Hospit-ai-lity: Increasing Rag Accuracy And Effectiveness On The Back-end For Better Guest Interactions

By Author: D Edward Levy
Total Articles: 27
Comment this article

With the global travel lockdowns far in the rear view mirror, the travel industry forges ahead and travel consumer preferences are shifting to more personalized experiences. A recent survey from Book4Time indicated that 61% of consumers were willing to spend more for personalized travel experiences. For travel providers, that is a significant statistic as the individual transaction value of the trip or engagement is the bread and butter of the business. But how do travel providers personalize and elevate guest experiences? The experience itself starts with the frontline team that services guests in varying capacities. Those individuals working along the frontlines must have essential and advanced guest service skills to really connect and engage guests while on the property or interacting in guest service areas. That’s where Walt and hospit-AI-lity come in. By delivering advanced guest service training in a shorter time frame and in a more effective manner, those guest facing workers will have the tools and skills they need to maximize each guest interaction and ultimately, elevate experiences and increase transactional value.

The ...
... Mechanics that Power Walt

General AI platforms sometimes suffer from what the industry calls ‘hallucinations’ where the responses given back to the user are wrong, slightly incorrect or may not even be related to the question being asked. Purpose built AI platforms can reduce and eliminate hallucinations by utilizing RAG pipelines for data sources. RAG stands for Retrieval Augmented Generation and aptly describes the system where the technology ‘retrieves’ the data presented and formulates responses based on those hard sources. The large language model, or LLM, is used only to assemble wording and sentences so that responses are coherent and ultimately, make sense. RAG based systems need the following elements to perform correctly:

A knowledge base or data sources that are assembled efficiently for readability
Loader scripts that can effectively ‘read’ and tokenize the presented data sources
Vectorization scripts that break down the data mathematically
An LLM for user input and model output

This is a simplified list for a complex process that compares user queries versus available data mathematically to achieve the best possible response. The major benefit of RAG frameworks are the ease of updating and upgrading data sources, enabling custom skills and development opportunities for existing travel workers.

Improving Retrieval for Specific Tasked Models

The First Generation of Walt is a simple RAG system to evaluate the core mechanics of the system. Additional methods and redundancies for Walt 2.0 will ensure higher performance, speed and accuracy in user interactions and future training roles:

Multi Query Checks – Walt will break down user responses and reword them to mimic the context the user presented. By using vectorization against his own formulated responses, Walt will better understand the context of the user’s original question.
Mathematical Back-checks – By integrating formulas such as cosine similarity and Euclidean Distance, Walt will be able to gauge accuracy of vectors more precisely when creating user facing responses.
Resource Handling – Optimized document structuring for improved readability and loading combined with higher performance ‘chunking’ strategies will increase Walt’s precision in tokenization and vector graphing. Better text chunking will also reduce latency.

Walt 2.0 will also have a user friendly, Chat GPT style user interface for the next generation evaluation. Deployment of Walt 2.0 is slated for mid January 2026.

Delivering Essential Training Content in Real Time

Walt’s role will be as a travel worker trainer, delivering essential and advanced skills for guest service representatives and other entry level workers within the organization. The evolution of Walt’s structure will see him minimized to a fully interactive mobile app, accessible from handheld devices and tablets. Training content, evaluations and media will be delivered directly through the app in a highly engaging and entertaining format. A local, server based LLM can be optimized for mobile device access and task specific usage in a limited environment ensures performance and speed. Successful evaluation and testing of mobile specific LLMs will offer another dynamic option for hosting Walt and the hospit-AI-lity training platform. What took weeks with traditional methods of training, takes only days with Walt’s high level of engagement, real time in the field learning and field and consistent feedback.

The 2nd Generation of Walt

The First Generation Walt evaluation is available and is a simple interaction portal where users can ask Walt questions about guest services to test the viability of the RAG pipeline. Walt 2.0 is currently in development and will have a host of improvements including:

2 way interactions and user facing questioning
A faster and more user friendly interaction portal with chat history
Graphics pull and display functionality for relevant infographics and diagrams
Increases in complexity to the knowledge base and addition of a basic training map and package

Along with the future minification of the technology and migration to a local LLM, Walt will be able to pull short videos and other media related to training content to improve training and skills delivery effectiveness. Walt will also have worker and trainee evaluation capabilities in future versions. Visit the TTS Evaluation Portal to give Walt a try and share your feedback. Access here: https://traveltechnologysolutions.net/EvalPortal/