Consecutive captures whose printed PAGE X OF Y markers advance under the same total are grouped into one multi-page PDF, with duplicate pages keeping the better read. Naming gains doc-title/party components and a scan-date fallback so nothing lands in UNSORTED. Vision LLM default switches to minicpm-v (faster, fewer false blanks on dense mono pages).
- vision/refine.py: tighten crops to the paper band (removes mat
margins and hands beside receipts) and inpaint border-connected
skin regions so fingers disappear from output
- llm/vision.py: identify documents with a local Ollama vision model
(qwen2.5vl); extracts vendor/date/total/form code and flags quality
issues (fingers, blur, glare); falls back to Tesseract when down
- pipeline: drop blank pages, dedupe consecutive captures of the same
document, record refine/LLM fields in report and export summary
- ocr/: Tesseract-based vendor/date/total extraction. PSM 6 (single
uniform text block) instead of the default fixes receipts where
item-name/price columns were otherwise split into separate blocks and
dropped. Vendor line picked by median (not max) word height within the
top fraction of the crop, breaking ties by topmost line - max() was
fooled by single descender/ascender glyphs (commas, parens) inflating
one word's bounding box.
- naming/: merge OCR metadata into YYYY-MM-DD_vendor.pdf filenames,
degrading gracefully to UNSORTED_<timestamp>.pdf; per-run dedup.
- pdf/: img2pdf-based multi-page-capable assembly (images_to_pdf).
- pipeline.run_export(): OCR + name + PDF each detected crop from a
detect() report, writes pdf/ and export_summary.csv.
- New `paperpod export <video>` CLI command.
- Verified against synthetic Home Depot/Metro/Petro-Canada receipts:
3/3 correct vendor, date, and total after tuning.
Known limitation (documented in README): no page-flip/multi-page
grouping yet (that's events/, still unbuilt) - every detected document
becomes its own single-page PDF.
ilia