BNP Paribas Virtual Assistant Chatbot

The BNP Paribas Virtual Assistant Chatbot is a production-ready RAG (Retrieval-Augmented Generation) system designed to help customers answer questions about BNP Paribas banking products and services. This project demonstrates a complete end-to-end implementation of a modern LLM-based application, from data collection to production deployment.

The system combines cutting-edge technologies including OpenAI GPT-4 for language understanding, FastAPI for backend infrastructure, Next.js for a responsive frontend, and Azure for scalable cloud deployment. What makes this project particularly interesting is the intelligent combination of web scraping, vector databases, and LLM orchestration to create a system that understands banking context and provides accurate, sourced answers.

When visiting the BNP Paribas website, customers are often overwhelmed with information. Finding answers to specific questions about banking products, services, or procedures requires navigating multiple pages or contacting customer support. This creates friction in the customer journey and puts unnecessary load on support teams.

The solution: A conversational AI assistant that understands BNP Paribas's entire knowledge base, retrieves relevant information instantly, and provides accurate, sourced answers in the customer's preferred language.

The system architecture follows a three-tier design: frontend user interface, backend API with RAG logic, and cloud infrastructure for scalability and reliability.

1. Data Collection & Preprocessing

The system extracts publicly available information from BNP Paribas web pages through responsible web scraping. The collected documents are then processed and chunked into manageable pieces for embedding.

2. Vector Database Construction

# Vector Database Creation Process
1. Document Loading
   └─ Load from rag_documents.json

2. Text Splitting
   └─ Chunk size: 1000 characters
   └─ Overlap: 200 characters

3. Embedding Generation
   └─ Model: paraphrase-multilingual-MiniLM-L12-v2
   └─ Supports French & English

4. Vector Storage
   └─ ChromaDB storage
   └─ Semantic indexing
   └─ Fast retrieval enabled

3. RAG Pipeline

When a user asks a question, the system:

1. Converts the question to embeddings
2. Retrieves top-k most relevant document chunks
3. Constructs a context window with retrieved documents
4. Sends context + question to OpenAI GPT-4
5. Returns AI-generated answer with source attribution

4. Testing & Validation

Testing with Postman - Query to `/query` endpoint

The system was thoroughly tested with example queries in both French and English, demonstrating accurate information retrieval and answer generation.

The application is deployed using a comprehensive CI/CD pipeline with automatic testing, building, and deployment to production.

Backend Deployment (Azure)

Frontend Deployment (Vercel)

GitHub Secrets & Environment

Sensitive information is securely managed through GitHub Secrets:

• • AZURE_CREDENTIALS - Service Principal JSON
• • OPENAI_API_KEY - GPT-4 API Access

Live Chatbot Demo

Production chatbot interface ready for user queries

Example Query Response

Real API response showing source attribution and detailed answer

Performance Metrics

The BNP Paribas Virtual Assistant Chatbot demonstrates a complete, production-ready implementation of modern LLM-based applications. By combining RAG techniques with cloud-native infrastructure and comprehensive CI/CD automation, we've created a system that's both powerful and maintainable.

Key takeaways for anyone building similar systems: invest in robust data preprocessing, implement proper source attribution for trustworthiness, use multilingual embeddings for international reach, and automate your entire deployment pipeline for reliability and speed.