Many real planning tasks include both harder “quantitative” restrictions (e.g. budgets or planning requirements) as well as softer “qualitative” goals (e.g. user preferences expressed in the natural language). Consider someone who is planning a one -week vacation. As a rule, this planning would only be subject to various clearly quantifiable restrictions such as budget, travel logistics and visiting attractions if they are open, in addition to a series of restrictions that are based on personal interests and preferences that are not easy to quantify.
Large -speaking models (LLMS) are trained on massive data records and have internalized an impressive amount of world knowledge, often an understanding of the typical human preferences. Therefore, they are generally good to take into account the parts of the travel planning that are not so quantifiable, e.g. However, they are less reliable in order to treat quantitative logistical restrictions, for which possible and current real information (e.g. bus trips, train plans, etc.) or complex interacting requirements (e.g. minimizing travel over several days). As a result, LLM-generated plans can temporarily include impractical elements, e.g. B. visiting a museum that is closed at the time of travel.
We recently introduced AI trip ideas in the search, a function that suggests the travel routes during the day in response to trip-planning questions. In this blog we describe some of the work that carried out one of the most important challenges when starting this feature: ensuring that the travel routes produced are practical and feasible. Our solution uses a hybrid system that an LLM uses to propose a first plan in combination with an algorithm, which is jointly optimized for the similarity to the LLM plan and the real factors such as travel material and opening times. This approach integrates the ability of the LLM to meet soft requirements with the algorithmic precision that is necessary to meet hard logistical restrictions.