fix: green the release gate — tesseract local-only, PP-OCRv5 availability, OCR blockId, reproducible models#3
Merged
Conversation
Land the first of the three-layer post-conversion verification system
(rule diff + SSIM + OCR readback). Introduces public/core/verification/
with shared block fingerprints, field-level diffSemanticDocs, and a
runVerificationStage orchestrator wired after the Repair Engine cycle.
- block-fingerprint.js: lift blockFingerprint/modelFingerprint out of
repair-engine (byte-for-byte identical) + getBlockKey/extractBlockFields
+ ROUND_TRIP_FORMATS single source
- rule-diff.js: diffSemanticDocs -> { identical, blockCounts, changed/
added/removedBlocks, fidelity, overallScore } with minor/major severity
- verification-stage.js: gating + same-format readback + md<->html
cross-format loopback + RULE_DIFF_DRIFT / RULE_DIFF_READBACK_FAILED
- format-registry: write qualityReport.ruleDiff + .verification envelope;
repair-engine roundTripDelta contract unchanged
- scripts/rule-diff-test.js (10 assertions) wired into npm test (21 total)
- guard scripts + MULTI_MODEL_ARCHITECTURE + DEVELOPMENT_TASKS synced
- two specs: P9-C overall + P9-C.1 sub-stage
npm test (21 scripts), git diff --check, release:prepare all pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Land the second of the three-layer post-conversion verification system.
SSIM uses visual-loopback semantics: for visual-preserving inputs
(pdf/png), rasterize the input page and output page and compare
structural similarity, writing qualityReport.ssim.
- ssim.js: self-implemented, zero-dependency SSIM core (rgbaToGrayscale,
resampleGrayscale box resample, computeSSIM windowed mean, compareImages)
- page-image-source.js: pixel-source abstraction (Node throws
VERIFICATION_IMAGE_SOURCE_UNAVAILABLE, browser auto-loads canvas impl,
setPageImageSource for tests); RASTERIZABLE_FORMATS = {pdf, png}
- page-image-source-browser.js: vendor pdfjs + canvas getImageData
- verification-stage: runSsimLayer + async runVerificationStageAsync that
merges the sync rule-diff base with the async SSIM layer; sync
runVerificationStage unchanged (qualityReport.ssim stays null)
- format-registry: extract _runRepairCycle + _assembleQuality; convert()
stays sync (rule-diff), convertAsync() uses async wrap (rule-diff + ssim)
- scripts/ssim-verification-test.js (12 assertions) wired into npm test (22)
- guards + MULTI_MODEL_ARCHITECTURE + DEVELOPMENT_TASKS synced
Rendering is stub-only this round (Node has no canvas); real PDF/PNG
fixtures and browser end-to-end deferred. No new npm deps.
npm test (22 scripts), git diff --check, release:prepare all pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Land the third and final layer of the post-conversion verification system.
OCR readback rasterizes the output (PDF) and re-reads it with the OCR
engine, comparing recognized text against the original SemanticDoc text
via a character-level multiset similarity, writing qualityReport.ocrReadback.
- ocr-readback.js: compareText (char-level multiset recall/precision/f1,
robust for CJK + OCR noise) + normalizeText + extractModelText +
runOcrReadbackLayer; reuses registered ocr-text engine + OCR pdf rasterizer
- verification-stage: runVerificationStageAsync dynamic-imports the readback
layer (keeps OCR off the sync convert path) and merges it into the envelope
- format-registry: _assembleQuality adds ocrReadback (null on sync path)
- scripts/ocr-readback-test.js (13 assertions) wired into npm test (23 total)
- guards + MULTI_MODEL_ARCHITECTURE + DEVELOPMENT_TASKS synced
Three layers (rule-diff + ssim + ocr-readback) now write a unified
qualityReport.{ruleDiff,ssim,ocrReadback} + verification envelope.
Rendering/OCR are stub-only this round; no new npm deps.
npm test (23 scripts), git diff --check, release:prepare all pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a programmatic sample-corpus generator to stress-test conversion, layout, and the three-layer verification across every supported format at varied sizes (large tier >= 3MB). Binaries stay git-ignored per the existing programmatic-fixture policy. - scripts/lib/sample-content.js: deterministic complex content builders (md/html/json/xml/csv/txt) with headings, nested lists, task items, aligned tables, multi-lang code, nested quotes, footnotes, images, CJK/RTL/emoji/entities; SIZE_TIERS + buildToTargetBytes - scripts/lib/png-encode.js: minimal node:zlib PNG encoder (no new deps) - scripts/generate-samples.js: emits md/html/txt/json/xml/csv natively, docx/pptx/epub/pdf/xlsx via project writers, png via encoder; writes MANIFEST.json with coverageGaps for doc/ofd (no writer); --tiers/--out - scripts/sample-corpus-test.js: fast in-memory round-trip gate (24th npm test) — no 3MB writes - samples:generate npm script; samples/generated/ git-ignored - samples/fixtures/README + DEVELOPMENT_TASKS documented Verified large tier: text formats >=3MB, pdf 19MB, docx 13.6MB, xlsx 16MB, epub 4.9MB. npm test (24 scripts), git diff --check, release:prepare pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…9-C.4)
The P9-C.1/2/3 layers compute qualityReport.{ruleDiff,ssim,ocrReadback}
plus a verification envelope, but transformContent discarded result.quality
and the old bottom drawer was already removed — so the core "post-conversion
verification" differentiator was invisible. Surface it.
- index.html: new collapsible #verificationReportPanel inside the output
panel (auto-repair verdict + rule-diff / SSIM / OCR-readback rows + warning
counts + active-layer badge); does not revive the removed drawer ids
- app.js: renderVerificationReport(quality) (textContent only, no innerHTML),
captures result.quality into currentConversionQuality, renders on both
text and binary paths, clears on reset
- styles.css: .verification-report/.verification-row/.verification-badge,
data-state ok/drift/skip coloring via :has() left border
- browser-smoke-test: assert #verificationReportPanel present; removed-drawer
negative assertions still hold
Display-only; no change to conversion core / verification-stage. Skipped
layers show their reason (honest: not-triggered != failed). No new deps.
npm test (24 scripts), git diff --check, release:prepare all pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
landing-view.js imports getKnownInputFormats from browser-transformer.js, but it was never re-exported (only defined in format-registry.js). In the browser this is a module-load SyntaxError that takes down landing-view.js and app.js, leaving the landing page blank below the header. Node-side tests never caught it because browser-smoke-test only asserted static HTML strings, not that the module graph loads. - browser-transformer.js: import + re-export getKnownInputFormats - browser-smoke-test.js: load browser-transformer/router/landing-view module graphs and assert key re-exports exist, so a missing re-export fails CI instead of silently blanking the UI npm test (24 scripts) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Research showed the roadmap's named advanced-OCR targets (PaddleOCR-VL / MinerU) are not embeddable under the project's constraints (browser/Tauri local, no cloud, 30-80MB default bundle, no Python runtime): the VLM has no mature ONNX/WebGPU path (~500MB + 1-2GB VRAM or vLLM), MinerU is a Python/vLLM tool. Per user confirmation, the built-in advanced-OCR target is now PP-OCRv5 (ONNX Runtime + WebGPU, WASM fallback); VLMs are marked as far-term/external. - specs: 2026-05-29-p9d-advanced-ocr-research.md (with sources) + 2026-05-29-p9d1-paddle-ocr-skeleton-design.md - paddle-ocr-engine.js: paddleOcrEngine (id paddleocr-v5, ocr-text/ocr-layout) implements OCREngine; isAvailable() false in Node; recognize() three-stage rejection (vendor-not-ready / model-missing / runtime-not-wired); markPaddleOcrVendorReady. No onnxruntime, no real inference this round. - paddle-ocr-bootstrap.js: registers engine (after tesseract) + PP-OCRv5 ONNX ModelManifest (engine paddleocr, int8, det/cls/rec) as not-downloaded - browser-transformer: import bootstrap + export paddle API - ocr-baseline-test: pickForTask fallback set + 35th block (paddle skeleton) - guards: local-security ALLOWED/STRICT; direction-test PP-OCRv5/ONNX/WebGPU - direction docs reframed: PP-OCRv5 built-in, VLM far-term/external (kept "default bundle excludes GB-scale models") Skeleton-first (mirrors P9-A.2 tesseract). npm test (24), git diff --check, release:prepare all pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Wire the ONNX Runtime layer for PP-OCRv5 advanced OCR, mirroring the tesseract P9-A.2 vendor+skeleton stage. Real det/cls/rec inference pipeline and CTC decode are deferred to P9-D.2.b (needs real models + dictionaries). - scripts/sync-onnxruntime-vendor.js: sync ort*.mjs + *.wasm from node_modules/onnxruntime-web/dist to public/vendor/onnxruntime/; exit 0 if the optionalDependency is absent (does not block install/release) - paddle-ocr-runtime.js: loadOnnxRuntime (same-origin vendor dynamic import, sets ort.env.wasm.wasmPaths same-origin, throws OCR_VENDOR_LOAD_FAILED in Node), pickExecutionProviders (navigator.gpu -> [webgpu,wasm] else [wasm]), createOcrSession/disposeOcrSession skeleton, PADDLE_VENDOR_PATHS - paddle-ocr-engine.recognize: third stage now loads the runtime; in a browser with vendor+models it reports pipeline-not-wired (P9-D.2.b) - package.json: onnxruntime-web optionalDependency + vendor:onnx + release:prepare runs the onnx vendor sync - browser-transformer exports the runtime API - guards: local-security recognizes public/vendor/onnxruntime/ + STRICT; direction-test asserts onnxruntime-web; release-readiness expects new release:prepare/vendor:onnx scripts - ocr-baseline-test 36th block (EP selection + Node vendor-load reject) Tauri CSP already covers ORT (wasm-unsafe-eval + worker-src blob: + connect-src 'self'); no change. npm test (24), git diff --check, release:prepare all pass. No forced install of onnxruntime-web. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
… fix Let users import PP-OCRv5 det/cls/rec ONNX models into the local cache from the security center, flipping paddleOcrEngine from model-missing toward ready. Reuses the tesseract tessdata local-import pattern (file picker + SHA-256 + IndexedDB) — the project's no-network STRICT guard forbids remote fetch, so "on-demand download" means local user import. - security-center.js: renderPaddleActions for engine=paddleocr rows (import det.onnx/cls.onnx/rec.onnx + clear); importPaddleModel (sha256 -> store at paddleocr/v5/<file> -> ensureProbe; only flips to AVAILABLE + sets vendor ready when all three present, else reports which are missing); clearPaddleModels; click delegation for data-import-paddle/data-clear-paddle - Fix latent bug: paddleOcrEngine/tesseractOCREngine stored readiness on a frozen instance prop, so ensureProbe()'s assignment throws "Cannot assign to read only property" in ES-module strict mode — which made the security center import flow fail silently (caught as "import failed"). Readiness now lives in module-level state; the engine objects stay frozen. - ocr-baseline-test: 37th block (availability flips with model presence + vendor flag) + ensureProbe no-throw assertion on the frozen tesseract engine - browser-smoke: assert #modelCacheFileInput present - docs/DEVELOPMENT_TASKS synced npm test (24), git diff --check, release:prepare all pass. No remote fetch; det/cls/rec inference + CTC decode deferred to P9-D.2.b. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Implement the real PP-OCRv5 inference pipeline as pure, fully-testable
functions plus an orchestrator that takes injectable sessions — so it runs
end-to-end in Node with mock sessions + synthetic tensors, no real models.
- paddle-ocr-pipeline.js: parseCharDictionary, preprocessForDetection
(ImageNet norm + multiple-of-32 + limit_side_len), preprocessForRecognition
(height 48, [-1,1]), dbPostProcess (threshold + 4-connected components +
axis-aligned bbox + box-score filter + scale-back + reading order),
ctcGreedyDecode (argmax -> collapse repeats -> drop blank -> map dict),
cropImageData, resizeRgba; runPaddlePipeline({ ort, det/cls/recSession,
imageData, dictionary }) -> OCRResult
- paddle-ocr-engine.recognize: in-browser decode (Image+canvas, no fetch ->
respects the no-network guard) -> load det/cls/rec buffers + optional dict
from cache -> createOcrSession x3 -> runPaddlePipeline; Node still rejects
at loadOnnxRuntime before any browser-only path
- browser-transformer exports the pipeline API
- scripts/paddle-ocr-pipeline-test.js (9 blocks incl. mock-session e2e
decoding "HI") wired into npm test (25th script)
- guards: local-security ALLOWED/STRICT; direction-test runPaddlePipeline/
ctcGreedyDecode
- resource-budget: public/core 256KB -> 320KB (pure-JS algorithm code, no
model weights; models stay in model-cache) with rationale
cls angle correction / minAreaRect+unclip / multi-column deferred. Real-model
end-to-end is browser/manual. npm test (25), git diff --check,
release:prepare all pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…eract)
Capstone for the PP-OCRv5 local advanced-OCR chain: prefer paddle over
tesseract when both are available, so PNG / scanned-PDF OCR stages
automatically use the higher-accuracy engine.
- ocr-engine.js pickForTask: priority-aware — sort task candidates by
priority (desc, default 0) and pick the first available; fall back to the
last-registered candidate when none are available (unchanged)
- paddle-ocr-engine priority 20, tesseract-engine priority 10
(placeholder stays 0) -> PP-OCRv5 wins when available
- PNG / scanned-PDF stages need no change: enhanceWithOCR /
runScannedPdfOCRStage resolve via pickForTask("ocr-text")
- ocr-baseline-test 38th block: priority selection unit + default-registry
preference flip (both available -> paddleocr-v5; remove paddle models ->
tesseract-zh-en)
The PP-OCRv5 chain (contract -> runtime -> model import -> inference
pipeline -> route preference) is now complete; real-model end-to-end is a
browser/manual step. npm test (25), git diff --check, release:prepare pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The PP-OCRv5 chain (P9-D.1..D.4) is fully covered in Node via mock sessions and pure-function tests; real ONNX inference can only run in the browser/Tauri (WebGPU/WASM). Document the manual steps: install+vendor onnxruntime-web, import det/cls/rec ONNX + dictionary via the security center, verify engine priority, real PNG/scanned-PDF recognition, the verification report, and the no-network guarantee. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Make advanced OCR actually usable (P9-D.1..D.4 were contract/runtime/pipeline
scaffolding with no real model). Install the real runtime + models and
auto-load them on launch — no manual import needed.
- onnxruntime-web installed (optionalDependency 1.26.0); sync-onnxruntime-vendor
trimmed to the minimal set (ort.min.mjs + ort-wasm-simd-threaded.jsep.{mjs,wasm},
JSEP build covers WebGPU+WASM, ~25MB; drops ~68MB of redundant variants)
- real PP-OCRv5 mobile ONNX (det 4.8MB / cls 0.58MB / rec 16.6MB) + dict
downloaded to public/vendor/paddleocr/ (from the OnnxOCR PP-OCRv5 ONNX repo)
- paddle-default-models.js: ensurePaddleDefaultModels() idempotently fetches
the same-origin /vendor/paddleocr/ models into defaultOCRStorage (IndexedDB),
marks the engine ready + probes -> advanced OCR works out of the box; missing
vendor is silently skipped (manual security-center import still works)
- app.js calls it fire-and-forget on init; browser-transformer exports it
- .gitignore excludes the heavy onnxruntime + paddleocr vendors (reproducible
via npm i + vendor script + local download; bundled into the build from disk)
- guards: paddle-default-models ALLOWED/STRICT; isLocalVendorAsset trusts the
onnxruntime vendor (its minified bundle has CDN strings; the no-network
guarantee comes from pinning ort.env.wasm.wasmPaths same-origin + Tauri CSP
connect-src 'self'); public/vendor budget 6MB -> 64MB with rationale
Real ONNX inference runs in the browser/WebGPU only; Node stays mock/pure-fn
covered (npm test 25 green). Browser e2e steps: docs/PP_OCRV5_BROWSER_VERIFICATION.md
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Render LaTeX math in document previews. The markdown inline tokenizer previously ate math (\frac -> rac via backslash escaping; _ -> <em>), so math is now a protected token. - inline-tokens.js: recognize $$...$$ (display) and $...$ (inline) BEFORE escape handling; capture content verbatim (no recursion/escaping). Inline heuristic (no inner-edge whitespace) excludes currency like "$5 and $10" - semantic-inlines.js: createInlineMath + math node rendered in plainText/markdown/html; HTML emits <span class="t2f-math" data-tex="raw" data-display=".."> with a no-JS fallback ($tex$); markdown round-trips - vendored KaTeX v0.17.0 (css + js + 20 woff2 fonts, ~592KB) at public/vendor/katex/; katex-render.js renderMathIn() typesets .t2f-math spans via the global katex (same-origin, zero network; silent fallback if katex absent) - index.html/preview.html load katex css + defer js; app.js (3 preview sites) + preview.js call renderMathIn after rendering - local-security: trust the katex vendor (its only http strings are W3C MathML/SVG namespaces, not network) - scripts/latex-math-test.js (7 blocks: tokenization with backslash+underscore preserved, currency exclusion, katex-targetable span, md round-trip, plain-text + factory, md->html no <em>) -> npm test (26th) npm test (26) passes; existing conversion snapshots survive the new token. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
OCR was unreachable in the running app: convertWithWorker sent conversions to a Web Worker that uses sync convertContent (no OCR stage), and OCR needs main-thread canvas decode + onnxruntime which a Worker can't provide. So image and scanned-PDF inputs never triggered OCR. Route png/pdf inputs to convertContentAsync on the main thread (text PDFs still take the normal path; images and scanned PDFs now run OCR), and ensure the bundled PP-OCRv5 models are loaded first (idempotent ensurePaddleDefaultModels). OCR has no dedicated button — it runs automatically when converting an image or scanned PDF to a text format. npm test (26) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…urate Validated the inference pipeline against the real PP-OCRv5 ONNX models with onnxruntime-node (dev-only). Findings: - rec is correct: decodes the PaddleOCR word_10 crop to "PAIN", output classes C=18385 exactly match the dictionary (blank + 18383 + space). RGB channel order + (x/255-0.5)/0.5 normalization are right (no BGR needed) - det is correct: produces a probability map; 16 text boxes on a real document - root cause of garbled output: dbPostProcess emitted tight axis-aligned bboxes without PP-OCR's "unclip" expansion, clipping character strokes -> wrong recognition (avgConf 0.41, CJK gibberish). Adding unclip (distance = area*ratio/perimeter, unclipRatio default 1.6) yields coherent, correct text on a full product-label document (avgConf 0.978) - dbPostProcess: add unclipRatio param + outward box expansion - pipeline unit test: assert exact bbox with unclipRatio:0, plus an expansion assertion - scripts/paddle-ocr-integration-test.js: runs the real rec model via onnxruntime-node on a committed fixture (samples/ocr/word-PAIN.png), asserts "PAIN" + C==dict; gracefully skips (exit 0) when the dev deps / models / fixture are absent, so npm test stays green in CI. Wired in (27th) Real rec hits 0.991 confidence locally. npm test (27) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ntrol Improve hard-case recognition and add explicit quality assessment, validated against the real PP-OCRv5 models with onnxruntime-node. - orientation: the cls model outputs [1,2] softmax over [0°,180°] (upright->[1,0], flipped->[0,1]). runPaddlePipeline now applies it via interpretClsOutput, rotating 180° crops before rec. Validated: a fully upside-down document is read correctly (all 16 lines flipped, avgConf 0.976) - vertical/sideways: tall boxes (h/w > verticalAspect) additionally try 90° cw/ccw and keep the highest-confidence read (robust for sideways labels; cost limited to the rare tall boxes) - rotateImageData180 / rotateImageData90(dir) / interpretClsOutput pure helpers (unit-tested: 180 is its own inverse, 90 swaps dims + maps corners, cls threshold branches) - quality control: runPaddlePipeline returns a quality summary (lineCount, averageConfidence, minConfidence, lowConfidenceLines, rotatedLines, grade high/medium/low); enhanceWithOCR records result.quality into metadata.modelReview.ocrQuality, alongside the existing per-line confidence + detectOCRLowConfidence validator + P9-C OCR-readback verification - pipeline test: rotation/cls/quality assertions (12 blocks total) Still limited: strong italic / ornate artistic text (needs minAreaRect + perspective-warp polygon boxes and a stronger rec model). npm test (27) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rescued) Per the steer — only convert the text content (no artistic-style preservation), focus on denoising. Add image denoising that cleans noisy/artistic-background images without hurting clean ones. - denoiseImageData: per-channel 3x3 median filter (edge-preserving, removes salt-and-pepper); estimateNoiseLevel: fraction of isolated pixels that jump far from their 3x3 median (separates speckle from text edges) - runPaddlePipeline options.denoise = "auto"(default)|true|false. AUTO denoises only when estimateNoiseLevel > threshold (default 0.05), because median filtering softens clean text (measured 0.974 -> 0.903), so clean images are never denoised - validated on real models: clean doc (noise 0.016) stays untouched at 0.974; 15% salt-and-pepper (noise 0.10) is denoised, recovering detection from 4 -> 16 lines and avgConf 0.692 -> 0.832 (heavy noise collapses detection; denoise rescues it) - quality summary adds denoised + noiseLevel - pipeline test: noise-estimate / median / auto-gating assertions (14 blocks) Explicitly not doing artistic-style preservation (text content only) or minAreaRect perspective. npm test (27) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Skew is the biggest remaining recognition gap: a +10° rotated document drops from 16 lines/0.974 to 8 lines/0.535 (detection collapses). Auto-deskew fixes it. - rotateImageDataByAngle(im, deg): arbitrary-angle nearest-neighbor rotation with canvas expansion + white background - estimateSkewAngle(probData, mapW, mapH): shear-projection histogram variance on the (binarized, downsampled) det probability map -> the angle that makes text rows most horizontal (the deskew rotation) - runPaddlePipeline: extract a detect(image) helper; options.deskew = true(default)|false. Estimate skew from the first det; if |est| >= minSkew (default 3°), rotate the image upright and re-detect, then proceed. Upright images estimate ~0 and skip the second det (zero overhead); rec always runs once - validated on real models: +10° doc recovers 8/0.535 -> 16 lines/0.970 (skewApplied=10), -8° recovers too, upright untouched (skewApplied=0, 0.974) - quality summary adds skewApplied - pipeline test: arbitrary rotation + skew estimation (slanted synthetic rows detected ~angle, flat rows ~0) — 16 blocks npm test (27) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Address "enhance recognition of the file's internal text format": group OCR lines (with bboxes) into a structured document instead of one flat paragraph. - ocr-structure.js: deriveOcrStructure(lines) sorts by reading order (y,x), detects headings by relative font size (line height >= 1.35x median -> heading, level 1-3 by ratio), splits paragraphs on large vertical gaps (> 0.7x median), joins same-paragraph lines (CJK tight, latin spaced); falls back to a single paragraph when bbox geometry is missing. blocksFromOcrResult walks pages - png-ocr enhanceWithOCR now emits structured heading/paragraph blocks (replacing the old "join all 16 lines into one paragraph"); removed the dead helper + unused import - validated on the real product label: 16 lines -> 4 blocks (title becomes a heading, body grouped into paragraphs) - browser-transformer exports deriveOcrStructure / blocksFromOcrResult - scripts/ocr-structure-test.js (7 blocks) wired into npm test (28th) npm test (28) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Make the OCR quality data (engine, confidence, grade, low-conf lines, skew, rotation, denoise) visible — it was computed but never shown. - fix: _runRepairCycle overwrote the OCR-stage modelReview with the repair engine's, dropping the ocr/ocrQuality sub-objects. Merge them back so the recognition quality reaches result.quality.modelReview - index.html: add an "OCR 识别质量" row (#verificationOcrRecognitionRow, hidden unless OCR ran) - app.js renderVerificationReport reads quality.modelReview.ocr + .ocrQuality and renders engine / lines / confidence / grade / low-confidence / skew / rotation / denoised, with grade-driven coloring - browser-smoke asserts the row; ocr-baseline adds a regression assertion that modelReview.ocr survives the default repair path npm test (28) passes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ification Bump version 2.2.0 -> 2.3.0 across package.json / tauri.conf.json / Cargo.toml / Cargo.lock (desktop-shell-test enforces the sync). - README: refreshed capabilities (real PP-OCRv5 OCR with orientation/skew/ denoise/structure/quality, LaTeX rendering, three-layer verification), roadmap, limitations, test/sample sections; badge -> 2.3.0 - CHANGELOG: [2.3.0] entry covering the OCR pipeline, LaTeX, verification, Repair Engine + model-cache, sample generator, the four latent-bug fixes, and the PP-OCRv5 direction pivot - RELEASE_NOTES_v2.3.0.md added - DEVELOPMENT_TASKS acceptance command -> trans2former-2.3.0 - gitignore the tesseract vendor too (reproducible, like onnxruntime/paddleocr) npm test (28), git diff --check, release:prepare (release/trans2former-2.3.0, git-ignored) all pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…lity, OCR blockId, reproducible models - sanitize tesseract vendor bundles (rewrite CDN defaults to same-origin, drop .map, post-assert) so local-security-test passes - security-center: mark PP-OCRv5 vendor-ready before ensureProbe so imported det+rec flips to available - make cls (direction classifier) optional across availability gates (runtime/pipeline already tolerated null cls) - assign stable ocr-block ids at append time + map OCR lines to blocks by trimmed-text containment so low-confidence repair targets resolve (png + scan-pdf); drop the never-incrementing scan-pdf loop - add reproducible 'npm run vendor:paddle' (pinned ppu-paddle-ocr-models + committed SHA-256 manifest), wired into release:prepare - parseCharDictionary: don't double-append space when dict already ends with one (fixes off-by-one vs ppu dict); never trim() so U+3000 token is preserved - resource-budget: correct public/vendor budget to account for tesseract (~30MB) it previously omitted - align README/.gitignore/PP_OCRV5_BROWSER_VERIFICATION docs to bundled+reproducible policy; update/extend tests Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 23cd73a68d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| export const PADDLE_OCR_MODEL_FILES = Object.freeze(["det.onnx", "cls.onnx", "rec.onnx"]); | ||
| // 必选集：det（DB 检测）+ rec（CTC 识别）。cls（方向分类）为可选——管线运行时已容忍其 | ||
| // 缺失（clsSession 为 null 时跳过 180° 校正），故不纳入可用性闸门。 | ||
| export const PADDLE_OCR_REQUIRED_FILES = Object.freeze(["det.onnx", "rec.onnx"]); |
There was a problem hiding this comment.
When PP-OCRv5 models are supplied through the Security Center, the UI only offers det.onnx, cls.onnx, and rec.onnx, and this readiness check marks the engine available after just det+rec. In that manual-import path, or if the bundled dict.txt fetch fails, recognize() falls back to dictionary = [], so ctcGreedyDecode cannot map any non-blank class to text and the OCR stage returns empty output despite the engine being selected as ready. Include dict.txt in the required set/import flow or fail readiness until it is present.
Useful? React with 👍 / 👎.
| // 先置位 vendor-ready（用户已选用 PP-OCRv5），再 probe；否则 ensureProbe 在 vendor | ||
| // 未置位时恒返回 false，状态永远翻不过去。真正的 onnxruntime 运行时加载仍在 | ||
| // recognize() 时把关。对齐 paddle-default-models.js / tesseract 导入流程的顺序。 | ||
| markPaddleOcrVendorReady(true); |
There was a problem hiding this comment.
In environments where the user imports det/rec but public/vendor/onnxruntime/ort.min.mjs is absent (the sync script explicitly skips when the optional dependency is missing, and the vendor dir is gitignored), this line still marks PP-OCR as vendor-ready; after ensureProbe() it outranks a working Tesseract engine, then recognize() fails with OCR_VENDOR_LOAD_FAILED and the OCR stage does not fall back. Keep Paddle unavailable unless the ONNX runtime bundle is actually loadable, or fall back to the next available engine on runtime-load failure.
Useful? React with 👍 / 👎.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
背景
评审报告列出 4 项问题(2×P1、2×P2),已逐条在代码上复现并修复;
npm test现全绿(29 项通过,EXIT=0)。修复
sync-tesseract-vendor.js复制后清洗 bundle(cdn.jsdelivr.net→ 同源/vendor)、不再复制.map、清洗后断言无残留远程协议。CDN 默认值本是死代码(运行时恒以同源路径覆盖)。security-center在ensureProbe()前先markPaddleOcrVendorReady(true),对齐paddle-default-models范式。ocr-block-<绝对索引>id,新增mapLinesToBlockIds(单调游标+修剪文本包含)在 png/scan-pdf 回填;删掉扫描 PDF 不自增的死循环。npm run vendor:paddle(钉定ppu-paddle-ocr-models+ 入库 SHA-256 manifest),接入release:prepare;cls方向分类设为可选(运行时本就容忍 null)。文档(README/.gitignore/验证清单)对齐。连带修复(被 P1 红掩盖、此次解开)
parseCharDictionaryoff-by-one:字典已以空格行结尾时不重复追加空格;保留全角空格 U+3000(不可trim())。真实模型集成测试解出 "PAIN"(conf 0.991,C 对齐)。resource-budget预算账目修正:public/vendor上限此前漏算 tesseract(~30MB),据实调整。验证
npm run vendor:tesseract重生清洗版 →local-security-test通过。npm run vendor:paddle→ 下载 det/rec/dict + SHA-256 校验通过;离线优雅跳过。npm test全绿(29/29)。docs/PP_OCRV5_BROWSER_VERIFICATION.md手验。注意
.gitignore),由vendor:paddle钉定来源 + SHA-256 重建,随release:prepare打包。🤖 Generated with Claude Code