Commit Graph
8 Commits
Author SHA1 Message Date
iamdoubz eab7c5a50a Paste an email's text instead of a photo; dates without a year go in the year nearest the scan's month 2026-11-09 20:59:25 -06:00
iamdoubz 9b1a0a99f2 Add Ollama top P, top K, thinking level, context size and keep-alive settings 2026-11-06 11:00:45 -06:00
iamdoubz 725e93c242 Make phase0/extract.py a thin wrapper around the package 2026-11-05 16:08:30 -06:00
iamdoubz a3a87d1b74 Move the grid row finder into the package 2026-11-05 16:08:29 -06:00
iamdoubz 3b0fb016ed Add EXTRACT_IGNORE list for labels that aren't food or events
Grade ranges like K-8 are always dropped; other labels (such as BHS) are now listed in EXTRACT_IGNORE instead of being built in. Matches whole entries in any case, in menu items and event titles.
2026-11-05 14:59:24 -06:00
iamdoubz 2347c2810d Add Phase 0 scoring script
Scores a run against the golden set: no-edit rate, recall, flagged errors, per-field accuracy, and menu days.
2026-11-05 14:50:26 -06:00
iamdoubz 2c449f0ac4 Add Phase 0 extraction script
Reads golden pages with a local Ollama model (default nemotron-3.5-bolt:e4b) or Claude. Grid pages are read one week row at a time and dated in code; EXTRACT_DOUBLE_READ=1 reads twice and flags disagreements.
2026-11-05 14:50:26 -06:00
iamdoubz f3eb67849c Add week-row finder for calendar and menu grids
Deskews photos, finds the grid's horizontal lines, skips banners and long arrows, and crops the weekday headings so each row can be read on its own.
2026-11-05 14:50:25 -06:00