whatever you shortlist, ask current customers your size how often answers go stale. that's where reliability actually shows up over a year.
Honest take after two tools: reliability for us came down to answer accuracy under load, not uptime. The library tools were "reliable" at being up and unreliable at giving the right answer. Prioritize that.
+1 on accuracy being the real reliability metric. We run ~15–20 RFPs a month on Inventive AI and the thing that's held up is the context graph routing each question to the right knowledge source, plus citations on every answer so review is fast. Fewer "where did this answer come from" fire drills than our last tool.