For structured documents with tables and inherited metadata, agentic search using deterministic operations outperforms embedding-based retrieval by 43 percentage points—the gain comes from the interface design, not from iteration or agent reasoning.
This paper challenges the standard retrieval approach for document-heavy domains like financial reports. Instead of embedding chunks and finding top-k matches, the authors propose READ: an agent that searches documents through deterministic operations (lexical search, structural navigation, span reads) exposed as tools.