I would count lengths and offsets with respect to characters (codepoints). Note that for fixed-length encodings such as "raw", "ascii", or "latin-1", this always means bytes.