System Design

​

Architecture Evolution

​

Node.js Backend Architecture

​

Core Application Stack

​

AI & Vector Layer

​

Security & Configuration

• API key encryption with internal crypto module
• JWT tokens for access and refresh authentication
• Environment-based configuration (.env files)
• Role-based access control

​

Real-time & Storage

• Socket.IO for live updates and streaming responses
• Redis Pub/Sub for socket scaling
• SMTP email integration
• MinIO / AWS S3 file storage with parallel processing
• Event streaming for optimized file uploads

​

Next.js Frontend Architecture

​

Framework & UI

​

Module Structure

​

AI Integration

• Socket.IO client for real-time AI responses
• Single event emission for all AI operations
• Backend-driven operation routing
• Asset serving through S3/MinIO

​

Request Processing Architecture

Client Request (Socket Event)
    ↓
[Node.js Backend] → Single entry point
    ↓
[LangGraph Router] → Intelligent operation detection
    ↓
[Operation Handler] → Execute appropriate flow:
    • Normal Chat → LLM call
    • Web Search → SearxNG + LLM
    • Image Generation → DALL·E
    • Document Chat → Pinecone + LLM
    • Agent Chat → Agent prompt + LLM
    • Agent + Document → Combined context + LLM
    ↓
[Response Stream] → Socket.IO streaming to frontend
    ↓
Response → Real-time display in UI

​

LangGraph Flow Management

​

Intelligent Backend Routing

• Normal chat queries
• Document-based questions
• Agent conversations
• Combined agent + document queries

• Web search integration
• Image generation requests
• Vision model processing

​

Operation Detection

• Tool requirements (web search, image generation)
• Document context needs
• Agent selection
• Model capabilities and limitations

​

Service Communication

• Frontend ↔ Node.js: Single Socket.IO event for all operations
• Node.js ↔ LLMs: Direct API calls through LangGraph
• Backend ↔ Database: MongoDB connections
• Task Processing: Bull + Redis queues
• Real-time Updates: Socket.IO + Redis Pub/Sub
• File Processing: Parallel S3 upload and vector embedding

​

File Upload Optimization

​

Parallel Processing Architecture

User Upload
    ↓
Event Streaming to Backend
    ↓
Parallel Processing (2 chunks at a time):
    ├─ Chunk 1 → S3 Upload
    └─ Chunk 2 → Vector Embedding (Pinecone)
    ↓
Optimized Speed & Performance

​

Web Search Architecture

​

SearxNG Integration

• Independence: No dependency on OpenAI’s search features
• Self-hosted: Complete control over search infrastructure
• Universal: Works with all models except GPT-4o latest, DeepSeek, and Qwen
• Privacy: No external search API dependencies

​

Deployment Architecture

• Development: Local Docker containers
• Production: Multi-container deployment
• Secrets: Environment variables and secret managers
• Monitoring: Socket.IO event tracking and logging