Fixed-size chunking is breaking my document context.

I’m building a RAG app for long technical manuals, but basic 512-token chunking is driving me crazy. It keeps slicing critical instructions right down the middle, so the LLM only gets half the picture and guesses the rest.

I know people talk about semantic or layout-aware chunking, but does it actually solve this in production? Or is parent-document retrieval the real way to go here? Would love to hear what worked for you guys.