Parent-document retriever
A RAG strategy where you match on small "child" chunks (precise search) but return the larger "parent" chunk (or the whole document) they belong to, to the model. Search small, deliver large.
Mechanic: split each document into small chunks for indexing and search, but keep track of which larger parent section (or whole document) each small chunk came from. When a small chunk matches the query, send the parent to the model instead of just the little piece → precise retrieval but full context. Win: avoids both the problem of too-large chunks (blurry matching) and too-small ones (missing context) — the same core idea as sentence window, but on a chunk/document hierarchy instead of sentences. Popular in LangChain ("ParentDocumentRetriever"). A standard technique in more advanced RAG pipelines. Related to sentence window and chunking.