Stop letting one model both search and read: splitting the jobs lifts F1 by 11 points and beats published RAG and GraphRAG on multi-hop QA

Open-domain multimodal document QA no longer asks one model to hold both retrieval and reading tools and schedule itself; the work splits the jobs into a search-only Locator agent, a code-scheduled bridge judge, an evidence pool that merges every search's ranking, a Deep Reader that only reads text, images and tables, and a synthesizer with a format checker. Earlier retrieval-augmented agents gave one model both kinds of tools and let it decide when to search and when to read. On the full open-domain M3DocVQA benchmark, a self-hosted Qwen3.6-35B-A3B (3B active) reaches F1 47.5, +11.0 over the published M3DocRAG system and +7.1 over one-shot retrieve-and-answer on the same index and model; with DeepSeek-V4-Flash it reaches 53.5.

The gain can be broken down by role: the Deep Reader is worth 3.8, 9.8 and 0 points on the three corpora, tracking how visual each corpus is. Merging both agents into one with all tools and twice the rounds loses 1.9 F1, because it keeps searching instead of reading. Per-job ablations on the full benchmarks give those numbers; a round-budget sweep shows a step at four Locator rounds and a plateau beyond; query latency scales with model calls rather than corpus size.

The pipeline targets open-domain question answering over large heterogeneous document collections, covering multimodal documents, multi-hop QA and single-document long-document benchmarks, and the three indexed benchmarks are served as read-only knowledge bases with free registration. Prompts and runner code are released, so a reader can reproduce the pipeline on similar document collections.

Per-job ablations show the Deep Reader contributes 3.8, 9.8 and 0 points on the three corpora, so the gain tracks how visual a corpus is; the round-budget sweep shows a step at four Locator rounds and a plateau beyond; query latency scales with model calls rather than corpus size. These are observations under the current setting, and the source does not state whether they hold for other corpora or models.