Chunking is splitting documents into pieces for embedding and retrieval: by paragraphs, sections, semantic boundaries, or token windows, usually with overlap. C
Splitting documents into pieces for embedding and retrieval, by paragraphs, sections, semantic boundaries, or token windows, usually with overlap. Chunk design determines what retrieval can find: too small loses context, too large dilutes relevance and wastes tokens.
It is the most underestimated lever in RAG: teams often blame models for failures that better chunking would have fixed. Structure-aware splitting, attached metadata, and parent-child schemes routinely improve retrieval more than changing the model does.