Chunking

Slicing documents into passage-sized pieces and filing each piece separately so it can be found later. Search hands back pieces, not whole books — chunking is the decision about how big a piece is.

Cut them too small and each one loses the thread of what it was saying. Cut them too big and the answer arrives buried in three paragraphs of unrelated text.

Why it matters: deeply unglamorous, and it decides more about answer quality than almost anything else. If an AI that reads your documents keeps answering from half a thought — right subject, sentence cut off at the knees — the chunking is wrong. That's a fixable setting, not the AI being poor.

Related: Retrieval · Knowledge base