knowledge-extraction

Load the PDF with markitdown to preserve tables and headings, falling back to pdftotext -layout if that fails. Then extract four domains: entities, relations, claims, and evidence, each with provenance pointing to the source sentence and section, and save the result as a KnowledgeItem JSON.

markitdown converts PDFs to Markdown, preserving table structure and heading hierarchy, which makes evidence extraction and section labeling more accurate. pdftotext -layout is the fallback for scanned, encrypted, or oversized PDFs where markitdown produces empty output.

No. It strictly processes one paper at a time. Cross-paper comparison, contradiction detection, and association discovery belong to the downstream association-discovery atom, keeping single-paper extraction free of cross-paper contamination.

Every claim is labeled with one of three confidence levels: stated for explicit assertions, suggested for implied findings, and speculated for author hypotheses. This prevents author speculation from being recorded as established literature fact.

For papers with no abstract, the skill records findings as 'Abstract not available' with methodology set to unspecified instead of failing. If both markitdown and pdftotext yield no text, the paper is flagged as unextractable rather than fabricating content.

It reconstructs a published paper's research design as sparse or dense idea documents plus an experimental log, excluding actual results. This creates benchmark cases for evaluating how research pipelines respond to underspecified inputs.