Jamie Craane personal blog

My views, thoughts and experiences about Fullstack software development.

Chunk Overlap Solves the Right Problem at the Wrong Time

Split on structure at ingest, and buy the context back at retrieval

Standard RAG advice says to overlap your chunks by 10 to 20 percent. My knowledge base uses zero overlap and expands a matched chunk to its enclosing section at query time instead. Here is why, and where it actually breaks down.