T4-R2 in five minutes
A chunk is a piece you store and retrieve.
Too large and the window fills with neighbors. Too small and the fact splits.
Overlap is a choice you record.
Headings often make better boundaries than a fixed character count.
There is no universal chunk size. Test it on your notes.
A policy split across two chunks can retrieve only the exception. Keep the rule and the exception in the same chunk or retrieve both on purpose.
Sources and go deeper
- Lewis and co-authors, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, 2020 paper last checked 2026-10
- Retrieval-augmented generation — Wikipedia wikipedia last checked 2026-10