Lesson 2.1

Chunking strategies

The retrieved passage is what the model sees.

14mAdvanced6.5k students

Overview

A chunk has to make sense alone

The retrieved passage is what the model sees. If a chunk boundary lands mid-argument, the model gets half a thought and answers from half a thought.

Splitting on document structure — headings, sections, paragraphs — beats fixed character counts, because those boundaries were chosen by an author to separate ideas. Fall back to a size limit only inside a section that is too long.

A modest overlap between adjacent chunks stops an answer that straddles a boundary from being lost by both. Too much overlap inflates the index and returns near-duplicates.

In this lesson you will:

  • Split on structure rather than character count
  • Choose a size that keeps a passage self-contained
  • Use overlap to avoid cutting an answer in half

Resources

Previous Lesson
Next Lesson
Chunking strategies — Retrieval-Augmented Generation in Practice — Vertex