Chunk retrieval
Updated 2026-09-15
Chunk retrieval is the step where a system splits documents into passages and retrieves the individual passages that match a query. The unit an engine works with is therefore a block of your page, not the page as a whole.
Why the unit matters
A passage is judged on its own: if it answers the query, names its subject and carries its context, it can be kept and quoted. If it says "as mentioned above, the second option is usually cheaper", it is unusable alone and gets dropped even when the page as a whole is the best source on the web. Writing for chunk retrieval means every block that could answer a question repeats the entity name, states its claim fully, and avoids depending on the paragraph before it.
Illustrative rewrite (example wording)
Dependent: "It costs roughly twice that at the next tier." Self-contained: "Otterly.AI's Standard tier costs $189/mo for 100 prompts, compared with $29/mo for 15 prompts on Lite." Same fact, but only the second survives being retrieved on its own.
Common mistakes
- Long comparison prose where the comparison only exists across paragraphs.
- Tables without a plain-language sentence stating the takeaway — table structure is frequently lost in retrieval.
- Splitting one answer across a heading and three paragraphs so that no single chunk is quotable.
Frequently asked questions
- How long should a chunk be?
- Write paragraphs that make sense alone rather than chasing a token count: engines chunk differently and change their splitting, so self-containedness is the property that survives, not a specific length.
- Do headings help retrieval?
- They help by scoping the passage that follows and by giving the model a label for it — a heading that states the question and a first paragraph that answers it is the most retrievable pattern available.
