Our Approach

Maximising the full scope and potential of a GenAI chatbot: The path to production

Equal Experts has previously engaged with Travelopia on a number of successful projects, which prompted the decision to work together again on this particular opportunity. When Equal Experts arrived at this engagement, Travelopia was already utilising a chatbot for a different brand. Although the chatbot was similar in nature to the one Enchanting Travels needed, it was developed as a proof-of-concept by internal Travelopia staff. Although the chatbot did function adequately, there was limited effort put into refinement, completeness, and implementing the full scope and potential of the chatbot. 

Equal Experts used the existing chatbot as a starting point, but to meet the requirements of Enchanting Travels, we needed to add more data sources to the chatbot to automate data ingestion, and ensure all the information held was searchable and retrievable. These were to be found in the original documents the coordinators would previously have manually searched to find information, and were available in many formats including PDFs and spreadsheets. 

The addition of these extra data points for the chatbot served to increase the accuracy of the responses, and allowed for citations and justifications for those results. The data is kept up to date and in sync by an internal team within Travelopia who simply upload documents to a cloud storage bucket (AWS S3), from which the data is then extracted automatically.

We began by writing a connector for each document type (PDF, Excel, PPT, etc.) that could read and interpret the data and store it within the LLM (therefore building context for the chatbot.) However, this has now been replaced by AWS Knowledge Base, which can automatically identify each file type, pass it through the LLM and store it accordingly. This effectively removes a lot of the manual developer effort involved in the process, and doesn’t require adaptation of original files. Although this didn’t present any real challenges, we were careful to trial various methods and models for accuracy and efficiency of results to reduce response time. We also made some visual alterations to the interface of Athena to reflect the Enchanting Travels brand. 

The chatbot’s responses were validated with manual testing. We referred to documents and outlined questions to check if the chatbot was responding with the expected answers. Since this was an internal tool, we were able to capture feedback from end users via weekly reports and a ‘thumbs up’ or ‘thumbs down’ reaction for  the chatbot’s response

A user-friendly, intuitive, and speedy customer service tool

In terms of functionality, the front end is similar to Chat GPT. The coordinator enters the customer’s question and Athena locates and presents the information immediately, alongside links to any documents in which the data is stored. It holds context throughout an interaction, so coordinators can request further information to narrow results if necessary.

The product captures and retains data around what information is being requested and whether requests are successfully fulfilled, which can be used to identify gaps in the knowledge base and enrich the data available. 

A pilot version of Athena was rolled out to a limited number of users within seven days.

The data is currently being ingested monthly by Athena, but this can be increased if necessary. We are now exploring ways of working with AWS, which would allow a fully automated method of Athena identifying and ingesting data.