requirements.

• Ubuntu 22.04 or Ubuntu 24.04

• Minimum NVIDIA driver - 580

• The notebook pins Docker to a known-good combination if the host Docker stack is outside the supported version range

Obtain NGC API Key

To obtain the required NGC API Key, follow the steps in:

• NGC User Guide

• Generating NGC API Keys

Deployment Profiles

The launchable notebook uses the same MODE, BP_PROFILE, STREAM_TYPE, SAMPLE_VIDEO_DATASET, and NUM_STREAMS values described in the Deployment Selection section above. In the notebook’s first code cell, set them to match your chosen profile.

See Deployment Options for full details on each profile.

Steps to Launch the Blueprint

1. Go to the Launchable page.

   

2. Click on ‘Deploy Launchable’ on top right.

3. Click on ‘Go to Instance Page’.

4. Click on ‘Open Notebook’ button after it is enabled (this could take a couple of minutes).

   

   • Close any warnings on the Jupyter Lab tab.

5. Navigate to and open the video-search-and-summarization/deploy/docker/scripts/deploy_warehouse_launchable.ipynb notebook.

   In Warehouse Blueprint 3.2, the compose files ship in-tree in the video-search-and-summarization repo — there is no separate compose tarball to download. On Brev, the repo is cloned onto the instance automatically; set DEPLOY_SOURCE_PATH in Section 1 to that path (typically ~/video-search-and-summarization).

6. In Section 1, set NGC_CLI_API_KEY and select your deployment profile (MODE, BP_PROFILE, STREAM_TYPE, HARDWARE_PROFILE, etc.) — see the table above.

7. Run each cell in order and follow the instructions in the notebook to access the Warehouse Blueprint UI through the HAProxy ingress (port 7777).

Note

The complete setup takes around 20 minutes on first run as it downloads containers, pulls warehouse app data from NGC, and (for perception) compiles the TensorRT engine. Subsequent runs are faster — images stay cached locally and the engine is cached under $VSS_DATA_DIR/models/.

Accessing the UI

Browser-facing services are reached through their service ports:

• Agent UI → <secure_link> directly (the HAProxy ingress serves the Agent UI at the root path)

• Kibana → <secure_link>/kibana/ (requires extended profile, MINIMAL_PROFILE="")

• VST UI → <secure_link>/vst

Note

The Brev Launchable has known limitations with VST video playback, 2D with Agent profile, and auto-calibration profiles. See Brev Launchable Known Limitations.

For the full set of service access points (including direct ports for Video Analytics API, MCP servers, and NIM endpoints), see Service Access Points.