The NSA records BILLIONS of telephone calls, emails, twitter posts, internet searches, IMs, Facebook posts, etc A DAY. Of that data approximately .5-1% is analyzed by automated tools for further analysis (officially The NSA is only allowed to analyze communications that have one leg outside the US border, but leaked information indicates that all information is analyzed for people of interest) That's upwards of 20 million a day IN REAL TIME. Although the FBI certainly doesn't have even near the NSA's computing power they wouldn't need it.
I believe Fuji's statement is correct, they would perform initial filtering to eliminate emails they've already seen and emails that only pertain to Weiner and his x-wife with no reference to any keywords, strings, aliases or logical connection to emails of interest. That just leaves only net-new emails of interest already sorted and key-worded to be manually examined by the agents. No way of knowing what the number was.
This isn't rocket science, most larger police forces have this technology in-house. Previously I sold Enterprise class content filtering devices that could in real time monitor and block end-user emails, IMs, web content and file transmission within a company IP infrastructure. These appliances could scale from 1K-5K items checked per second and be be further scaled using load balancer technology. 1K/sec = 3.6M/hour. These devices didn't just look for keywords, they were context sensitive and were able to flag and block search parameter aliases. IBM offers even more power AI-based tools that perform similar functions, but their analysis algorithms are significantly more powerful than the appliances I sold.