Milestones
List view
Week 1 (May 25–31) — Structured, section-labeled, provenance-linked document representation Marker handles PDF-to-structured-content conversion, achieving strong accuracy on scientific papers with multi-column layouts and complex table structures. The work in this phase is building a lightweight adapter layer on top of Marker's JSON block tree output, producing a document representation the extraction chain can actually use. The adapter's core job is threefold: classify each block by section (Methods, Results, Abstract, etc.) using Marker's existing section hierarchy metadata; convert Marker's table representations into a structure that preserves header relationships, row groupings, and footnotes since these are critical for correct trait extraction downstream; and retain figure image data for potential vision-model processing later. Section classification will combine Marker's structural metadata with lightweight keyword matching; this is more reliable than raw text keyword scanning because Marker has already identified block types via layout detection. Table extraction deserves particular attention. Agronomic paper tables vary significantly in structure, and the fidelity of the table representation directly affects trait extraction quality. Initial evaluation during community bonding will characterize where Marker's table handling works well and where additional preprocessing is needed.
Overdue by 3 month(s)•Due by May 31, 2026•0/1 issues closed