How literature relationships are built
A six-stage extraction pipeline turns full-text literature into grounded, citation-traceable ingredient-outcome relationships.
1
Ingest - study PDFs, slides, and symposium transcripts are loaded, including scanned documents.
2
Segment - text is broken into overlapping windows so claims spanning paragraph boundaries are still captured.
3
Index - each segment is encoded so relevant passages can be retrieved for ingredient and outcome queries.
4
Extract - the model pulls out ingredient, outcome, direction, confidence, and grounded mechanism summaries.
5
Standardise - varied terminology is mapped to consistent labels so findings can be compared across papers.
6
Aggregate - matching ingredient-outcome claims are merged across sources, with agreement and contradictions surfaced.
Directional and evidence-weighted - claims reflect how many independent studies agree, with contradictions made explicit.
Grounded mechanisms - explanations come from the study text itself, not free-form model speculation.
Full citation traceability - each claim can link back to a source, page, DOI, and year.
Example: "Pea protein -> Expansion: increase in 5 studies, 0 contradictions. Mechanism: higher protein content promotes melt viscosity, driving radial puffing."