More or less a workaround for the context window limitations of today’s Large language models: given the context of the conversation, fetch the most relevant information (from a too-large set of tokens) and supply it along with the request.