What Parent-Child Retrieval Actually Trades Off

A team's retrieval eval shows their vector search is finding the genuinely correct section for almost every query — precision on "is the retrieved chunk topically right" is high. But answers are still frequently incomplete, because each indexed chunk is only ~150 tokens and often ends right before the sentence that would have completed the answer. They're considering switching to parent-child retrieval. What does this pattern actually change?

A. It indexes small chunks for precise vector matching, but returns the larger surrounding "parent" section to the LLM at generation time, so matching stays precise while the model gets enough context to answer. B. It replaces the cross-encoder re-ranking pass with a two-level retrieval model that scores chunks at multiple granularities. C. It fixes retrieval accuracy by embedding each document at several resolutions and averaging the similarity scores. D. It removes the need for metadata filtering by nesting chunks inside a document hierarchy instead.