Mane - Hair Care Shop

deployment Link - https://neusearch-ei5c888y3-vandit98s-projects.vercel.app/

An e-commerce platform with an AI-powered shopping assistant that helps users find the right hair care products.

What it does

• Product Catalog: Browse 48+ hair care products scraped from Traya.health
• AI Chat Assistant: Ask questions like "I have dry scalp, what should I use?" and get personalized recommendations
• Product Details: View full product info including price, description, and features

Tech Stack

• Frontend: React + TypeScript + Vite
• Backend: FastAPI + PostgreSQL (Supabase)
• AI: Google Gemini for chat and embeddings
• Search: pgvector for semantic product search

Quick Start

1. Clone and setup backend

cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# Add your credentials
echo "DATABASE_URL=your_supabase_url" > .env
echo "GEMINI_API_KEY=your_gemini_key" >> .env

# Run
uvicorn app.main:app --reload --port 8000

2. Setup frontend

cd frontend
npm install
npm run dev

3. Load products

curl -X POST http://localhost:8000/api/scraper/run

Open http://localhost:3000

How the RAG works

1. User asks a question
2. System searches products using text matching (or vector search if embeddings are available)
3. Relevant products are sent to Gemini as context
4. Gemini generates a helpful response with product recommendations

API

Endpoint	Description
GET /api/products	List products
GET /api/products/:id	Get product
POST /api/chat	Chat with assistant
POST /api/scraper/run	Scrape products

Scraping

Products are fetched from Traya.health's Shopify API (/products.json). This gives us structured JSON data directly - no HTML parsing needed.

Folder Structure

backend/
  app/
    api/        # Routes
    models/     # Database models
    services/   # Business logic
    scraper/    # Product scraper

frontend/
  src/
    components/ # UI components
    pages/      # Page views
    services/   # API calls

Problem i faced

The Render free tier only provides 512 MB RAM, which isn’t enough to load the sentence-transformers model (it requires ~800 MB with PyTorch). So instead, I use an LLM to handle the conversation and then generate the recommended product.

Locally, the suggestion workflow creates an embedding of the user’s query, refines the query, and then performs a semantic similarity match between embeddings to identify the top 3 results.

Screenshot