Interview questions
The agent asks these one at a time, in your language, and skips every question that your files or the conversation already answered. The interview explains the rules.
| # | ID | Ask | Store |
|---|---|---|---|
| 1 | org.users_total |
How many people will use the platform? | whole number, 1 or more |
| 2 | org.departments |
How many departments or teams need their own budget or data boundary? | whole number, 1 or more |
| 3 | org.isolation |
May departments see each other’s prompts and data? | none (they may), per_department (they may not), multi_level (they may not, and data also has classification levels) |
| 4 | org.audit_required |
Must every request be attributable to a person and retained? | true or false |
| 5 | org.downtime_tolerance |
How much downtime is acceptable? | accepted, planned_only, none |
| 6 | org.ops_capacity |
Who will operate the system? | part_time_one, small_team, platform_team |
| 7 | org.kubernetes_experience |
Does the operating team run Kubernetes today? | true or false |
| 8 | workload.usage |
What will people use it for, in rough shares? | map of chat, code_assistant, rag, batch to shares that sum to 1 |
| 9 | workload.models |
Which model or models will be served? | list, see below |
| 10 | workload.context_tokens |
How long is a typical request of each group, prompt and history together, in tokens? Ask it when the groups differ. Optional. | map of chat, code_assistant, rag to a whole number of tokens, see below |
| 11 | hardware.gpus |
Which GPUs exist or will be bought? | list, see below; [] when there is no GPU |
| 12 | hardware.ram_gb |
How much system RAM does the server have? Ask only when there is no GPU. | number |
| 13 | hardware.power_budget_kw |
How much power is available for this system? Optional. | number |
| 14 | existing.identity |
Is there an identity provider already? | none, ad_ldap, oidc |
| 15 | existing.api_gateway |
Is there an API gateway already? | none, kong, other |
| 16 | existing.kubernetes |
Is there a cluster platform already? | none, kubernetes, openshift |
| 17 | existing.registry |
Is there an artifact registry already? Optional. | none, harbor, nexus, artifactory, other |
| 18 | transfer.channel |
How do files enter the network? | removable_media, data_diode, staging_host, other |
| 19 | transfer.cadence |
How often can updates be carried in? Optional. | ad_hoc, monthly, quarterly |
Usage types: chat is conversation in a chat window. code_assistant covers agentic coding tools, which keep a request in flight most of the time; people who only use inline completion in an editor are counted under chat. That is a deliberate overestimate, not a sourced placement: the notes of sizing.code_assistant.duty_low and sizing.code_assistant.duty_high in catalog/rules.toml show that inline completion keeps a request in flight only a few percent of the time, far below code_assistant and far below the duty of chat. rag is question answering over documents. batch is unattended jobs.
The formats of workload.models, workload.context_tokens and hardware.gpus are in skills/airgap-architect/references/interview.md.