Workflow Boundaries

figrecover is designed around auditable approximate recovery, not silent fully automatic reconstruction.

Current durable boundary:

  • users supply a cropped chart image;

  • users supply plot-frame pixel bounds and axis data bounds;

  • deterministic extractors recover coloured line, scatter, or bar series;

  • filled area charts can be treated as line extraction using the top or bottom edge of a coloured region;

  • outputs include recovered values, pixel coordinates, confidence, and diagnostics.

Digitization records can carry source provenance, including image ID, document ID, figure ID, figure label, source PDF, source page, crop bounding box, extraction tool, tool version, and extractor settings. CSV exports can include these provenance fields when combining results across documents.

Phase 3 document preparation renders PDF pages and records candidate figure regions before chart digitization happens. This layer remains deterministic: PDF rendering uses PyMuPDF behind the optional pdf extra, and figure candidates are stored as explicit manifest entries.

The first supported document-preparation workflow is:

  1. Render selected PDF pages to ignored local image artifacts.

  2. Create or load a JSONL figure candidate manifest.

  3. Crop manifest candidates into figure images.

  4. Pass prepared crops to the calibrated extraction core.

Candidate manifests can be created manually or by parser adapters. The first adapter boundary is intentionally narrow: adapters emit canonical FigureCandidate records. The PyMuPDF image-block adapter can identify embedded PDF images and map their page-space bounding boxes onto rendered-page pixel coordinates. These are candidates, not guaranteed charts.

Figure candidate manifests are auditable work queues, not proof that a crop contains a recoverable chart. Each entry should preserve source document, page number, bounding box, candidate source, confidence, image paths, and any caption text available from a parser or manual review.

Current diagnostic codes include:

  • no_pixels_matched when a series colour is not found;

  • matched_pixels_clipped_to_plot when matching pixels fall outside the calibrated plot frame;

  • low_confidence_extraction when an extractor produces sparse or empty results;

  • ambiguous_components_filtered when small scatter components or bar runs are filtered;

  • calibration_plot_bounds_missing when extraction falls back to the full image because no plot-frame bounds were supplied.

Optional layers:

  • heavier document parser adapters for candidate figures and captions;

  • batch corpus orchestration with resumable workers;

  • local VLM chart metadata proposals;

  • FEMIC/FHOPS export adapters.

QA overlays, quality metrics, and review manifests are now available for human-in-the-loop inspection after digitization. They are intentionally separate from extraction so downstream systems can require explicit acceptance before using recovered tables as model inputs.

Local VLM assistance can propose chart type, labels, legends, series colours, calibration hints, and warnings. VLM outputs are proposals. They should help reduce manual work, but calibrated geometry remains the preferred numeric source whenever it is feasible. VLM records preserve prompt, backend, model, raw response, parsed proposal, and diagnostic metadata so downstream review can audit how a proposal was produced.