Multilingual RAG

Table of Contents

• 🎥 Demo Video

• 🏗 Architecture Overview

  • 1. Document Ingestion & Extraction
  • 2. Embedding & Vector Storage
  • 3. Query Processing & Retrieval
  • 4. RAG Generation Layer
  • 5. User Interface & Deployment

• 🚀 Quick Start

  • Online Demo
  • 1. Prerequisites
  • 2. Environment Configuration
  • 2. Running with Docker
  • 3. Running Locally (From Source)
  • 7. Project Structure
  • 9. LLMOps Evaluation

---

🎥 Demo Video

---

🏗 Architecture Overview

The Multilingual RAG System is designed for Arabic and English document understanding. It separates document processing, storage, and querying into clear layers to ensure scalability, multilingual support, and high-speed retrieval.

1. Document Ingestion & Extraction

• Converts raw PDFs/DOCs into structured text blocks with metadata.
• Supports complex layouts (tables, multi-column text, footnotes).
• Automatically detects Arabic (RTL) vs. English (LTR) content.
• Outputs clean, chunked text ready for embedding.

2. Embedding & Vector Storage

• Converts text chunks into multilingual semantic embeddings.

• Stores vectors in Qdrant Cloud with:

  • Language-specific collections (Arabic / English)
  • High-speed vector similarity search
  • Cloud scalability for large datasets

3. Query Processing & Retrieval

• Detects query language automatically.
• Retrieves top semantic matches from the relevant collection.
• Aggregates context for accurate, language-specific responses.

4. RAG Generation Layer

• Powered by Groq-accelerated Llama models.

• Produces multilingual answers enriched with:

  • Contextual reasoning
  • Source citations (page numbers & document names)

5. User Interface & Deployment

• Streamlit web app for interactive queries and document uploads.
• Streamlit Cloud deployment with secure secrets management.

---

🚀 Quick Start

Online Demo

Try the live app: https://multilingual-rag-app.streamlit.app/

---

1. Prerequisites

• Docker installed

• A free Qdrant Cloud account

• API keys for:

  • Groq (LLM)
  • Qdrant (vector DB)
  • LlamaCloud (PDF extraction)

Required API Keys

• Qdrant Cloud (Vector Database) Go to Qdrant Cloud → Create a free account → Create a new cluster → Copy your cluster URL and API key

• Groq (LLM Provider) Visit Groq Console → Sign up and create API keys (We use two keys for load balancing Arabic and English)

• LlamaCloud (PDF Extraction) Go to LlamaCloud → Sign up and get your API key (Required for document parsing)

---

2. Environment Configuration

Create a .streamlit/secrets.toml file in the project root with your API keys:

# Qdrant Cloud Configuration
qdrant_url=https://your-cluster-url.qdrant.tech:6333
qdrant_api_key=your-qdrant-api-key

# Groq API Keys (for LLM)
groq_api_key=gsk_your-groq-api-key-for-arabic
groq_api_key_1=gsk_your-groq-api-key-for-english

# LlamaCloud API Key (for PDF extraction)
llama_cloud_api_key=llx_your-llamacloud-api-key

---

🚀 Running the RAG App

1. Running with Docker (Prebuilt Image)

You can pull and run the prebuilt image from Docker Hub directly.

Windows PowerShell

# Run prebuilt image from Docker Hub with secrets mounted
docker run --rm -p 8501:8501 `
  -v ${PWD}/.streamlit/secrets.toml:/app/.streamlit/secrets.toml `
  ymack/my-rag-app:latest

Linux / macOS

docker run --rm -p 8501:8501 \
  -v $(pwd)/.streamlit/secrets.toml:/app/.streamlit/secrets.toml \
  ymack/my-rag-app:latest

Then open your browser to: http://localhost:8501

---

2. Running Locally (From Source)

If you prefer to build and run the image locally:

# Clone the repository
git clone https://github.com/yusufM03/multilingual-rag.git
cd multilingual-rag

# Build the Docker image locally
docker build -t my-rag-app .

# Run the container with secrets mounted
docker run --rm -p 8501:8501 \
  -v $(pwd)/.streamlit/secrets.toml:/app/.streamlit/secrets.toml \
  my-rag-app

---

Make sure .streamlit/secrets.toml exists before running the container.

7. Project Structure

yusufm03-multilingual-rag//
├── .streamlit/
│   └── secrets.toml
├── config/
│   ├── __init__.py
│   └── settings.py
├── docs/
│   ├── Agentic RAG Task.pdf
│   ├── overview.png
├── Documents/
├── models/
│   └── rag_response.py
├── src/
│   ├── __init__.py
│   ├── app.py
│   └── core.py
├── utils/
│   ├── language_detector.py
│   └── store_logs.py
├── .dockerignore
├── .gitignore
├── dockerfile
├── readme.md
└── requirements.txt

---

9. LLMOps Evaluation

Integrated with Weights & Biases (W&B) Cloud for:

• Version control of prompts & models
• Retrieval monitoring & drift detection
• Feedback-driven optimization
• Regression testing & batch evaluation

User Query → Qdrant Retrieval → LLM Response
       ↓                   ↓
  Retrieve Top-K        Compute Evaluation
   (context)        ┌──────────────┐
        └─────────▶│ Auto Metrics │
                   │ + W&B Logging│
                   └──────────────┘
                           ↓
              Continuous Monitoring & Improvement
                           ↓
                 Batch Regression Testing
                           ↓
               Model/Prompt Updates & Deployment