Question about the condition representation

[teass-t]

First of all, thank you for your excellent work.

I noticed that in the figure above, c_i is described as the vector obtained by feeding the input sequence into the LLM and aggregating its output. However, I am confused about how a single vector can be processed by a Transformer encoder.
Therefore, as shown in the figure below, it seems that the condition is actually obtained from the answer, which is generated by the structure_response function.

def structure_response(target, item_list):
title = str(np.array(item_list['title'])[int(target)])
category = str(np.array(item_list['category'])[int(target)])
response =f"The user might lean towards ({category}) products based on an analysis of their purchase history, like {title}."
return response