API reference: AI assistant

Send the conversation ({ messages: [{ role, content, images? }] }; roles user/assistant, last must be user). The server grounds the model on a live, read-only data digest of YOUR org (real spend figures and trends, inventory, the network topology graph, alerts and Terraform state) plus your org's most-similar prior Q&A and indexed cost/topology insight documents (RAG), calls the configured OpenAI-compatible backend, and returns the answer. A user turn may carry up to 4 image attachments (png/jpeg/webp/gif, ≤5MB each, base64 + mediaType) forwarded as multimodal content — transient, never stored. Audited assistant.query.

{ "messages": [ { "role": "user", "content": "Which VPCs have no NAT gateway?" } ] }

{ "text": "…", "model": "gemini-…", "usage": { "…": "…" } }

The DEPLOYER assistant. Send the conversation ({ messages: [{ role, content }] }) describing what to deploy; the server grounds the model on YOUR org's real blueprint catalog, connected cloud accounts and Terraform folders, then RESOLVES the model's answer against that same real data and returns a validated draft proposal ({ blueprintId, accountId, workspaceId|null, workspaceName, vars, missing[] }) plus the prose reply. THIS ROUTE MUTATES NOTHING — no workspace, no plan, no apply: the draft is handed to the deploy wizard, which runs the unchanged plan → policy → cost → approve → apply pipeline. A blueprint, account or folder the model names that does not exist comes back as { proposal: null, unresolved: { kind, detail } } — never a substitute. Audited assistant.deploy_proposed.

{ "messages": [ { "role": "user", "content": "a t3.small web server in eu-west-1" } ] }

{ "reply": "…", "proposal": { "blueprintId": "aws/compute-instance", "accountId": 12, "workspaceId": null, "workspaceName": "web-staging", "vars": {}, "missing": [] }, "unresolved": null }