Requirements

CLI Tools

Install these on the machine you will run Helm from:

Tool	Minimum version	Install
`kubectl`	1.28+	kubernetes.io
`helm`	3.12+	helm.sh

Cluster Requirements

Kubernetes

Property	Requirement
Kubernetes version	1.28+
CNI	Any (Calico, Cilium, Flannel, etc.)
Ingress controller	`ingress-nginx` (IngressClass `nginx`)
Storage	A default StorageClass with `ReadWriteOnce` support

OpenShift

Property	Requirement
OpenShift version	4.12+
Ingress	OpenShift Router (built-in)
Storage	Default StorageClass with `ReadWriteOnce` support
SCC	`anyuid` SCC for pods that need it (PostgreSQL, MinIO)

Node Sizing

Without GPU inference (vLLM disabled)

Component	CPU request	Memory request	Storage
backend	250m	512Mi	—
frontend	100m	256Mi	—
dashboard-connect	100m	256Mi	—
postgresql	250m	512Mi	10 Gi PVC
qdrant	500m	1Gi	10 Gi PVC
minio	250m	512Mi	100 Gi PVC
otel-lgtm	500m	1Gi	25 Gi PVC total
Total	~2 vCPU	~5 Gi	~145 Gi

A single node with 4 vCPU / 8 Gi RAM and 200 Gi available disk is sufficient for a minimal deployment.

With GPU inference (vLLM enabled)

The vLLM pod must be scheduled on a GPU node. The recommended model (Qwen3.5-9B-AWQ) requires:

Resource	Minimum	Recommended
GPU	1× NVIDIA GPU with 16 Gi VRAM	1× A10G 24 Gi (e.g. `g5.2xlarge` or on-prem equivalent)
CPU	4 vCPU	8 vCPU
RAM	20 Gi	28 Gi
Disk (model weights)	30 Gi	80 Gi PVC
NVIDIA driver	525+	535+
CUDA	11.8+	12.x

The GPU node must run the NVIDIA device plugin DaemonSet so nvidia.com/gpu is visible as a schedulable resource. See GPU Setup.

Image Registry Access

All Cobi application images (hellocobi/*) are hosted on Docker Hub as private images. You need:

Docker Hub credentials with pull access to the hellocobi organization.
A Kubernetes Secret of type kubernetes.io/dockerconfigjson in the target namespace.

kubectl create secret docker-registry dockerhub-secret \
  --docker-server=https://index.docker.io/v1/ \
  --docker-username=<username> \
  --docker-password=<password-or-token> \
  --docker-email=<email> \
  --namespace <your-namespace>

Reference it in your values file:

global:
  imagePullSecrets:
    - dockerhub-secret

Hugging Face Token

vLLM downloads model weights from huggingface.co at startup. You need a Hugging Face account and an access token with read access to the model repository:

Create a token with read scope.
Pass it via vllmstack.servingEngineSpec.modelSpec[0].hf_token in your values file.
For air-gapped clusters, pre-download the model weights and serve them from a local cache volume.

Persistent Storage

All stateful components use ReadWriteOnce PersistentVolumeClaims. The default StorageClass is used unless you specify storageClass in each component’s values. For on-premises clusters without a cloud storage provisioner, common options are:

Provisioner	Notes
`rancher.io/local-path`	Single-node dev/staging; data is local to the node
`nfs.csi.k8s.io`	Multi-node HA; requires an NFS server
OpenEBS	Block storage for bare-metal clusters
Longhorn	Distributed block storage for bare-metal clusters

Set the storage class globally per component:

postgresql:
  primary:
    persistence:
      storageClass: "local-path"

qdrant:
  persistence:
    storageClass: "local-path"

minio:
  persistence:
    storageClass: "local-path"

Getting Started

Self-Hosted Deployment

Data Sources

Channels

API

CLI Tools

Cluster Requirements

Kubernetes

OpenShift

Node Sizing

Without GPU inference (vLLM disabled)

With GPU inference (vLLM enabled)

Image Registry Access

Hugging Face Token

Persistent Storage

​CLI Tools

​Cluster Requirements

​Kubernetes

​OpenShift

​Node Sizing

​Without GPU inference (vLLM disabled)

​With GPU inference (vLLM enabled)

​Image Registry Access

​Hugging Face Token

​Persistent Storage

CLI Tools

Cluster Requirements

Kubernetes

OpenShift

Node Sizing

Without GPU inference (vLLM disabled)

With GPU inference (vLLM enabled)

Image Registry Access

Hugging Face Token

Persistent Storage