ai-assist /chat

Endpoint

POST http://localhost:8080/chat
Content-Type: application/json

This is the assistant's streaming chat endpoint, driven by the Vercel AI SDK streamText loop. The browser's useChat hook posts to it directly (bypassing the studio proxy), so identity, ACL and project key travel in the request body.

Request body

{
  "messages": [ /* UIMessage[] — full conversation from useChat */ ],
  "sessionId": "client-generated-uuid",   // server creates the session if new
  "provider": "openai",                    // optional: "openai" | "anthropic"
  "context": {
    "userEmail": "agent@example.com",
    "userRole": "Support Agent",
    "projectKey": "your-ct-project-key",
    "businessType": "b2c",                  // "b2c" | "b2b"
    "pageContext": null,                    // what the rep is looking at
    "proactiveHint": null,
    "workingMemoryBlock": null,
    "canViewTickets": true,
    "canCreateTickets": false
    /* …further per-capability ACL flags */
  }
}

context and ACL

context carries the session identity and the access-control flags. ACL defaults to least privilege: reads true, writes false. Capability flags such as canViewTickets / canCreateTickets gate what tools the assistant may call. When omitted, userEmail, userRole, projectKey and businessType fall back to safe defaults.

Response

A streamed response consumed by useChat — assistant tokens interleaved with tool-call results. The stream ends when the loop hits its step budget or a tool emits an approval card (so a destructive action waits for the rep's next turn).

Behavior notes

• Rate limiting is per userEmail, an in-process sliding window.
• Memory: Redis working memory is injected into the prompt; on finish, the working-memory state and the user + assistant turns are persisted to MongoDB.
• Tools: commerce tools, ticket tools and render-only UI tools, all sharing one per-request approval gate.