In practice, chunk overlap is typically set to 10%–20% of the chunk size; too small will cause cross-chunk semantic loss, while too large will introduce noise and inflate storage costs.