Helm deployment#

This guide covers deploying VSS developer profiles on Kubernetes using Helm. It is an alternative to the Docker Compose method used in Quickstart and the Agent Workflows guides.

Install steps, values files, GPU sizing, ingress setup, and troubleshooting for each profile are documented in the chart README.md under deploy/helm/developer-profiles/ in the video-search-and-summarization GitHub repository.

Available developer profiles#

  • base

  • alerts

  • search

  • lvs

Prerequisites#

  • A Kubernetes cluster (version 1.34 or greater) you can reach with kubectl

  • Helm 3.x

  • NVIDIA GPU Operator on the cluster

  • NVIDIA NIM Operator when deploying in-cluster LLM and VLM models

  • StorageClass - for PVC provisioning

  • NGC API key — for image pull and NIM model access

GPU requirements depend on the profile and values you choose (for example, in-cluster NIMs vs remote LLM/VLM endpoints). See the GPU requirements section in the chart README for the profile you are deploying.

Steps to deploy a profile#

  1. Clone the blueprint repository:

    git clone https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization.git
cd video-search-and-summarization
git checkout tags/dev-26.06.1-3
git lfs install
git lfs pull

  2. Change into the Helm developer-profiles directory:

    cd ./deploy/helm/developer-profiles

  3. Open the chart README.md for the profile you are deploying and follow the installation and uninstall instructions there:

    Workflow

    Developer Profile

    Chart documentation

    Q&A and Report Generation

    base

    dev-profile-base

    Alerts

    alerts

    dev-profile-alerts

    Search

    search

    dev-profile-search

    Video Summarization

    lvs

    dev-profile-lvs

Note

After helm upgrade --install, wait until all pods in the release namespace are in Running state and show Ready (for example, kubectl get pods -n <NAMESPACE>) before opening the UI or sending requests. Image pulls, persistent volume binding, and in-cluster NIM model download and warm-up add startup time. For some profiles, bringing every pod to Ready can take 20–25 minutes depending on cluster resources and network speed.