Installation overview
Olympusx AI installs into a namespace of a Kubernetes cluster you own. This section is for the platform team that installs it: it covers what you need, the decisions to make first, the install itself, and the checks after it.
Two install paths
Section titled “Two install paths”Both paths deploy the same platform from the same chart, and both produce the same objects with the same names. What differs is who keeps them in place.
| The operator | The Helm chart | |
|---|---|---|
| How it installs | You add the Olympusx AI operator catalog to Operator Lifecycle Manager (OLM), install the operator, and create one Olympus resource |
You run helm install with a values file |
| Who keeps it running | A controller reconciles continuously: a component somebody deletes comes back | helm upgrade applies changes when you run it |
| What you can set | A curated set of fields, each one a supported decision (Configuration) | Every chart value |
| Registers its own cluster | Yes, by default | Only when you set clusterRegistration.enabled: true |
| Upgrades | Through OLM, with automatic or manual approval | helm upgrade |
Pick one per namespace. Do not install with Helm and then create an Olympus resource over it: the two take turns rewriting the same objects. To move from Helm to the operator, run helm uninstall first (Uninstall), then create the resource.
Which path on which platform
Section titled “Which path on which platform”| Platform | Use | Why |
|---|---|---|
| OpenShift 4.20 or later | The operator | It installs from the Software Catalog, detects OpenShift, renders Routes and removes the pinned user IDs that restricted-v2 rejects. OpenShift |
| Oracle OKE | The Helm chart | Three settings an OKE install needs have no field on the Olympus resource: the user ID model servers run as, the model download proxy, and Workload Identity for OCI Generative AI. Oracle OKE |
| Other Kubernetes (on-premises, EKS, AKS, GKE) | The Helm chart | Self-hosted model servers need a user ID that only the chart can set on clusters that assign none. Kubernetes |
| OpenShift 4.14 to 4.19 | The Helm chart, with the chart’s OpenShift values | The operator needs Kubernetes 1.33, which arrives with OpenShift 4.20. OpenShift |
The operator also runs on vanilla Kubernetes 1.33 or later with OLM installed. Choose it there only when no self-hosted model will run on a cluster that assigns no user ID, for example when every model comes from a cloud provider.
A cluster that only serves models for an Olympusx AI installed elsewhere runs the operator without an Olympus resource. It uses the operator’s other resource, ModelDeployment, which keeps each model’s workload healthy from inside the cluster.
What gets installed
Section titled “What gets installed”| Component | What it does | Replicas |
|---|---|---|
| Control plane | Tenants, models, capacity, pricing, billing, audit, and the management API | 2 |
| Gateway | The OpenAI-compatible inference endpoint: authentication, entitlement, budgets, content policy, routing, metering | 2 |
| Console | The web interface. It also proxies the management API, so the browser stays on one origin | 2 |
| Metering and budget workers | Write the usage ledger and enforce budgets | 2 and 1 |
| Code parser | Splits source files into functions and types for the Code Assistant | 1 |
| Document converter | Converts PDF, DOCX, PPTX, XLSX and scanned images for knowledge pipelines | 1 |
| Object store | S3-compatible storage for platform files | 1 |
| Database, cache, identity service | Bundled by default, each replaceable by your own | 1 each |
| Guardrails service | PII detection and content safety. Off by default | 2 |
| Tracing store | Traces of agent runs and assistant answers. Off by default | 3 workloads |
| Reranker | Reorders retrieval results. Off by default | 1 |
GPU capacity, model servers and vector databases are not part of the install. You register clusters and deploy models from the console afterwards.
The install, in order
Section titled “The install, in order”- Requirements: versions, cluster, storage, network, sizing.
- Before you start: the decisions, and the ones that cannot change later.
- The install for your platform: Kubernetes, OpenShift or Oracle OKE. For a cluster with no internet access, read Air-gapped installation first.
- Configuration: every field and value you set.
- First sign-in: the first platform administrator and the setup checklist.
- Verify the installation.
Then Quickstart takes you from an empty platform to a first answer.