Architecture

Browser / WhatsApp / Telegram / Gmail / Twitter
        │
        ▼
  Express.js Server (port 5001)
  ├── REST API (~160+ endpoints)  ←── JWT Auth middleware
  ├── WebSocket Server            ←── Real-time inbox updates
  ├── Webhook handlers            ←── Gupshup / Wasimple / Meta / Stripe
  └── Static file server          ←── Vite-built React app
        │
        ├── MySQL 8 (Drizzle ORM)
        ├── Redis (session cache, job queue)
        ├── AWS S3 (media + training files)
        ├── Qdrant (vector store for RAG)
        └── AI Agent Process (port 5555)
              └── LangChain + Gemini / OpenAI

sahyogi/
├── client/src/           # React frontend
│   ├── pages/            # Route-level components
│   ├── components/       # Shared UI components
│   ├── hooks/            # Custom React hooks
│   └── lib/              # API clients, utilities
│
├── server/               # Express backend
│   ├── routes.ts         # All API routes (160+ endpoints)
│   ├── db.ts             # MySQL connection pool
│   ├── jwt-auth.ts       # JWT middleware
│   ├── storage.ts        # DB query/mutation layer
│   ├── stripe.ts         # Stripe client + webhooks
│   ├── websocket.ts      # WebSocket server
│   ├── cron.ts           # Scheduled jobs
│   └── utils/            # vectorSearch, encryption, events
│
├── shared/
│   ├── schema.ts         # Drizzle schema + Zod types (source of truth)
│   ├── gupshup.ts        # Gupshup API client
│   └── wasimple.ts       # Wasimple API client
│
├── agent/                # Standalone AI agent (port 5555)
└── db/                   # Migrations + seed

WhatsApp user sends message
        │
        ▼
Gupshup/Wasimple/Meta webhook → POST /api/webhooks/...
        │
        ▼
Bot lookup (channel → bot mapping)
        │
        ▼
AI Agent (LangChain + Gemini)
  ├── Tool calls (Airtable, webhooks, API tools)
  └── RAG: Qdrant vector search → context injection
        │
        ▼
Response sent back via WhatsApp provider API
        │
        ▼
Message stored in MySQL → WebSocket push to inbox

User uploads PDF / URL
        │
        ▼
File stored in AWS S3 / URL scraped
        │
        ▼
processTrainingFiles.ts / processTrainingUrls.ts
  → Text chunked → Embeddings via Gemini/OpenAI
  → Stored in Qdrant vector collection
        │
        ▼
On bot message → vectorSearch.ts queries Qdrant
  → Top-k chunks injected into LLM prompt