Need help for a prompt-based matching of many items

[k-b-1]

I need your help to develop a matching procedure that will find the best match of a source element to a single target element out of 1000 targets.
I can't define a prompt that contains all the targets because that would exceed the context size. But I can iterate over the targets and compare the source element to a target element. The prompt contains three parts: the actual instruction on how to compare and what the output is, the source element, and the target element.
If I did this, llama.cpp would have to tokenize and decode the first two parts over and over. Is there a way to fix the cache for the first two parts and just swap the third part?
The second possible improvement could be that I can preprocess the target element and insert the target element as an already tokenized or even embedded element?
I would be very grateful if you could point me in the right direction or even show me some sample code.
Thank you!