receiving, through a user interface, a first query from a user associated with a vehicle dealership;

applying the first query to a first trained supervised machine learning model configured to predict an intent of the first query, the first trained supervised machine learning model trained using training samples that show one or more inputs as labeled by a corresponding intent, the training samples each labeled using an intent label of at least three candidate intent labels, the predicted intent predicted to correspond to a given label of the at least three candidate intent labels, each of the at least three candidate labels corresponding to different respective databases of the vehicle dealership storing different data from one another;

applying the first query to a second trained supervised machine learning model configured to predict a set of entities of the first query, at least a portion of the set of entities comprising one or more words of the first query, the second trained supervised machine learning model trained using historical user input paired with one or more labels indicating one or more entities;

generating a normalized representation of the first query that includes both the predicted intent and the predicted set of entities, wherein a format of the normalized representation of the first query is database language agnostic, and wherein generating the normalized representation comprises filtering out a subset of metadata from an interim representation of the first query that precedes the normalized representation of the first query, the filtering out based on an importance value for each metadata of the subset being below a threshold level of importance;

translating the normalized representation of the first query into a second query having a format compatible with a language of a database of the vehicle dealership, the database selected from the different respective databases based on the predicted intent;

fetching data from the database of the vehicle dealership associated with the predicted intent and the predicted set of entities using the second query;

providing, on the user interface, the data for display to the user;

receiving feedback from the user by way of the user interface that reflects an error in the data; and

re-training the first trained supervised machine learning model based on the feedback, thereby resulting in a different intent label being predicted responsive to again receiving the first query.