Available Features

Day-to-day use: asking questions, optional cluster introspection, and feedback/transcripts.

Asking questions

Open the Lightspeed widget (bottom-right corner of the OpenShift console) once OpenStackLightspeed is Ready:

  • “How can I spin up a VM using the OpenStack CLI?”

  • “Why would a Nova compute service show as down?”

Answers are grounded via RAG, with references you can verify. By default, grounding comes from the Offline knowledge portal (always deployed, no credentials needed to browse — see Configuration for the free vs. keyed tiers). The bundled community documentation is also available, but only if you set dev.okpRagOnly: false.

Multi-model request routing

When you configure multiple entries in spec.models[], each entry defines:

  • a model alias (spec.models[].name)

  • a generated OGX provider ID: provider-<alias>

  • the upstream provider model name (spec.models[].modelName)

In request payloads for /query and /streaming_query, use these fields to select a non-default model:

  • provider: the generated provider ID (provider-<alias>)

  • model: the model alias (<alias>)

Example (explicit model selection):

{
  "query": "How can I check Nova services?",
  "provider": "provider-another-model",
  "model": "another-model"
}

If provider and model are omitted, Lightspeed uses spec.defaultModel (and its corresponding provider provider-<defaultModel>).

Cluster introspection (optional)

Enabling the rhoso_mcps dev flag (Configuration) gives the assistant read-only tools to inspect your actual OpenStack/OpenShift resources instead of relying on docs alone.

  • Strictly read-only by default — only list/get/describe-style openstack and oc commands are exposed as tools; nothing that creates, updates, or deletes resources is available to the assistant out of the box.

  • Introspection stays local to your cluster; only the query and retrieved context go to your LLM provider.

  • Credentials are automatic — the operator provisions a scoped Keystone Application Credential when an OpenStackControlPlane is detected.

Disabled by default; still evolving.

Quota enforcement (optional)

OpenStack Lightspeed can enforce token quotas per user and across the whole cluster using lightspeed-stack’s built-in quota system. The operator manages the quota storage automatically, so no additional setup is needed.

Quota enforcement is opt-in: it is disabled until you configure at least one limiter. You can combine per-user and cluster-wide limiters; requests must satisfy each configured limiter. See Quota enforcement for the configuration and an example.

Feedback and transcripts

  • dataverseExporter.feedback.enabled (default true) — thumbs-up/down on responses.

  • dataverseExporter.transcripts.enabled (default false) — full conversation transcripts.

Both configured on the CR (Configuration). Used to improve answer quality — disable either if that doesn’t fit your data policy.