Adds a step-through tool for creating 100% correct reference data (ground truth)
with position information. Users scan a page, review each vocabulary entry with
image crops, confirm or correct the OCR text, and save the result as JSON.
Backend: extract_entries_with_boxes() helper + 3 endpoints (extract-with-boxes,
ground-truth save/load). Frontend: GroundTruthPanel component with SVG overlay,
ImageCrop, keyboard shortcuts (Enter/Tab/arrows), and tab navigation in page.tsx.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New OCR method using classical Computer Vision: high-res rendering (432 DPI),
deskew, dewarp, binarization, projection-profile layout analysis, multi-pass
Tesseract OCR with region-specific PSM, and Y-coordinate line alignment.
Includes bugfix for convert_pdf_to_image call (line 869) and 39 unit tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Ollama crashes when two concurrent vision requests hit the 32B model
(compare-ocr + analyze-grid). The grid analysis was redundantly calling
Ollama again even though compare-ocr already extracted all vocabulary.
- compare-ocr now saves vocabulary in session for reuse
- analyze-grid builds grid from session data (no Ollama, instant response)
- Grid button disabled until comparison results are available
- Added export-to-editor functionality
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Restores the full OCR comparison tool with vocab extraction,
green grid overlay, labeling, and LLM comparison features.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Restore the SDK context, types, and component files to the HEAD version
since newer pages (company-profile, import) depend on these API changes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
