Parent-Child Chunks
Also called Parent-Child Chunks
Search on small, precise chunks but hand the model the larger passage they came from.
Think of it like
Using the index of a book to find the exact line, then reading the whole paragraph around it so the sentence makes sense.
Example
A search matches one tight sentence about refund windows, but the system feeds the model the full refund section so it has the surrounding conditions too.
How it actually works
There’s a tension in chunk size: small chunks retrieve precisely but lack context, large chunks give context but retrieve fuzzily. Parent-child splits the difference — index tiny "child" chunks for matching, but store a pointer to the bigger "parent" passage. On a hit, you return the parent. You get sharp retrieval and rich context without choosing one chunk size for both jobs.
For product teams
Answers stop being technically-correct-but-missing-the-caveat, because the model sees the whole relevant passage.
For engineers
Index small child chunks with references to larger parent segments; retrieve on children, return the deduplicated parents to the context window.
Related
- Chunking — A more nuanced take on chunk size.
- Context window — Controls what actually reaches the model.
- Hierarchical Retrieval — One rung of hierarchical retrieval.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome