Citaim
AI & agents

Chunking & passage retrieval

Also known as: passage indexing, text chunks

Retrieval systems do not read your page — they read paragraph-sized slices of it, and each slice has to make sense alone.

Last revised 2026-09-07. Also available as markdown — request this URL with Accept: text/markdown.

In plain English

Before a page can be retrieved it is cut into chunks: a few hundred words each, roughly along headings and paragraphs. What gets compared to a question, and what ends up quoted, is a chunk — not your page.

That single fact explains most of the difference between pages that get quoted and pages that do not. A paragraph beginning "This is why it matters" is meaningless once separated from the paragraph above it, and separated is exactly how the machine will see it.

How to fix / set it up

  1. Reread your key pages one paragraph at a time, with the rest covered. Any paragraph that stops making sense needs its subject restored.
  2. Replace pronoun openings ("It", "This", "They") with the actual noun in paragraphs that carry facts.
  3. Put the answer sentence first and the reasoning after it.
  4. Repeat units, currencies and dates inside the sentence that carries the number.

The technical detail

Write self-contained paragraphs: name the subject rather than referring back to it, keep one idea per chunk, and put the conclusion at the top of the passage rather than at the end of the section.

Headings do real work here, because they usually survive as the chunk's context. A heading that names the question gives every paragraph beneath it a topic even after slicing.

Tables and lists chunk badly when the meaning lives in the header row far above. Repeat the unit or the subject in the cell where it matters — "£29/month" beats "29".

FAQ

Does this mean short pages are better?

No — long pages are fine and often better, provided each section stands on its own. The enemy is dependency between paragraphs, not length.

How long should a section be?

Aim for a complete thought in roughly 80–200 words. That happens to sit inside typical chunk sizes, which is a useful accident rather than a rule to obsess over.

Related

← All topics in the knowledge base