There is a practical limit that we can't cache results for too long; Search engine users are particularly sensitive to stale data, especially around current events. Without a holistic and realiable way to know when the cache ought to be invalidated, our caching is mostly focused on mitigating "abuse", e.g., someone / bunch of people spamming the same search in a short timespan; no sense in repeating all those upstream calls.

Most "cost saving engineering" is involved in finding cases/hueristics where we only need to use a subset of sources and omitting calls in the first place, without compromising quality. For example, we probably don't need to fire all of our sources to service a query like "youtube" or "facebook".

Marginalia data is physically consolidated into the same infra that we use for small web results in our SERP, but also among other small scale sources besides those two. That line is simply referring directly to https://kagi.com/smallweb (https://github.com/kagisearch/smallweb).