When I started working with Getaway to develop scoring algorithms for flights and hotels, I knew we had to approach the task out of the box. The challenge was to create a system that could evaluate and rank thousands of travel options according to complex criteria. Instead of traditional methods, we chose advanced language models such as ChatGPT and Llama.

How was the process of working with language models? We break down our approach in detail.

Problem: Challenges of creating business requirements using large language models ​

Organising travel on the Getaway platform required flexible and precise algorithms to score flights and hotels. To this end, we used advanced language models such as ChatGPT and Llama to create a unique scoring system. The challenge was to adapt the algorithm to the specific needs of Getaway and its users, taking into account parameters such as the number of connecting flights, departure time, hotel location and quality of amenities.

Creating a scoring system for flights required a thoughtful approach. The main goal was to create an algorithm that could score different flight and hotel offers based on a number of factors – not only such as price, but also the quality of the amenities offered and the specifics of the trip.

This is where the ability to test different approaches and quickly iteratively adjust the context of the prompts was key.

Let’s talk business.

#1 Working with prompts – context and iterations ​

At the initial stage, we focused on designing prompts that appropriately contextualised the scoring requirements. For flights, key factors included:

• Price,
• Total flight time,
• Number of transfers,
• Departure/arrival times,
• Number of airports visited.

Using an iterative approach, we developed the MVP of the scoring algorithm in just two days, which allowed us to quickly get an initial picture of the model's effectiveness and conduct the first tests.

#2 Implementation of a reward and punishment mechanism ​

Following the development of the prompting system, we started to implement a reward and punishment system. It allows the user's preferences for a particular flight or hotel to be taken into account. For example:

• Flights with a number of connecting flights higher than the average received negative points,
• More expensive flights or hotels, exceeding the average prices for the date, were 'penalised' with a percentage reduction in value.

This solution increased the precision of the results and allowed the results to be dynamically adapted to the user's preferences.

By introducing penalties for offers that deviate from user preferences, we have achieved a system that is tailored to the real expectations of Getaway customers, which has had an impact on their satisfaction with the results.

Let’s talk business.

#3 Testing and adjustment of the mechanism at the UAT stage ​

In the final stage, we implemented the algorithm on a test environment (UAT) and conducted intensive unit and integration tests. The scoring functionality was further optimised based on the test results, ensuring that it was in line with real user needs. This phase also allowed for the calibration of scales and parameters.

• Software development: Optimising code for fast query processing.
• Design and maintenance of prompts: Preparation of ready-made sets of prompts for future use.
• Reward/punishment mechanism: Clarification of thresholds and penalties for less desirable parameters.
• Unit and integration tests: Ensuring consistency of results across all use cases.

Let’s talk business.

Effect: an integrated scoring system for TravelTech ​

By combining large language models and advanced scoring techniques, Getaway has gained a scalable and dynamic recommendation system for flights and hotels. Customisation mechanisms allow users to make decisions faster, and the entire platform has become more personalised and user-friendly for end users.

Getaway's turnkey scoring system is prepared for future extensions, making it flexible and resilient to changes in user preferences and new developments in the TravelTech industry.