What is semantic chunking?
Strong answer: chunking by topic or meaning boundaries rather than only size or separators.
Follow-up: when is it better than recursive chunking?
Interview Questions: Semantic and Hierarchical Chunking
Practice explaining advanced chunking without over-selling it.
What is hierarchical chunking?
Strong answer: storing relationships between document, parent sections, and child chunks so retrieval can expand context.
Follow-up: what metadata is required?
What is parent-child retrieval?
Strong answer: retrieve precise child chunks, then fetch parent context when needed.
Follow-up: what can go wrong with parent expansion?
When is simple chunking enough?
Strong answer: when fixed or recursive chunking passes golden retrieval, citation, and faithfulness tests.
Follow-up: why avoid over-engineering?
How do you evaluate hierarchy?
Strong answer: measure child recall, parent expansion precision, citation support, context utilization, and answer faithfulness separately.
Follow-up: what metric catches parent flooding?
What should citations point to?
Strong answer: exact child evidence where possible, while parent context supports interpretation.
Follow-up: why are parent-only citations weak?
What is the access-control risk?
Strong answer: parent expansion can pull restricted context unless both parent and child pass permission filters.
Follow-up: where should filters run?
How do you version advanced chunking?
Strong answer: store chunker version, hierarchy version, expansion rule version, index version, parent id, and child id.
Follow-up: what breaks if this is missing?