How phrugsa-limbunlom/picksmart works
Project type and purpose
• PickSmart is a full-stack, distributed AI-powered shopping assistant.
• It consists of:
• A React-based SPA (“chatbot-app”) for user interaction.
• A FastAPI backend (“chatbot-server”) implementing a Retrieval-Augmented Generation (RAG) pipeline with AI agents.
• Kafka for event streaming.
• MongoDB as a vector store.
• External AI services (Groq API, Tavily search API, LangChain, LangGraph).
File structure
Top-level directories:
• chatbot-app – React frontend (public, src, Dockerfile, package.json)
• chatbot-server – Python backend (agent, processor, service, data, utils, main.py, Dockerfile, model.yaml)
• kafka – Kafka broker config (Dockerfile, config/)
• docker-compose.yaml – brings up all containers in a network
Patterns:
• Clean separation frontend vs. backend.
• Backend is layered: service → processor → agent → data.
• Event-driven: Kafka broker between producer & consumer.
• RAG: vector store for embeddings in MongoDB.
Config & deployment:
• Dockerfiles per component.
• docker-compose orchestrates multi-container network.
Additional README insights
• Tech stack: React, FastAPI, Kafka, MongoDB, LangChain, LangGraph, Tavily, Groq.
• Agent workflow: Query Analysis → Online Shop Search → Analysis & Ranking → Product Source Search (implemented in SearchAgent.py & orchestrated via LangGraph).
Elements
• Clear labels: “React Frontend”, “FastAPI Backend”, “Kafka Broker”, “MongoDB Vector Store”, “Groq API”, “Tavily API”, “LangGraph Orchestrator”
– HTTP: Frontend → Backend → Frontend
– API calls: Backend → External AI services
– Data flow: Backend → MongoDB
– Events: Backend ↔ Kafka
– Blue rectangles: services/components
– Clouds: external APIs
– Sequence numbers: agent workflow steps
Final advice
1. High-level overview (user → frontend → backend → data stores → external services)
2. Detailed backend sub-architecture (layered services & agent state graph)
• Environment boundaries: Docker Compose network, container per service
• Keep it detailed but avoid clutter by grouping related modules.
• Iterate with dev team to validate component names and data flows before finalizing.