Hello, I'm Khusyasy

< back |home/projects/custom-chatbot-rag
LLM-RAG Powered Customer Service Chatbot for Telkom
20 Sep 2024

LLM-RAG Powered Customer Service Chatbot for Telkom

Developed a prototype WhatsApp chatbot using a Large Language Model (LLM) with Retrieval-Augmented Generation (RAG) to provide efficient and accurate customer service.

Overview

  • Role: Data Analyst / AI Programmer Intern
  • Problem/Goal: This project is aimed to increase the efficiency of Telkom's customer service by developing a 24/7 automated chatbot. The main goals is to improve customer experience by providing fast and accurate responses to customer questions that is easily accessible via WhatsApp.

Key Features

  • Used Retrieval-Augmented Generation (RAG) to ensure responses are based on a specific factual data.
  • Built a custom knowledge base in Markdown format.
  • Integrated with WhatsApp for easy access by customers.
  • System architecture designed to be deployable in company's intranet.

Tech Stack

  • Core: Python, Langchain, Huggingface Transformers
  • AI/LLM: Azure OpenAI, RAG (Retrieval-Augmented Generation)
  • Database: ChromaDB (for vector storage)
  • API: Meta / WhatsApp API

Development Process

At the start of the project, the first step I do is researching and planning the chatbot's architecture, then collecting and cleaning Telkom's data to build the knowledge base. I then developed the RAG system, where user queries are embedded and used to search for relevant information in ChromaDB. This retrieved context is then passed to the Azure LLM, along with the user's question, to generate an accurate response. This main pipeline is made using the langchain library.

Documentation

Retrieval-Augmented Generation (RAG) System Flowchart

RAG system is built by combining Retrieval and Generation:

  • Retrieval: process used to search relevant information from the knowledge base.
  • Generation: process for LLM model to generate answers based on the data found.

System Architecture

Knowledge Base Flowchart

Chatbot System Flowchart

Results

Chatbot Interaction Example

  • The RAG-based LLM chatbot was successfully developed as a prototype and can be accessed via WhatsApp for ease of use.
  • The chatbot can accurately answer specific questions related to Telkom, tested manually with 67 questions showed that the chatbot achieved an 88.81% accuracy rate.
  • The chatbot average reponse time is around 6.27 seconds.
  • This prototype can serve as a important foundation for further development and implementation of LLM-based chatbots in Telkom's customer service.