Lesson 7.1: Project Overview and Requirements

System Requirements

Functional:

• Support 10k concurrent users
• Sub-100ms query latency for features
• Store 1M+ documents for RAG
• Log all model predictions
• Track user conversations
• Real-time analytics dashboard

Non-Functional:

• ACID transactions for critical data
• Horizontal read scaling (replicas)
• Partitioning for time-series data
• Proper indexing (all queries under 100ms)
• Data retention policies

Architecture Diagram

User Query
    ↓
[API Layer]
    ↓
┌─────────────────────────────────────┐
│  PostgreSQL (Primary)               │
│                                     │
│  ┌──────────────────────────────┐  │
│  │ RAG System                   │  │
│  │ - documents                  │  │
│  │ - chunks (with embeddings)   │  │
│  │ - permissions                │  │
│  └──────────────────────────────┘  │
│                                     │
│  ┌──────────────────────────────┐  │
│  │ Feature Store                │  │
│  │ - user_features (cached)     │  │
│  └──────────────────────────────┘  │
│                                     │
│  ┌──────────────────────────────┐  │
│  │ Conversation Memory          │  │
│  │ - conversations              │  │
│  │ - messages                   │  │
│  │ - agent_entities             │  │
│  └──────────────────────────────┘  │
│                                     │
│  ┌──────────────────────────────┐  │
│  │ Model Logs (Partitioned)     │  │
│  │ - predictions_2024_01        │  │
│  │ - predictions_2024_02        │  │
│  └──────────────────────────────┘  │
└─────────────────────────────────────┘
           │
           ↓
    [Read Replicas]
    - Analytics
    - Dashboard

Key Takeaways

• Capstone project builds a complete AI-powered customer support system
• Four main components: RAG system, feature store, conversation memory, model logs
• Functional requirements focus on scale (10k users, 1M+ documents)
• Performance requirements demand sub-100ms latency for features
• Architecture uses PostgreSQL primary + read replicas for horizontal scaling
• Time-series data requires partitioning for model predictions
• Production readiness includes ACID transactions, proper indexing, and retention policies