Sentence window retrieval finds relevant sentences then expands to surrounding context for generation. Mechanism: Index individual sentences, retrieve matching sentences, expand each to include N sentences before/after, provide expanded windows to LLM. Why sentences?: Maximum retrieval precision - no irrelevant content in retrieved unit. But single sentences often lack context for understanding. Window expansion: Retrieved sentence + K previous + K following sentences. Typical K = 2-5 depending on document type. Implementation: Store sentence index with document position, retrieve top-k sentences, fetch surrounding context from position, merge overlapping windows. Comparison to parent document: More flexible window size, adapts to local context needs, but requires position tracking. Best for: Documents where key information is localized, QA over factual content, precise citation needs. Trade-offs: Index size (many more vectors), position metadata storage, merge complexity. Variations: Variable window based on paragraph boundaries, semantic window (expand to related sentences via embedding similarity). Clean separation of retrieval precision and context completeness.