Paste text. See four chunking strategies side by side. Overlap regions are highlighted.

Source text

The art of chunking text for retrieval-augmented generation is deceptively important. A chunk that is too small loses context; a chunk that is too large dilutes relevance and wastes tokens.

Fixed-size chunking is the simplest approach. You pick a token count, then slice your text into pieces of that length. It works, but it commonly cuts mid-sentence and mid-thought, leaving the retrieval model to guess where one idea ends and another begins.

Recursive chunking starts with larger separators (double newlines, then single newlines, then sentences, then words) and only falls through to the next level if a chunk is still too big. This preserves more structure but is harder to reason about for novel text shapes.

Sentence-based chunking respects natural language boundaries. It groups sentences until a target size is reached. Good for prose, less useful for code or tabular data.

Paragraph-based chunking is the coarsest of the four. It treats double-newline-separated blocks as atomic units. Great for well-structured documents, weak when paragraphs are very long or very short.

Overlap, sometimes called stride, addresses the boundary problem by including the tail of the previous chunk at the start of the next. Typical values are 10 to 20 percent of chunk size. More overlap means more redundancy in your vector store but better context preservation across chunk boundaries.

Target chunk size (tokens)

Overlap (tokens)

Chars/token ratio

Reading the output. Highlighted regions are bytes shared with the adjacent chunk (overlap). Stats show: chunk count, mean/min/max tokens, total tokens (corpus size + overlap inflation). For a fair comparison, watch how each strategy handles sentence boundaries near your target size.

Picking a strategy

Source type	Suggested	Why
Long-form prose	Recursive or sentence	Preserves sentence-level coherence.
Technical docs with headings	Paragraph	Headings act as natural boundaries.
Source code	AST-based (not shown)	Splitting on tokens breaks scope. None of the four shown handle code well.
Mixed-quality scrapes	Fixed with high overlap	Bad input deserves redundant chunks.