I've done a basic RTE in React for a client. I've done three now and RTEs are always annoying but the difficulty in this case is largely in how you model it. I chose to create a JS object per block element (usually a paragraph) consisting of the text as a string and a set of text ranges (start/end/type). Input was via hidden text field that's updated as you change blocks and the mapping from x+y screen coordinates (click, arrow up/down across block boundaries) to text field offset is doable by writing a shim for caretPositionFromPoint.

I started with just using contentEditable but you have even more exciting cross browser issues, the output is both inconsistent between browsers and terrible, and you can't do controlled input because updating the editable content DOM at all breaks undo.

With the virtual DOM model, updates are roughly O(n) where n is the size of the virtual DOM. The way to handle very large amounts of input is to window the input data and only vdom render/diff the subset of the data that's currently visible.