Match chunking strategy to content type — and use hierarchical parent-child for technical docs

FreeAcademy·medium signal

There's no universal chunk size: use fixed-size (256–512 tokens, 10–20% overlap) for prose, semantic chunking (split on embedding drift) for dense technical text, structural chunking that respects markdown/HTML for documentation, and hierarchical parent-child where you embed small chunks for precise matching but return the larger parent for context. Picking per content type, rather than one global setting, is a cheap retrieval-quality win that compounds with re-ranking.