Adds a badge/nav header, the four architecture and sequence diagrams from the
handover set, a port table, a screenshot gallery, collapsible gotchas and
troubleshooting sections, and a documentation index.
Corrects two inaccuracies: endpoint resolution tries LAN first and ngrok second
(not the reverse), and the poll loop runs ~4.3 minutes (not 2). Documents three
security facts that were previously unrecorded: seeded passwords reset on every
backend restart, JWT_SECRET falls back to a constant in source, and the LAN leg
is plain HTTP.
Force-added: these paths are gitignored, so they stay ignored for new files
unless added the same way. Committed on request so the working data is not
lost during the migration off this machine.
PFM-Scanner-Ringkasan-Scope.pdf/.odt — 20 pages covering the flow, folder map,
ports, how to run each part and from which directory, how login and upload move
data, and what is built versus what isn't.
src/build_ringkasan.py draws its four diagrams as native shapes (same approach as
build_deck.py), crops them to img/ring-*.png, and builds the document in one run.
Grilled 2026-07-16 with the user; full decision record in
docs/expiry-tracking-plan.md. Core reframe: expiry is captured once per
batch at DO intake (staff-typed on the stock-entry confirmation page, from
the physical packs), so the cashier scan only MATCHES OCR fragments against
the 1-3 known in-stock batch dates instead of free-reading damaged
dot-matrix prints (proven model-capability ceiling, 2026-07-15). Fallback:
auto-FEFO + 'inferred' flag, zero cashier interaction. No cloud, ever.
- docs/expiry-tracking-plan.md: architecture, matching algorithm spec
(resolveExpiryFromEvidence), schema/API deltas, phases 1-3, testing plan
- backend plans §13 (13.1-13.4): matcher util + offline tuning, route
wiring + expiry_source provenance, multi-frame union, dot-matrix
recognizer fine-tune
- root plans §10 (10.1-10.3): cashier fast path, inferred badge +
end-of-day review, burst capture for mounted camera
- stock-feature-plan.md: extension note (batch dropdown becomes the
manual-override path)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Q8TumxFDnyVnfsR3mxPXfX
Split the expiry-date extraction cascade out of classify_ocr_server.py into
config/date_extract.py (pure regex, importable/testable without loading
models). Three behavioral fixes, offline-regressed against all 79 captured
OCR line-sets and sanity-verified live on the two target images:
- Guard the 012/112 month-misrecognition cleanup rules: they fired on
perfectly valid dates too (BB 01122026 = 01/12/2026 matches 0+112+2026)
and mangled them into 7-digit junk that parsed as 00/22/26. Skipped when
the line already contains a valid date. Fixes image 11.
- Exclude store price-tag lines (Printed:.., Rp...) from the keyword-less
stages so a shelf label's print timestamp can't shadow the real date
printed on the package. Fixes image 71 (09/04/2027).
- Validity-gate the lenient stage (day<=31, month<=12, year 2020-2039) so
garbled digit runs return empty instead of junk like 1/3/06 or 11/1/01.
Also: clamp /probe-ocr crop box to image bounds (PIL pads out-of-bounds
crops into a gigapixel canvas -> DecompressionBombError), and update
CLAUDE.md's Graphify section - the global Claude Code skill integration was
installed 2026-07-15 at the user's explicit request.
Full-batch measurement of these fixes (expected 79.7% -> ~80.6%) is still
pending - the run was stopped twice at the user's end; re-run
scripts/accuracy-check-scan.mts next session before building on this.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gr6HH7JrdsXX8AARejQboM
2026-07-15 re-run of the 79-image benchmark: the tiled full-res pass, VL
text-line evidence, and VL SKU retry deployed at end of 2026-07-14 produce
ZERO flips vs the 79.7% baseline (only img1's expiry pred changed to a
lenient-parse garble). Detail dump: product_scan_detail_20260715.json.
Probe endpoint POST :8120/probe-ocr (temporary, remove before production):
OCRs a crop of a container-local image through arbitrary preprocessing
recipes (scale/blur/close/CLAHE/threshold) and alternate readers - VL
layout-parsing with/without layout detection, direct VLM chat on :8118,
lazily-loaded PP-OCRv5 server det/rec. Probing the dot-matrix expiry crops
of img2/img41 with every combination shows none of the stack's models can
read these prints (VL tags them as pictures, VLM hallucinates, v5-server
skips them) - the 21 expiry non-detections are a model capability ceiling,
not a pipeline bug.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gr6HH7JrdsXX8AARejQboM
Accuracy work on the 79-image product-scan validation set (user goal: 90%):
- classify_ocr_server.py: 0/90/180/270-degree expiry-date search (stops at
first hit, 0-degree fallback); classification decoupled onto the upright
image (rotated frames regressed DINOv2 -6pts until this); cross-line date
stitching; tiled full-res OCR pass (defeats the 4000px downscale that
killed small inkjet dates); VL-pipeline expiry fallback with
keyword-anchored anti-hallucination guard; VL text lines merged into
text_lines + VL SKU retry. Visualization endpoints removed entirely
(Visual/Spotting grids - unused by frontend, 3x per-scan GPU cost).
- product-scan.ts: coverage-normalized OCR-evidence re-ranking of DINOv2
top-K (tuned offline: +8/-0 on top-1 misses), re-ranked class mapped to
sku_master by SKU prefix; classifier timeout 90s->240s for fallback paths.
- Frozen benchmark: product-test-images-fixed/ (79 renamed images) +
freeze/seed/build-undetected/capture/experiment scripts; labels trimmed to
the 79 validation entries (training rows kept in .bak-with-training);
5 TRAINED-ON SKUs replaced with fresh held-out photos.
- manual-label-scan page: shows last batch-test AI prediction under every
field by default (new /api/product-scan-results); serves the fixed folder;
fixed total hydration failure via allowedDevOrigins 127.0.0.1.
- Measured (all-79, zero failures): sku/name 87.3%, expiry 64.6%, overall
79.7%. Tiles/VL-evidence/VL-SKU deployed but not yet batch-measured.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gr6HH7JrdsXX8AARejQboM
train_classifier.py's split_dataset() previously shuffled and split
individual image files, letting an augmented copy (photo_aug_2.jpeg) land
in validation while its near-duplicate source stayed in training -
inflating val accuracy with memorization rather than measuring real
generalization. Now groups by source photo (stripping _aug_N) before
shuffling and splitting 80/20.
Also records the in-progress effort to retrain the product classifier
against the full 81-class/2,493-photo foto-kemasan-v2 dataset (up from the
16 classes/118 photos the deployed model was actually trained on) - see
plans/next-enhancements.md task 2.5 and the accompanying iteration-log
entry for the real, currently-observed numbers (DINOv2 index rebuilt:
2493/2493 images; classifier training: in progress, ~32s/epoch observed).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xsxk4ZkDQVVaLUcixDcqb5
Ports the DO-harness's auto-diff-vs-previous-run reporting into
accuracy-check-scan.mts: prints a per-field, per-split (Training/
Validation) delta against the last product_accuracy_history.jsonl entry
and calls out regressions/improvements explicitly, plus classifier
method distribution and average confidence as informational context.
Also adds the real held-out validation photo set into
sources/product-test-images/ (75 photos, one per current SKU class) with
its README documenting the drop-photo -> label -> re-run workflow, so the
harness's Validation Set split actually has images to score against.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xsxk4ZkDQVVaLUcixDcqb5
Adds tasks 1.4-8.5 across existing sections 1-8 (each traced to a specific
file/line, not invented busywork) plus a new section 9 (Stocks Menu &
DO-to-Stock flow) seeded from a user-directed, grilled ad-hoc feature
request - full design doc in docs/stock-feature-plan.md, status: planned,
no code yet.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xsxk4ZkDQVVaLUcixDcqb5
Records the AST-derived code-graph tooling (query/explain/affected commands
against graphify-out/graph.json) so future sessions know it's available
and why it's not wired in as a global skill/hook.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xsxk4ZkDQVVaLUcixDcqb5
The Flutter app is used by Prima Fresh Mart store staff (petugas toko) on
duty at each store to receive/confirm deliveries and log products - not by
the delivery drivers themselves. "Driver" only appears correctly now as a
data field on the DO document (who drove the delivery), never as the app's
operator. Fixes wording across README.md, CLAUDE.md, and
screenshots/v2/WORKFLOW.md; also corrects CLAUDE.md's stale "up to 2 min"
poll-timing note to match the real ~4.3min/130-retry value.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JSgYVUWJTGMk7SZqqHH8xy
- Auth section described a hardcoded admin/password stub; it's now real
JWT + bcrypt on /api/v1/* (verified against the actual route handlers).
- Known Limitations listed the pending-queue and non-transactional item
writes as still-open bugs; both were fixed 2026-07-08 - marked resolved.
- Corrected the poll-loop timing (2s x 130 retries = ~4.3min, not "2 min").
- Noted the confirmation-gate step now in the PUT flow, and that Product
Scan classification is DINOv2-primary (YOLO fallback), single-pass at
upload time - not the separate LAN-dependent call it used to be.
- Linked screenshots/v2/WORKFLOW.md + the new slide deck as the visual
app walkthrough reference.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JSgYVUWJTGMk7SZqqHH8xy
Captures the full DO Scan and Product Scan flows end-to-end against the live
backend (real OCR extraction, real DINOv2 product classification, a genuine
network-timeout failure, and a live validation-gap finding) for use in
presentations. WORKFLOW.md documents every screen/button/algorithm, and
Prima-Mart-Scanner-Workflow.pptx turns it into a 23-slide deck.
Also updates AppConfig's hardcoded LAN IP (192.168.70.4 -> 192.168.100.20)
to match the current dev machine's address, discovered while reproducing the
upload flow against the real backend.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JSgYVUWJTGMk7SZqqHH8xy
Fixes reported from APK field testing: DO/Product scan mode was inconsistent
between the camera drawer and documents screen (now one shared provider,
with an orange/green color cue); unconfirmed scans leaked into history with
placeholder data before the user tapped confirm (backend now gates
GET /documents on a new `confirmed` column, flipped only by PUT); and
Product Scan ran the GPU classifier twice, once at upload and again on
review (now a single pass at upload, persisted and read directly by the
editor). Also removes the unused "Hubungkan ke PO" field and fabricated
PO/SO/DO placeholder values from the Product Scan flow, closes out the
per-document-polling and save-recovery tasks (6.1/6.3), and splits several
touched files to stay under the repo's 256-line guideline.
Full detail in docs/iteration-log.md and backend/docs/iteration-log.md.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Backend (app-pfm-ocr-v2/backend):
- Product/SKU scan feature complete: trained DINOv2 index (118 reference
photos, 16 SKU classes) and YOLO classifier (83.3% top-1 val accuracy),
fixed scripts/install-pipeline.sh (was missing ultralytics/torch), fully
browser-verified end-to-end on /scan-pfm. Mobile m-scan-pfm page cancelled
(Flutter app handles mobile; web UI is desktop-only for pipeline testing).
- Fixed a real data-loss bug: Save Ground Truth (scan-pfm and the DO-flow's
manual-label) was silently writing into the pfm-web-app container's
ephemeral filesystem instead of the host, because /sources wasn't
bind-mounted in docker-compose.yml. Added the mount, recovered an
orphaned entry.
- accounts.password is now bcrypt-hashed (bcryptjs, idempotent migration
in db/init.ts) instead of plaintext; login route compares hashes.
- /api/v1/documents/* (list, PUT, upload) now enforces real 401 auth,
matching what the Flutter client already sends. The "classic" routes
deliberately stay open — they're dev-only web UI with no login flow and
won't exist in production.
- OCR accuracy investigated end-to-end: real baseline is 95.10% overall
(target met; accuracy_report.md was stale at 75.04%, now flagged). Fixed
one genuine parser.ts bug (SO/DO field duplication in the global fallback
regex); remaining gaps are OCR/layout-model limitations, not parser bugs.
- Adopted a standalone copy of the fhanyuh/agents-settings e/n workflow
scoped to backend/ (AGENTS.md Part A/B split, SKILLS.md, plans/, docs/),
independent of the root copy which now covers Flutter only.
- next-implementation.md deleted; content folded into
backend/plans/next-enhancements.md for traceability.
Root:
- Adopted fhanyuh/agents-settings kit (AGENTS.md, SKILLS.md, plans/,
docs/feature-list.md), scoped to the Flutter app only.
- Pending documents queue now persists to Hive (lib/core/storage) instead
of memory-only, surviving an app kill mid-upload.
Removed backend_backup/ (stale Express/Prisma prototype, superseded by
pfm-web-app) and the completed plans/next-enhancement-plan.md checklist.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Documents the 89.4% overall accuracy (up from 82.2% baseline, 95% target),
grounded in the actual logged run in accuracy_history.jsonl rather than a
vague claim, plus which fields the sanitize/triple-check layer rescues over
raw regex. Notes that current gaps are partly a photo-capture SOP floor
(DO paper photographed on top of other documents confuses auto-deskew),
not just a parsing/OCR ceiling.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eRAsLqN9Sg1b9YPz22Lxz
Replaces stale setup instructions (single hardcoded API URL, old migration
story) with the dual-mode ngrok/LAN config, the demo/production compose
override, and a "what to consider" section grounded in real issues hit this
session: the duplicate backend/docker-compose.yml project-name collision,
store_master shipping unseeded, and the debug-signed APK.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eRAsLqN9Sg1b9YPz22Lxz
Reliability/PoC hardening: dedupe uploads by file_hash, surface editor sync
failures instead of a false success SnackBar with a retry-without-re-OCR path,
bound the OCR pipeline fetches with timeouts, share a single ApiClient/Dio
instance app-wide, tune capture JPEG quality, and add an opt-in
docker-compose.demo.yml for a production-mode run ahead of client demos.
Translate all Flutter-side user-facing text (screens, validators, SnackBars,
the printed delivery receipt, and shared API error messages) to Bahasa
Indonesia.
Also includes in-progress OCR parser/accuracy-tuning work from the same
session: table column/unit normalization fixes, store/customer master data,
accuracy history log, and test-image renaming/cleanup.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eRAsLqN9Sg1b9YPz22Lxz
