Why it matters

Small chunks retrieve precisely; large chunks give LLM context. Best of both.

Advertisement

The architecture

Index child chunks (~200 tokens).

Store parent linkage.

Retrieve child → return parent.

Parent-child chunksSmall child chunksprecise retrievalParent linkin metadataReturn parentfor LLM contextAlso called 'small-to-big'; LangChain has built-in support
Parent-child.
Advertisement

How it works end to end

Child: 200 tokens for retrieval.

Parent: 800-2000 tokens for generation.

Metadata links child → parent.