Start a useful conversation#

  1. Select the cluster you want to discuss and open the AI chat panel. A standalone chat route also exists at /ai-chat-box.
  2. Choose an available model where the interface offers model selection. Model availability depends on administrator configuration and policy.
  3. Ask a scoped question: for example, “Inspect deployment web in namespace staging and explain why its pods are not ready.”
  4. Review the tool calls and returned resource information before following a recommendation.
  5. Continue in the same conversation when clarifying the same investigation; use the history panel to return to an earlier thread.

See the illustrated AI platform setup guide for the provider, model and workspace-policy controls used to make models available.

Tool evidence and decisions#

The assistant can use configured tools for resource reads, descriptions, logs, cluster overview and Prometheus queries. Other implemented tools can execute commands or apply, patch and delete resources when authorization and policy permit them. Read the proposed tool and arguments whenever a pending action requests your decision.

A request-choice or form interaction asks for information needed to continue. Answer with the intended cluster/resource context. A suggested action is not evidence that the action has already run; inspect the resulting tool output and live resource state.

A practical investigation sequence#

  1. Ask for a read-only summary of the affected workload.
  2. Request relevant events and logs rather than a general guess.
  3. Ask the assistant to explain the smallest proposed change and its expected effect.
  4. If you authorize a mutation, review its resource identity and manifest/arguments.
  5. Verify the result in the normal resource page and compare it with the original symptom.

Data and permission boundaries#

Messages and tool results can be sent to the configured inference provider. Use a provider and policy suitable for the information involved. Self-hosting the console does not automatically make an external AI endpoint private.

AI tool access remains constrained by the configured policy and resource authorization. A chat response cannot grant missing permissions. Treat generated explanations as recommendations to verify against the actual cluster state.