Chatbot Implementation for KPO Success

Introduction

Our customer is a leading global KPO that helps organizations deliver exceptional customer and employee experiences, driving business growth. They provide comprehensive customer service solutions, including call center support and digital assistance via chatbots.

Project Requirements

1. Development of a RAG-based Chatbot

This project requires a RAG-based Chatbot, where the contact center personnel do not need to manually refer to large knowledge-based folders of User Manuals but instead can directly query the AI Chatbot. This chatbot should not only retrieve the relevant information but also generate a solution-oriented output based on the user’s query.

2. Integration with of chatbot

Our customer stores a large knowledge base of multiple user manuals in a Microsoft SharePoint folder. They require that the chatbot can directly access and query these SharePoint files. Additionally, they require that any changes in the SharePoint files database should reflect in the chatbot’s knowledge database

3. An AWS-focused Solution

Solution to be hosted on AWS Cloud and use AWS-provided services. 

Solution/ Implementation

Our goal was to give an AWS-focused RAG solution that can be integrated with a chatbot. For this, we divided our problem into 3 sub-problems. 

1. SharePoint to AWS synchronization

To keep our solution AWS-focused, we decided to use AWS S3 as our primary datastore. For this, we needed a solution to keep both our S3 bucket and SharePoint folder synchronized. 

For this, we used Microsoft Power Automate, AWS API Gateway, AWS Appflow, and AWS Lambda. In Microsoft Power Automate, we set up an HTTP POST request that would be triggered on any change detected in the SharePoint folder. This POST request would be captured by AWS API Gateway, which in turn would trigger an AWS Lambda function, executing an on-demand AWS AppFlow flow between the SharePoint folder and our S3 bucket. This flow was configured to be incremental, thus saving us time and resources. 

2. Vector Database and Retriever Setup

The heart of a RAG solution is a vector database and a retriever. We used AWS Bedrock Knowledge base alongside its built-in Pinecone vector store to set this up.

We created a Pinecone index using Amazon Titan Embeddings V2 and a cosine similarity metric, attaching our S3 bucket as our primary data source. We used default PDF parsing alongside a fixed-size chunking strategy with 1000 max tokens and 20% overlap. 

3. Output Generation

Using the boto3 Python client, we integrated our AWS Knowledge Base with a custom chatbot that used OpenAI’s GPT-4o model.

Result 

1. Enhanced Query Resolution Efficiency

A significant reduction in query resolution time, as the chatbot reduced the need to manually search through documents

2. Seamless Integration with SharePoint

We were able to provide a seamless synchronization between SharePoint and Chatbot, with incremental document updates.

3. Security against malicious input

Using the AWS Bedrock AI Guardrails, we were able to provide peace of mind to the client against malicious and harmful prompt inputs and AI outputs.

Conclusion

The implementation of the RAG-based chatbot solution exemplifies how AI-driven innovations can enhance operational efficiency and user experience. By leveraging AWS services and integrating seamless synchronization with SharePoint, we provided a robust, secure, and scalable solution that significantly reduced query resolution time and automated knowledge retrieval processes. This project demonstrates the value of tailored AI solutions in addressing specific business challenges, paving the way for continued digital transformation..