I am working on an ecommerce application where I have eaters and restaurants. Suppose that I am able to generate vector for eaters somehow, then can I find vector for restaurant using weighted average of eaters who ordered in last 90 days? I assume so but are there any counter arguments to this approach.

I plan to put these eater and restaurant vectors in vector DB. When an eater opens the application I need to rank these restaurants according to cosine similarity(will use approx NN). Do note that the first step would be retrieval and will be handled by a different system. Only retrieved candidates need to be ranked via similarity vectors