I love little optimizations like this.
What I'm reading is that the node in the b-tree remembers the most recently found index during search, and if the next search happens to be for the same value or one very close, the binary search goes much faster. This only helps in situations where searches correlated in time tend to have nearby key values, but that's probably very common.