In the case of the relay search, it doesn't involve a graph—it's just a simple list, since the number of relays is very small and is not likely to increase significantly. The similarity calculation is the same, but it's done by iterating over a simple array.

On the front end, a POST request is made to the relay, which eliminates any relays that don't respond to the request, listing only the active relays. This ensures that only active relays will appear in the search results. Another thing I implemented is a 2-second timeout for the requests made to the relays, which also ensures that only the relays closest to you, and thus more relevant, will appear.

I only use a graph for the users. The relays are mapped in kind 3 events, which contain a list of relays and are sent to the engine that saves them in a list loaded into RAM. The graph is also loaded into RAM at the application startup, and everything runs in memory.