A RAG application retrieves a relevant heading but misses the paragraph defining its exception. The text was split into tiny chunks without overlap or document structure. Which change should be tested first?

1. Fine-tune the generator immediately without measuring retrieval.
2. Test structure-aware or appropriately larger chunks and overlap against a representative retrieval set.
3. Set every document as one chunk regardless of model context limits.
4. Raise output token limits while keeping the missing evidence out of retrieval.