Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
78 lines (58 loc) · 2.33 KB

ARCHITECTURE.md

File metadata and controls

78 lines (58 loc) · 2.33 KB
Raw

Agora Conversational AI Demo — Architecture

This quickstart keeps the web UI and backend responsibilities separate. The Next.js app owns the browser-facing /api/* URLs, and next.config.ts rewrites them to the Python FastAPI service that owns token generation and agent lifecycle.

Python-Backed Request Flow

Browser
  ↓
Next.js app
  ↓
/api/* rewrites through AGENT_BACKEND_URL
  ↓
FastAPI service
  ↓
Agora Cloud Services
  • web owns the browser UI and the /api/* entrypoints
  • server owns the actual token generation and agent start/stop logic
  • this is the mode used by bun run dev

Shared Conversation Flow

1. Connection

Frontend: GET /api/get_config
  → Generate Token007 config for a user UID, agent UID, and channel
  → Frontend joins RTC and logs into RTM

2. Agent Start

Frontend: POST /api/startAgent { channelName, rtcUid, userUid }
  → Build agent session
  → Scope remote_uids to the requesting user
  → Start session and return agent_id

3. Conversation

User audio → RTC
  → Managed ASR, LLM, and TTS pipeline
  → Agent audio + RTM transcript events
  → UIKit transcript and visualizer in the web app

4. Agent Stop

Frontend: POST /api/stopAgent { agentId }
  → Stop session directly or through stateless fallback
  → Client cleans up RTC and RTM state

API Endpoints

Endpoint Method Description
/get_config GET Generate connection config (Token007, channel, UIDs)
/startAgent POST Start the agent session
/stopAgent POST Stop the agent by agent_id

Frontend calls these as /api/*. Next rewrites those calls to AGENT_BACKEND_URL; the Next app does not run token or AgentKit logic in-process.

Authentication

Token007 (AccessToken2) — generated from AGORA_APP_ID + AGORA_APP_CERTIFICATE only. No API_KEY/API_SECRET needed. The SDK handles token generation and API auth internally.

Detailed Documentation