FAQ

Is this production-ready?

It is a reference implementation and customer template, not a managed product. The customer values still need identity, secrets, ingress, transport encryption, storage, observability, backup, model selection, capacity tests, and current release evidence. See Production readiness.

What does the project install?

The local Argo CD profile installs the platform services and several lab add-ons. The customer profile installs a smaller application set and expects the operator to provide platform operators, observability, ingress, and backup integration. Architecture lists the difference.

Does the quickstart work offline?

No. The first run downloads tools, manifests, images, charts, Python packages, and an Ollama model. Inference uses the local Ollama pod after setup. An offline deployment needs internal mirrors and preloaded artifacts.

Do I need a GPU?

Not for the local path. The local smoke test uses qwen2.5:0.5b with Ollama on CPU. The checked-in customer vLLM profiles expect GPU resources and must be resized for the target model and nodes.

Is in-cluster traffic encrypted?

Not by default. The data plane uses HTTP. NetworkPolicy restricts which pods can connect but does not encrypt packets. See Security overview.

Is the agent workspace a separate kernel sandbox?

Only when the cluster supplies an isolation runtime and sandbox.runtimeClassName selects it. Otherwise the workspace uses a restricted container/pod security boundary on the node's normal container runtime.

How is tenant identity enforced?

A sandbox-bound key record or verified JWT tenant claim can bind the gateway request to a tenant. RAG can verify its own JWT as well. Without those bindings, X-Sandbox-ID is trusted caller input and is not a safe multi-tenant boundary under a shared key.

Does the gateway implement the full OpenAI or Anthropic API?

No. It implements the routes in the checked-in OpenAPI contract. Chat completions can stream; legacy completions, Anthropic Messages, and Responses cannot in this release. See Scope and non-goals.

What do the checked-in evidence files prove?

Files named sample-* prove report shape and gate behavior. The non-strict gate may use them. They do not prove the state of the current checkout, a release, or a customer cluster. Generate fresh reports and use make release-gate-strict for a handoff.

Does the audit chain prevent log tampering?

It makes edits and reordering detectable in an exported chain. Durability and rollback detection require external log retention and a trusted chain-head anchor. Each gateway process/replica has its own chain.

What does the regulated-offline profile guarantee?

It renders a tenant namespace without external CIDR egress. It does not air-gap the whole cluster or configure private registries, model mirrors, internal identity, or cluster-wide egress policy. See the restricted-egress example.

How do I upgrade or roll back?

Change the immutable CUSTOMER_REVISION, review the rendered changes, and let Argo CD reconcile. Roll back by returning to the prior tag. Follow the upgrade runbook; stateful schema or collection changes may require a separate data rollback.

Where should I report a security issue?

Use the private process in SECURITY.md. Do not put secrets, customer data, private prompts, or exploit details in a public issue.