See how a chunking strategy splits your text

Paste text and compare fixed-character, fixed-token, sentence-aware, and paragraph/recursive chunking strategies as coloured blocks, with overlap highlighted and size stats.

Runs entirely in your browser - nothing you paste is ever sent anywhere.

Background: the chunking lesson in /learn and why your RAG pipeline is confidently wrong.

Text to chunk

Retrieval-augmented generation grounds a model's answer in real documents instead of only what it memorized during training. The pipeline usually has three stages: chunking, embedding, and retrieval.

Chunking splits a long document into smaller pieces so each one can be embedded and searched independently. Choose a chunk too large and you waste context on irrelevant text; choose one too small and you lose the surrounding context a fact needs to make sense.

Overlap between consecutive chunks helps with the second problem. If a sentence explaining a term sits right at a chunk boundary, a small overlap means the next chunk still carries enough of that context to be useful on its own, even though it costs some duplicate storage and embedding calls.

There is no single correct chunk size. It depends on the embedding model's context limit, the structure of the source documents, and how precise the retrieval needs to be. Measuring retrieval quality on your own documents beats guessing from a blog post, including this one.

Chunk size: 400 chars

Overlap: 50 chars

highlighted text is shared with the previous chunk (overlap)