Our client, a leading company in the education sector, faced the challenge of managing a high volume of customer support requests. With a growing customer base, they struggled to provide timely and personalised support to their users.
They were looking for a solution that could automate parts of their customer support system while maintaining the company's tone, style, and quality of interaction.
What we did
✔︎Developed an AI customer service agent using retrieval augmented generation.
✔︎The process involved chunking Knowledge Base articles, creating vector embeddings, and finding the top 5 matches for user queries.
✔︎Utilized the Prompt Builder to generate ChatGPT API prompts.
✔︎Extensive testing and fine-tuning were performed against diverse customer queries.
The AI agent, developed as a Minimum Viable Product (MVP), surpassed our client's expectations in just 10 weeks.
It achieved an impressive response accuracy of 96%, far exceeding our project target of 50%. The AI agent transformed the user experience by providing timely and personalised support, effectively streamlining customer service operations for the company.
correct AI response. Eclipsing project target of 50%.
How we did it
We developed the AI customer service agent using a technique called retrieval augmented generation, utilising Python, a large language model (GPT-3.5), and a vector database.
This AI-driven conversational agent was designed to extract relevant information from the company's knowledge base and craft articulate responses to user inquiries. Our primary objective was to harness Generative AI to improve efficiency and scalability, all while maintaining a high level of personalised engagement.
The client's extensive Knowledge Base was divided into manageable chunks, processed using OpenAI's Embedding API, and the resulting vectors were stored in a Qdrant vector database.
When user queries were received, we employed cosine similarity to identify the top 5 most relevant matches from the Knowledge Base vector embeddings. These matches were then used to generate prompts for the ChatGPT API. The API produced the final responses, which were meticulously validated against established business rules before being delivered to users.
Our approach included rigorous testing, fine-tuning, and the creation of a testing environment to ensure the AI agent's efficiency and accuracy in responding to a wide range of customer queries.
Download our case study
Start a conversation
Take the first step by speaking with one of our data experts today.