Document Splitting

Splitting (chunking) cuts long documents into pieces suitable for retrieval: embedding models have limited input, and the goal of retrieval is to inject only the relevant content into context. Three strategies: fixed size with overlap (simple, but cuts structure); recursive by natural boundaries (standard choice); semantic (cut where the similarity between adjacent sentences drops). Typical size 256-1024 tokens, overlap 10-20%. An inherent downside: the chunk loses connection to the original context – this is healed by context-aware retrieval.

Related: RAG, [Context-aware retrieval], [Dense embedding]