The parent-child retrieval design is helpful for prospectus tables, but parent expansion should preserve the child's exact supporting row as well as the surrounding section. Otherwise a relevant fee row can be buried among several categories again when the larger parent reaches generation.
I would keep a row identifier and page location with each child hit, then show which fields the answer used. A test with two adjacent fee categories and one amount split across a page boundary would check the extraction repair and the expansion together. That makes the citation inspectable at the same granularity as the number being answered.
Ahmet Özel
AI Engineer. Computer Vision, RAG and LLM agents.
The parent-child retrieval design is helpful for prospectus tables, but parent expansion should preserve the child's exact supporting row as well as the surrounding section. Otherwise a relevant fee row can be buried among several categories again when the larger parent reaches generation.
I would keep a row identifier and page location with each child hit, then show which fields the answer used. A test with two adjacent fee categories and one amount split across a page boundary would check the extraction repair and the expansion together. That makes the citation inspectable at the same granularity as the number being answered.