Search the child, return the parent.
A port of LangChain’s ParentDocumentRetriever runs here on a 2,184-character page of the Larkspur Cloud handbook. Children are what the index searches; parents are what the model reads. Move both sizes and watch the answer, and the token bill, change.
The last parent stop drops parent_splitter, so the whole page is the parent. Both splitters are recursive with no overlap; the index returns the top 2 children by TF-IDF cosine, as in the M10 baseline.
- Answer whole, children only
- Answer whole, parents returned
- Tokens sent, children only
- Tokens sent, parents
Index map of the page
matched childparent returnedstored, not usedthe answer
Searched: the top 2 children
Cosine similarity, TF-IDF
| Rank | Child | Score | Characters | Parent | Text |
|---|
Returned: what the model reads
Matched child underlined; answer marked
All three golden questions at this setting
| Golden question | Children only | Tokens | Parents | Tokens |
|---|---|---|---|---|
| Answers found |
“Children only” is plain chunking at the child size: the same two matches, sent as they are. Tokens use the four-characters-per-token rule of thumb; the footer gives the mean.