Parent-Child Chunking

Parent-Child Chunking (Hierarchical Document Chunking) is a two-tier (or multi-tier) tree-structured RAG chunking paradigm; a raw document is first partitioned into large, context-rich segments (Parent Chunks, e.g., 1000-2000 tokens per section), and each parent chunk is subdivided into multiple smaller segments (Child Chunks, e.g., 100-200 tokens); only child chunks are vector-embedded with metadata pointers (`parent_id`); during search, high-precision child vector hits automatically resolve to their encompassing parent document for LLM generation.

Hierarchically organized technical manuals, API documentation, legal policy codes, and multi-chapter corporate knowledge bases.

Small chunks offer superior semantic retrieval accuracy but starve LLM generation of surrounding context; large chunks preserve context but dilute vector similarity; Parent-Child unifies fine-grained retrieval with coarse-grained rich generation context.

Compare Parent-Child Chunking vs Sentence-Window Chunking in context structure and multi-level inheritance?

When multiple child chunks hit the same parent document, how to deduplicate and score-weight the returned parent?

How to leverage native Markdown heading trees (H1->H2->H3) to automatically build multi-tier Parent-Child tree node hierarchies?

Explain the architecture of storing parent documents in a DocStore (Mongo/Redis) synchronized with child vector indexes?

How to ensure child chunk boundaries do not cut through table rows or fenced code blocks within parent chunks?