Abstract: A system and method are provided for supporting dense retrieval of a template (e.g., a document template) for responding to a query or other textual input. The templates and past queries that were responded to using the templates are stored. A machine-learning model for matching a new query to the most appropriate template is trained using a selected subset of the stored queries as training queries. For each of one or more training batches or phases, multiple stored templates are selected (e.g., randomly) then, from among all training queries that the selected templates were used for, the same number of queries are selected (e.g., randomly), such that they represent the distribution of the training queries among the selected templates. A unique loss function is computed that leverages similarities and differences not only between each selected training query and each selected template, but also between different queries and between different templates.