Skip to content

api.md

Latest commit

 

History

History
66 lines (54 loc) · 2.36 KB

api.md

File metadata and controls

66 lines (54 loc) · 2.36 KB
Raw

API Reference

POST /api/chat

Main conversation endpoint — runs the full agent pipeline (lib/orchestrator.ts).

Streaming: send Accept: text/event-stream to watch the pipeline live. The response is a Server-Sent Events stream:

data: {"type":"step","step":{"agent":"Sentiment","decision":"30/100 · negative","ms":412}}
data: {"type":"token","text":"Hi Priya, "}
data: {"type":"reset"}            ← partial text obsolete, a tool phase started
data: {"type":"token","text":"I've initiated your refund"}
data: {"type":"done","result":{ ...full JSON result below... }}

Without the header, the endpoint returns plain JSON (used by scripts and integrations).

Request:

{
  "message": "string",
  "customerId": "uuid",
  "conversationId": "uuid | null",
  "brandId": "string | null"
}

Response:

{
  "reply": "string",
  "conversationId": "uuid",
  "sentiment": { "score": 0, "label": "string", "churnRisk": false, "buyingIntent": false },
  "isEscalated": false,
  "memoriesUsed": 0,
  "memoryMode": "semantic | recency",
  "action": { "action": "string", "message": "string", "data": {} },
  "guardrails": [
    { "action": "process_refund", "allowed": true, "rule": "all-checks-passed", "reason": "string" }
  ],
  "injectionFlagged": false,
  "detectedLanguage": "english | hindi | hinglish",
  "trace": {
    "totalMs": 0,
    "steps": [ { "agent": "Sentiment", "decision": "string", "ms": 0 } ]
  }
}

guardrails lists every action the model proposed with its policy verdict; trace is the per-agent decision timeline that also persists to agent_traces.

GET /api/dashboard

All Mission Control data: stats (incl. avgPipelineMs), sentiment breakdown/trend, recent conversations, per-customer health trajectories (score, trend, series, factors), guardrails (stats + event feed + active policy), agentTraces, and the resolutionLedger.

POST /api/outreach

Triggers the proactive outreach scan (delayed + overdue orders). Protected by CRON_SECRET.

POST /api/escalation

Generates an AI escalation briefing for human handoff. Takes conversationId.

GET /api/conversation

Returns full message history for a conversation.

Demo utilities

  • POST /api/demo/simulate — seeds a delayed-order scenario for the dashboard demo
  • POST /api/demo/reset / POST /api/demo/reset-chat — restore the demo data to a clean baseline