Chunking splits documents into smaller segments for retrieval. Small chunks can produce specific matches but lose context; large chunks preserve context but can blur several ideas in one embedding. section 3.3.2
A two-stage strategy searches small chunks, then retrieves the larger parent document for context. Chunking is not automatic: short, self-contained items can remain intact as natural chunks. section 3.3.2 section 11.3.2
Choosing granularity for financial content
A 50,000-word Form 10-K may need splitting. An 80-character headline or 250-character summary is already a self-contained unit; splitting it further can fragment meaning without improving retrieval. section 11.3.2
