Chunking

Whole documents are too long to retrieve and pass to a model, so they are split into chunks. How you split matters: chunks that cut through an argument or a table lose meaning, and chunks that are too large dilute relevance.

Good chunking follows document structure such as headings, sections and clauses, uses overlap so ideas aren't severed at boundaries, and keeps a link to the parent section so more context can be added at answer time. The right size is found by evaluation, not by default settings.