Statement Of Need

Scientific and professional technical documents often publish figures without the original source tables used to generate them. Those figures may still contain useful quantitative evidence for modelling, review, meta-analysis, or historical reconstruction, but copying approximate values by hand is slow, inconsistent, and hard to audit.

figrecover addresses this gap by providing an open Python workflow for recovering approximate tabular data from chart images while preserving provenance, calibration decisions, diagnostics, and review status. The package is designed for power users who need repeatable recovery workflows across technical PDFs, not for silent fully automatic reconstruction.

The core design principle is conservative: deterministic algorithms should own numeric extraction whenever they can work. Optional local VLMs can propose chart metadata, legend entries, axis labels, series colours, and calibration hints, but those proposals are not authoritative recovered data by default.

The intended use cases include:

• recovering modelling metadata from historical technical reports;

• extracting approximate line, scatter, bar, or simple area-chart values from prepared figure crops;

• reviewing recovered values with overlays and diagnostics before downstream use;

• processing larger PDF corpora into auditable page, crop, result, and review artifacts;

• exporting accepted recovered values into modelling workflows such as FEMIC and FHOPS.

figrecover should be understood as data recovery and evidence management software. It does not recreate hidden source tables perfectly, and it does not make arbitrary PDF-to-table recovery automatic. Its purpose is to make the recoverable portion of figure-derived data explicit, reviewable, and reusable.