Improvements And Upgrades For The 2nd Generation Of Walt

By Author: D Edward Levy
Total Articles: 53
Comment this article

Travel training is a critical operational facet of any travel business and gaps in service and quality can quickly generate negative impacts to the business and revenue. Having skilled and well trained frontline teams is crucial to the success of any travel brand as guest experiences can dictate and influence travel brand perception and positioning. Realizing the need for fast and effective training in the travel industry, TTS is building hospit-AI-lity, a next generation travel training platform that delivers faster and more effective travel training. At the core of hospit-AI-lity is Walt, the lead travel trainer and face of hospit-AI-lity. The first generation evaluation of Walt is currently online and TTS is putting the finishing touches on Walt 2.0 with major improvements and upgrades in both data processing and the evaluation experience.

Formatted and Curated Data Sources

RAG, or Retrieval Augmented Generation, provides the framework for Walt and at the heart of any RAG system is the knowledge base. This is the curated data that the model uses as reference when generating output back to users. Knowledge bases ...
... can be made up of any documents such as PDFs, Word and even images but the most common document format is Markdown. Documents created in Markdown are easier to read and process for RAG based systems and their actual structure can be key to speed and accuracy of the model’s output. Even using simple formatting elements, such as headings and subheadings can improve system readability, vectoring and overall response accuracy. For the next generation of Walt, a curated basic training map will be installed for the next evaluation phase, increasing the size and detail from the first generation of data sources. For travel businesses, the curated data source is an advantage as the documents can easily be changed, upgraded or removed to optimize Walt’s responses based on what is needed at the current time.

Loads and Chunks

Key processes in RAG systems are ‘loading’ and ‘chunking’ and refer to how the system acquires data and prepares it for analysis. When loading, the system actively looks for main folders and scans for documents within the folder and sub folders. The technology identifies the type of document and reads the data in its entirety. Chunking is the next process where the system breaks the entire document into manageable parts for the next process of embedding and vectoring. For Walt 2.0, document loading strategies and folder structures have been optimized for faster speeds and higher quality reads. Chunking scripts and commands have been upgraded to reflect sentence size within the user created documents and a more linear method to paragraph structure has been adopted to take advantage of better chunking strategies. More efficient and concise ‘chunks’ will lead to more effective embedding and vectorization, the gateway to the comparison and response generation processes.

Better User Evaluation Experience

The next generation of Walt will get a brand new user evaluation portal to reflect the new TTS schema and Bootstrap integration. Similar to ChatGPT, Walt will employ a long form scrolling perspective so that queries and responses remain in the window and can easily be checked visually and in real time. A separate ‘document referencer’ window will display the document used in responses to improve evaluation and calibration. The Document Referencer is for evaluation only and will not appear in any user facing forms of Walt but will be available in a management console. Walt 2.0 will also see the introduction of random questions and quizzes that are based on the entire user interaction as well as the knowledge base. This is the first phase of Walt’s evaluation capabilities when training travel workers in the real world.

Walt 2.0 Leads the Way

Walt 2.0 represents the evolution of the technology and is the gateway to eventual real world scaling and deployment. Future versions of Walt will migrate to a localized large language model for improved speed and elimination of cloud LLM costs. Scaled versions of Walt will see the technology in mobile app format and accessible on mobile devices. Trainees will be able to access Walt on their own smartphones or business provided tablets for on the go and real time interactions. Continued evolution will realize Walt’s evaluation capabilities to ‘quiz’ and ‘check-in’ with trainees as their training progresses. Better response memory and advanced vectoring and redundancy checks will ensure smooth and accurate data flows between Walt and knowledge bases. To check out the latest available version of Walt, visit the Evaluation Portal to get access and take him for a spin.