Compare visual and textual page-selection methods

Establish the relative performance of visual page retrieval and textual page selection for evidence localization on the LitTraceQA task, particularly in terms of accuracy and cost at scale.

Background

The paper evaluates textual page-selection strategies such as BM25, dense retrieval, reciprocal-rank fusion, deterministic anchors, and showing all pages. It identifies visual page retrieval methods such as ColPali and M3DocRAG as stronger potential baselines, but does not evaluate them because showing every page already achieved 100% gold-page recall on the tested documents. The unresolved comparison concerns whether visual retrieval offers meaningful cost or scalability benefits while preserving evidence-localization accuracy, especially for papers too long to fit within the model context.