📚 RAG ChatBot in E-commerce

My Team: Cao Minh Tri - Tran Duong Tuan

Link demo: https://www.youtube.com/watch?v=21EZ339xix8

---

1. Project Overview

This project implements a Retrieval-Augmented Generation (RAG) chatbot that can answer user queries by combining semantic search with a large language model.
The system retrieves the most relevant document chunks from a PDF knowledge base and uses a generative AI model to produce context-aware answers.

The pipeline includes:

• Extracting text from PDF files.
• Splitting the text into manageable chunks.
• Creating vector embeddings for semantic similarity search.
• Using FAISS to store and query these embeddings efficiently.
• Providing the retrieved context to a large language model for answer generation.

---

2. Motivation

• Provide accurate information to customers when making online purchases.
• Reduce time and cost of customer support operations.
• Allow fine-tuning of the LLM for specific business needs, enabling greater personalization.

---

3. Technologies Used

Programming Language

• Python 3

Core Libraries

Library	Purpose
transformers	Load and use the google/gemma-2-2b-it model for text generation.
faiss-cpu	Vector database for fast nearest neighbor search.
PyPDF2	Extract text from PDF documents.
nltk	Sentence tokenization.
pandas	Data handling and storage of chunks + embeddings.
numpy	Vector manipulation.
gradio	Graphic user interface.

Environment

• Google Colab (with GPU support).
• Google Drive (for storing and loading PDFs).
• Hugging Face model access via personal token.

---

4. How It Works

1. Data Loading

   • PDF files are loaded from Google Drive.
   • Text is extracted using PyPDF2.

2. Text Splitting

   • Documents are split into sentence-based chunks using nltk.sent_tokenize.
   • Each chunk size is limited to ~1000 characters for better processing.

3. Embedding Creation

   • Each chunk is converted into a dense vector embedding using the all-MiniLM-L6-v2 model from SentenceTransformers.

4. Indexing with FAISS

   • All embeddings are stored in a FAISS IndexFlatL2 index for similarity search.

5. Retrieval

   • When a user asks a question, the query is embedded and matched to the top k most similar chunks.

6. Generation

   • The retrieved chunks are passed as context to the google/gemma-2-2b-it model.
   • The model generates a final, context-aware answer

---