🧘 Yoga Wellness RAG Application

An intelligent Retrieval-Augmented Generation (RAG) system providing safe, evidence-based yoga guidance powered by vector search and AI.

---

✨ Key Features

• ✅ Semantic Search: 100 curated articles with 384-dimensional vector embeddings
• ✅ Medical Safety Detection: Identifies 8 health condition categories
• ✅ Source Attribution: Transparent citations for every response
• ✅ Complete Logging: MongoDB tracking with embeddings and feedback
• ✅ Free Operations: Local embeddings (no API costs or quotas)
• ✅ Modern Stack: React + Vite frontend, Node.js + Express backend

---

🛠️ Tech Stack

Component	Technology	Purpose
Vector DB	Pinecone	Semantic search (384d, cosine)
Embeddings	Transformers.js	Local model (all-MiniLM-L6-v2)
Generation	Groq (Llama 3.1)	Fast AI responses
Database	MongoDB Atlas	Query logging & analytics
Backend	Node.js + Express	REST API
Frontend	React + Vite	Modern chat interface

---

Detailed Pipeline

┌─────────────────────────────────────────────────────────────────────────┐
│                         USER QUERY                                       │
└─────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  1. SAFETY DETECTION                                                     │
│     • Keyword-based check (50+ medical terms)                           │
│     • LLM-based safety analysis for unsafe queries                      │
│     • Generates safe alternatives if health condition detected          │
└─────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  2. EMBEDDING GENERATION                                                 │
│     • Model: all-MiniLM-L6-v2 (Transformers.js)                         │
│     • Dimensions: 384                                                    │
│     • Runs locally (zero API cost)                                      │
└─────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  3. VECTOR SEARCH (Pinecone)                                            │
│     • Cosine similarity matching                                         │
│     • Returns Top 5 relevant chunks                                      │
│     • Includes metadata (title, source, page)                           │
└─────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  4. LLM RESPONSE GENERATION (Groq)                                      │
│     • Model: llama-3.1-8b-instant                                       │
│     • Safe queries: Normal conversational response                      │
│     • Unsafe queries: Safety pivot with alternatives                    │
└─────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  5. RESPONSE TO USER                                                     │
│     • AI Answer + Source Citations                                       │
│     • Safety Warning Banner (if unsafe)                                  │
│     • Logged to MongoDB for analytics                                   │
└─────────────────────────────────────────────────────────────────────────┘

Simplified Flow

User Query → Validation → Safety Check → Embedding (384d) → Pinecone Search → LLM Response → MongoDB Log → Response + Sources

Component Architecture

graph TB
    subgraph Frontend["🖥️ Frontend (React + Vite)"]
        UI[App.jsx - Chat UI]
        API[api.js - API Service]
        SW[SafetyWarning.jsx]
        SL[SourcesList.jsx]
    end

    subgraph Backend["⚙️ Backend (Node.js + Express)"]
        Server[server.js]
        Routes[routes/ask.js]

        subgraph Services["Services"]
            Embed[local-embeddings.js]
            Retrieve[retrieval.js]
            Generate[generation.js]
            Safety[safety.js + smart-safety.js]
        end

        Models[models/QueryLog.js]
    end

    subgraph External["☁️ External Services"]
        Pinecone[(Pinecone Vector DB)]
        MongoDB[(MongoDB Atlas)]
        Groq[Groq LLM API]
    end

    UI --> API
    API --> Server
    Server --> Routes
    Routes --> Embed
    Routes --> Safety
    Routes --> Retrieve
    Retrieve --> Pinecone
    Routes --> Generate
    Generate --> Groq
    Routes --> Models
    Models --> MongoDB

    style Frontend fill:#e3f2fd
    style Backend fill:#f3e5f5
    style External fill:#e8f5e9

Loading

Performance Breakdown

Step	Component	Time
1	Input Validation + Safety Check	~30ms
2	Local Embedding Generation	~400-500ms
3	Pinecone Vector Search	~100-200ms
4	Groq LLM Response	~800-1500ms
5	MongoDB Logging	~100ms
Total	End-to-End	~1.5-2.5s

---

📚 Knowledge Base

Source: Common Yoga Protocol - Ministry of Ayush, Government of India

100 Articles Covering:

• Surya Namaskar (12 positions)
• Standing Asanas (15)
• Seated Asanas (12)
• Prone & Supine Asanas (18)
• Pranayama (8)
• Meditation (10)
• Safety Guidelines (15)
• Philosophy & Fundamentals (10)

---

🛡️ Safety System

Medical Categories Monitored

1. Pregnancy - prenatal, postpartum
2. Cardiovascular - heart disease, hypertension
3. Spinal - herniated disc, sciatica
4. Joint - arthritis, knee/hip issues
5. Respiratory - severe asthma, COPD
6. Chronic Pain - fibromyalgia
7. Neurological - stroke, seizures
8. Recent Surgery/Injury

Safety Response

• ⚠️ Warning banner displayed
• 📋 Detected conditions shown
• 💡 Safe alternatives provided
• 👨‍⚕️ Professional consultation recommended

---

🚀 Setup Instructions

Prerequisites

• Node.js 18+
• MongoDB Atlas account (free)
• Pinecone account (free)

Installation

1. Clone Repository

git clone https://github.com/VIDHITTS/Yoga-App.git
cd Yoga-App

2. Backend Setup

cd backend
npm install

Create .env file:

PORT=5000
MONGODB_URI=your_mongodb_uri
PINECONE_API_KEY=your_pinecone_key
PINECONE_INDEX_NAME=yoga-wellness-local
PINECONE_DIMENSION=384
GEMINI_API_KEY=your_gemini_key_optional

3. Initialize Vector Database

node scripts/setup-local-embeddings.js
node scripts/ingest-local.js

Expected: ✅ Successfully uploaded: 100 articles

4. Start Backend

node server.js

5. Frontend Setup (new terminal)

cd frontend
npm install
npm run dev

6. Access Application
Open: http://localhost:5173

---

📁 Project Structure

Yoga-App/
├── backend/
│   ├── config/              # Database & Pinecone setup
│   ├── data/                # 100 articles JSON
│   ├── models/              # MongoDB schema
│   ├── routes/              # API endpoints
│   ├── scripts/             # Setup & ingestion
│   ├── services/            # Core RAG logic
│   └── server.js
├── frontend/
│   ├── src/
│   │   ├── components/      # React UI components
│   │   └── services/        # API integration
│   └── vite.config.js
├── knowledge/               # Source PDF
└── README.md

---

🧪 Testing

Normal Query

curl -X POST http://localhost:5000/api/ask \
  -H "Content-Type: application/json" \
  -d '{"query": "Benefits of Surya Namaskar?"}'

Safety Detection

curl -X POST http://localhost:5000/api/ask \
  -H "Content-Type: application/json" \
  -d '{"query": "I am pregnant, can I do yoga?"}'

Expected: Warning banner + safe alternatives

---

📊 Performance

Metric	Value
Avg Response Time	2.4s
Embedding Generation	~500ms
Vector Search	~100ms
Articles	100
Dimensions	384

---

🎥 Demo Video

[Demo Video Link]

Demonstrates:

• Normal query flow
• Safety detection
• Source citations
• MongoDB logging

---

📝 Key Decisions

Why Local Embeddings?

• Zero API costs
• No quota limits
• Privacy-first
• Sufficient quality for domain content

Why Fallback System?

• Ensures availability
• Graceful degradation
• Users always get info

Why Pinecone?

• Managed service
• Fast similarity search
• Excellent metadata support

---

🎓 Learning Outcomes

• RAG architecture design & implementation
• Vector databases (Pinecone)
• Local ML models (Transformers.js)
• Safety engineering for wellness apps
• Full-stack React + Node.js development
• Production API practices

---

📖 Citations

Source: Common Yoga Protocol, Ministry of Ayush, Government of India

---

👤 Author

Vidhitt S
GitHub: @VIDHITTS

---

Built with ❤️ for safe & accessible yoga education