Sentence window
A RAG strategy that indexes and matches on individual sentences (high precision in the search), but then gives the model a window of surrounding sentences around the hit (rich context). It resolves the conflict between sharp matching and sufficient context.
Mechanic: the document is split into individual sentences that are embedded and searched (a small, precise unit matches the query well). When a sentence hits, not just it is fetched but also, say, the three sentences before and after — the window is sent to the model so it has context to answer from. Win: you get the search precision of small chunks AND the readability/context of large ones, without compromising. Related to parent-document retrieval (match small, return the larger parent). One of the most used "advanced RAG" techniques, popularized by LlamaIndex. Related to parent-document retriever and chunking.