Generated from folio-assistant-core/skills/deep-document-research.md — do not edit here.

✎ Edit this page’s source

Deep document research

Renders methodologies/doc-researcher.md — Dong et al., arXiv:2510.21603v1 — as something an agent here can run. Read the methodology node for what the paper claims and what it measured; this is the part you act on.

It assumes a corpus, and does not build one. library-ingestion acquires and files; ingest-* derives structure; the L1 completeness gate says whether an entry is fit to use. This begins after all three.

The loop, and the only two ways out

plan → [ search → refine → evaluate ] → report
              ↑                 |
              └── σ < τ and t < T_max

Stop when σ ≥ τ, or when t = T_max. Both, always.

• Threshold alone never terminates on a question the corpus cannot answer, and a corpus that cannot answer a question is the normal case rather than the exception.
• Cap alone stops an answerable question early, and the answer it returns looks exactly like a complete one.

A loop carrying one of them is a different method with a failure mode the paper’s does not have. If you implement only one, you have not implemented this.

Say which one you stopped on. An answer that hit T_max is an answer with known gaps; one that reached σ ≥ τ is not. A report that does not distinguish them hands the reader a confidence nobody established — the same shape as a QA sweep reporting clean because it could not look.

The four roles are four QUESTIONS, not four programs

	asks
plan	which documents could possibly bear on this, at what granularity, and what are the sub-questions?
search	what does the corpus return for this sub-question at that granularity?
refine	of what came back, what is actually relevant and not a duplicate?
report	what does the accumulated evidence say, and where can a reader check each claim?

One actor may perform all four. The deep-researcher role is one swimlane because the accountability is one: somebody stands behind the synthesis. Do not read the paper’s four agents as four roles — this repository’s role model is explicit that a role is an accountability, not a step.

Choose the granularity per question, and say you chose

The method’s substance is that a corpus parsed at several granularities lets a LATER step pick. A broad question is answered from summaries; a specific one from chunks. Picking once, in the pipeline, answers every future question the same way — which is the condition this method exists to escape.

This checkout is not there yet. l1-blocks.ts produces blocks and sections exist; there is no summary level and no per-query choice. Until there is, say which granularity you searched rather than implying the corpus offered one. That is the honest form of a capability that is missing.

Three things NOT to adopt with it

1. Machine descriptions as fact. The paper converts figures to text with a VLM and uses the output. Here an agent’s description is a draft and only a person confirms it (apply-image-verdicts.ts, schemas/narrative.ts): “an uncited narrative is indistinguishable from a transcription of the source.” Search over drafts if you must — and never report a draft as what the figure shows.
2. A sufficiency function you invented. The paper gives σ a symbol and a threshold and never says how it is computed. If you need one, propose it and have it adjudicated; do not bury a judgement in a number.
3. The paper’s numbers. 50.6% and “3.4× baselines” were measured on the authors’ own benchmark over their own 304 documents. They say nothing about this corpus, and quoting them here would be evidence for a claim nobody tested.

Say what you did not find

The one failure this method makes easy: an iterative search that ends quietly looks identical to one that succeeded. Report the sub-questions that returned nothing, and whether that means not in the corpus or not found by this search. Those are different, and only the first is a fact about the folio.

See also

• doc-researcher — the method, what it measured, and where this rendering stops.
• library-ingestion, document-intake — how the corpus this reads gets there.
• literature-search — searching for sources not yet held, which is the other direction.

Processes that run this skill

This skill has its own process: Deep document research.

process	step(s) that name it
Deep document research	Filter the corpus, choose the granularity, decompose the question; Retrieve for each sub-question; Deduplicate, and keep what is actually relevant; Synthesise, cite to a location, and say what was not found