The stages-not-strategies correction is the one I needed a year ago. The listicles compare them because both answer "how do I split", but layout-aware runs at parse time against the physical page and hierarchical runs after you already have a heading tree, so asking which to pick is a category error.
The condition I would move earliest is whether the heading structure is trustworthy. On PDFs where headings are a visual convention rather than tagged structure, hierarchical chunking inherits whatever the parser guessed, and a wrong parent silently attaches a clause to the wrong section. From the outside that looks exactly like a retrieval miss, so you go tune the retriever. I now eyeball the heading tree on a sample before trusting parent-child retrieval at all.
The stages-not-strategies correction is the one I needed a year ago. The listicles compare them because both answer "how do I split", but layout-aware runs at parse time against the physical page and hierarchical runs after you already have a heading tree, so asking which to pick is a category error.
The condition I would move earliest is whether the heading structure is trustworthy. On PDFs where headings are a visual convention rather than tagged structure, hierarchical chunking inherits whatever the parser guessed, and a wrong parent silently attaches a clause to the wrong section. From the outside that looks exactly like a retrieval miss, so you go tune the retriever. I now eyeball the heading tree on a sample before trusting parent-child retrieval at all.