A policy assistant retrieves isolated sentences from a 78-page handbook and misses exclusions that live in neighboring sections. Hierarchy-aware chunking: child precision plus parent context The common trap is optimizing for embedding convenience while destroying the document structure that gives clauses their meaning. Parse structure Extract title, heading path, tables, captions, effective dates, and appendix boundaries before chunking. Structure is metadata the retriever and generator can use. If you throw it away, the model must infer authority from partial text. Create child chunks Make searchable child chunks around 250-450 tokens, respecting paragraphs and list boundaries. Child chunks give precise embedding…
Read the full lesson
Sign up free: one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in