Sentence-Window Chunking

Sentence-Window Retrieval (Small-to-Big Sentence Window) is an advanced RAG chunking architecture that cleanly decouples the 'Retrieval Unit' from the 'Generation Context Unit'; during indexing, documents are parsed into standalone single sentences (or tiny 2-3 sentence chunks) and embedded into the vector store; metadata retains pointers to the surrounding context window (e.g., surrounding 3 preceding and 3 succeeding sentences); upon retrieving the high-precision sentence match, the system dynamically expands and loads the wider sentence window into LLM context.

Domain-specific question-answering requiring pinpoint factual retrieval combined with comprehensive surrounding paragraph context for generation.

Large chunks (1024 tokens) dilute embedding representations across broad background noise, reducing cosine similarity for precise queries; sentence-level indexing ensures maximum similarity specificity, while surrounding window expansion provides rich context for coherent generation.

How does Sentence-Window resolve the fundamental contradiction between embedding information dilution and generation context deficiency?

When retrieval hits multiple adjacent sentences in the same paragraph, how to execute window merging to prevent duplicate context injection?

Analyze the trade-offs of `window_size` parameter (preceding/succeeding $k$ sentences) on LLM token budget vs answer completeness?

Quantify the vector count inflation and storage overhead of Sentence-Window compared to traditional overlapping paragraph chunks?

How to decouple architecture by offloading expanded window texts to Redis/SQL while vector DB only indexes single sentences?