Knowing that the more rules sets you add, the longer it takes to scan a message, what rule sets do you recommend? I have found several sites with sure sets such as The Rules Emporium and the SA wiki but I am certain I do not need every rule set from those sites.
For what little it's worth, this is my general advice on improving hit rations (covers add-on modules and rulesets). Bear in mind this is just my advice, and I'd really suggest trying things out one-at-a-time so you know how well each works for you:
1) Make sure you have Net::DNS installed, this is an optional perl module, but it allows the RBLs and SURBL checks to run.
(If you run spamassassin --lint -D you can watch SA test for DNS functionality.)
2) I've grown to like DCC quite a bit, but never trust it entirely. It does FP, but this is rarely a problem if you don't jack the score up.
3) Same goes for Razor. Just make sure it's 2.61 or better.
4) I really like the "Random" rules from rulesemporium.com. Not very high hit rate, but very good S/O on these.
5) feed your bayes DB as much fresh meat as you can (if you can).
For those with 2.6x versions I add antidrug, backhair and surbl to my recommendations, but 3.0 comes with these.