Separating search from reading improves document question answering

Document QA asks a system to do two unlike things: search broadly for relevant material and stop to inspect a bounded passage, table, or figure.

It remains unclear whether assigning those jobs to separate, tool-limited stages can prevent search from crowding out reading while still recovering missing documents.

KnowForge separates finding evidence from reading it, forcing missing-document checks so a document-QA system can inspect more of what it retrieves.

The paper compares the full KnowForge pipeline with one-shot search-and-answer on the same index, model, and 2,441-question M3DocVQA test. M3DocVQA asks questions over 3,368 PDFs whose evidence can be text, tables, or images; the full pipeline scores 47.5 F1 versus 40.4, a 7.1-point gain.

That comparison links the gain to the changed workflow: the Locator searches, a bridge check seeks a missing hop, and the Deep Reader opens retrieved material before synthesis.

A role control found that a single all-tool agent searched more but read less and scored 1.9 F1 lower, supporting the authors' interpretation that job separation helps keep reading in the loop.

The evidence is bounded to the tested models, multi-round tool protocol, benchmark conditions, model-judge scoring for one transfer test, and one latency setup; it does not establish universal behavior or deployment performance.