Dear Miksame,

Thank you for reaching to me.

First of all, there is, to my knowledge, no semantic limitations to what you input to a Siamese network model: it can be a text, an abstract vector that you generated yourself... The only difference will be in the layers of your Neural Networks: will your vector have any notion of ordering (i.e is the 1st component of your vector related to the 2nd, in a Markov Chain kind of way):

- if there is any type of ordering, then you'd use 1D CNNs or RNN cells to process your input vectors and transform them into embeddings that will be used for comparison.

- if there is no notion of ordering, you might as well then just go for fully-connected layers. In the latter case, I agree that you would be taking less advantage of the representation capabilities of Neural Networks.

Additionnally, for this task, I'd advise you to look at pseudo-siamese networks: as you are not, strictly speaking, comparing two entities meant to have similar structure (a document and a query vector, while related, will never be considered "equals" I think), then you should opt for an architecture where you first use distinct layers to transpose both the document and the query to vectors with more related semantical meanings. You'd then use a siamese network at this point on. I hope that I have proved myself clear but I am not quite sure so feel free to contact me on LinkedIn for easier exchanges !