@effortlessmotion/pdf-to-md 0.1.8 → 0.2.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (108) hide show
  1. package/README.md +5 -5
  2. package/dist/cjs/analyze.d.cts +3 -1
  3. package/dist/cjs/analyze.d.ts +3 -1
  4. package/dist/cjs/analyze.js +38 -100
  5. package/dist/cjs/analyze.js.map +1 -1
  6. package/dist/cjs/audit.d.cts +6 -5
  7. package/dist/cjs/audit.d.ts +6 -5
  8. package/dist/cjs/audit.js +368 -60
  9. package/dist/cjs/audit.js.map +1 -1
  10. package/dist/cjs/cli/audit.js.map +1 -1
  11. package/dist/cjs/cli/extract.js +13 -48
  12. package/dist/cjs/cli/extract.js.map +1 -1
  13. package/dist/cjs/cli/index.js.map +1 -1
  14. package/dist/cjs/cli/render.js.map +1 -1
  15. package/dist/cjs/cli/types.js.map +1 -1
  16. package/dist/cjs/cli.js.map +1 -1
  17. package/dist/cjs/collect.js +91 -35
  18. package/dist/cjs/collect.js.map +1 -1
  19. package/dist/cjs/dict.js.map +1 -1
  20. package/dist/cjs/emit.js.map +1 -1
  21. package/dist/cjs/engines/tesseract.js.map +1 -1
  22. package/dist/cjs/errors.d.cts +2 -0
  23. package/dist/cjs/errors.d.ts +2 -0
  24. package/dist/cjs/errors.js.map +1 -1
  25. package/dist/cjs/extract.d.cts +20 -9
  26. package/dist/cjs/extract.d.ts +20 -9
  27. package/dist/cjs/extract.js +333 -106
  28. package/dist/cjs/extract.js.map +1 -1
  29. package/dist/cjs/geometry.js.map +1 -1
  30. package/dist/cjs/index.js.map +1 -1
  31. package/dist/cjs/lines.js +21 -10
  32. package/dist/cjs/lines.js.map +1 -1
  33. package/dist/cjs/markdown.js +140 -54
  34. package/dist/cjs/markdown.js.map +1 -1
  35. package/dist/cjs/page-marker.d.cts +1 -1
  36. package/dist/cjs/page-marker.d.ts +1 -1
  37. package/dist/cjs/page-marker.js +39 -8
  38. package/dist/cjs/page-marker.js.map +1 -1
  39. package/dist/cjs/pdf-open.js.map +1 -1
  40. package/dist/cjs/raster.js.map +1 -1
  41. package/dist/cjs/render-analysis.d.cts +4 -2
  42. package/dist/cjs/render-analysis.d.ts +4 -2
  43. package/dist/cjs/render-analysis.js +5 -3
  44. package/dist/cjs/render-analysis.js.map +1 -1
  45. package/dist/cjs/report.d.cts +2 -2
  46. package/dist/cjs/report.d.ts +2 -2
  47. package/dist/cjs/report.js +27 -8
  48. package/dist/cjs/report.js.map +1 -1
  49. package/dist/cjs/resolve.js.map +1 -1
  50. package/dist/cjs/text-analysis.d.cts +12 -0
  51. package/dist/cjs/text-analysis.d.ts +12 -0
  52. package/dist/cjs/text-analysis.js +175 -0
  53. package/dist/cjs/text-analysis.js.map +1 -0
  54. package/dist/cjs/types.d.cts +40 -3
  55. package/dist/cjs/types.d.ts +40 -3
  56. package/dist/cjs/types.js +2 -3
  57. package/dist/cjs/types.js.map +1 -1
  58. package/dist/cjs/validate.js.map +1 -1
  59. package/dist/esm/analyze.d.ts +3 -1
  60. package/dist/esm/analyze.js +18 -89
  61. package/dist/esm/analyze.js.map +1 -1
  62. package/dist/esm/audit.d.ts +6 -5
  63. package/dist/esm/audit.js +236 -29
  64. package/dist/esm/audit.js.map +1 -1
  65. package/dist/esm/cli/audit.js.map +1 -1
  66. package/dist/esm/cli/extract.js +10 -23
  67. package/dist/esm/cli/extract.js.map +1 -1
  68. package/dist/esm/cli/index.js.map +1 -1
  69. package/dist/esm/cli/render.js.map +1 -1
  70. package/dist/esm/cli/types.js.map +1 -1
  71. package/dist/esm/cli.js.map +1 -1
  72. package/dist/esm/collect.js +64 -15
  73. package/dist/esm/collect.js.map +1 -1
  74. package/dist/esm/dict.js.map +1 -1
  75. package/dist/esm/emit.js.map +1 -1
  76. package/dist/esm/engines/tesseract.js.map +1 -1
  77. package/dist/esm/errors.d.ts +2 -0
  78. package/dist/esm/errors.js.map +1 -1
  79. package/dist/esm/extract.d.ts +20 -9
  80. package/dist/esm/extract.js +167 -76
  81. package/dist/esm/extract.js.map +1 -1
  82. package/dist/esm/geometry.js.map +1 -1
  83. package/dist/esm/index.js.map +1 -1
  84. package/dist/esm/lines.js +14 -10
  85. package/dist/esm/lines.js.map +1 -1
  86. package/dist/esm/markdown.js +110 -27
  87. package/dist/esm/markdown.js.map +1 -1
  88. package/dist/esm/page-marker.d.ts +1 -1
  89. package/dist/esm/page-marker.js +15 -9
  90. package/dist/esm/page-marker.js.map +1 -1
  91. package/dist/esm/pdf-open.js.map +1 -1
  92. package/dist/esm/raster.js.map +1 -1
  93. package/dist/esm/render-analysis.d.ts +4 -2
  94. package/dist/esm/render-analysis.js +9 -5
  95. package/dist/esm/render-analysis.js.map +1 -1
  96. package/dist/esm/report.d.ts +2 -2
  97. package/dist/esm/report.js +12 -7
  98. package/dist/esm/report.js.map +1 -1
  99. package/dist/esm/resolve.js.map +1 -1
  100. package/dist/esm/text-analysis.d.ts +12 -0
  101. package/dist/esm/text-analysis.js +81 -0
  102. package/dist/esm/text-analysis.js.map +1 -0
  103. package/dist/esm/types.d.ts +40 -3
  104. package/dist/esm/types.js +4 -4
  105. package/dist/esm/types.js.map +1 -1
  106. package/dist/esm/validate.js.map +1 -1
  107. package/package.json +1 -1
  108. package/skills/pdf-to-md/SKILL.md +39 -66
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/emit.ts"],"sourcesContent":["// emit.ts — OCR-path pass 3: lines -> markdown blocks.\nimport { getDict } from './dict.ts';\nimport type { OcrLine, PageOCR, Tuning, Word } from './geometry.ts';\nimport { bucket } from './geometry.ts';\n\n// Join a wrapped line onto the open paragraph, closing a line-final hyphen\n// only when the dictionary confirms the joined form is a word.\nfunction hyphenJoin(acc: string, next: string): string {\n if (acc.endsWith('-')) {\n const withoutDash = acc.slice(0, -1);\n const pen = withoutDash.at(-1);\n if (pen && /[A-Za-z0-9]/.test(pen)) {\n const head = withoutDash.split(' ').pop() ?? '';\n const firstTok = next.split(' ')[0] ?? '';\n const filtered = [...firstTok].filter((c) => /[A-Za-z-]/.test(c)).join('');\n const tail = filtered.split('-')[0] ?? '';\n const close = getDict().has((head + tail).toLowerCase());\n return (close ? withoutDash : acc) + next;\n }\n }\n return `${acc} ${next}`;\n}\n\n// Blocks with too few word-like tokens to trust as prose (see\n// UNSTRUCTURED_WORDLIKE_FRACTION below) are labeled `> [review: reason] text` instead of guessed at.\nconst REVIEW_REASON = {\n unstructuredLabels: 'unstructured labels',\n} as const;\n\n/** Share of word-like tokens (3+ letter run) below which a block counts as\n * unstructured. Measured (`review-blocks.test.ts`): junk 0.000-0.100 vs prose 0.743-0.857, cut 0.4. */\nexport const UNSTRUCTURED_WORDLIKE_FRACTION = 0.4;\nfunction isUnstructured(text: string): boolean {\n const tokens = text.split(/\\s+/).filter(Boolean);\n if (tokens.length === 0) return false; // never mark an empty block\n const wordLike = tokens.filter((t) => /[A-Za-z]{3}/.test(t)).length;\n return wordLike / tokens.length < UNSTRUCTURED_WORDLIKE_FRACTION;\n}\n\nfunction reviewLine(reason: string, text: string): string {\n return `> [review: ${reason}] ${text}`;\n}\n\n// One classified block, carrying constituent words for analyze()'s\n// block->word linkage; discriminated on `kind` so `reason`/`level` are compile-time guaranteed present on their variants.\nexport type EmitBlock =\n | { kind: 'heading'; text: string; level: number; words: Word[] }\n | { kind: 'paragraph'; text: string; words: Word[] }\n | { kind: 'quote'; text: string; words: Word[] }\n | {\n kind: 'review';\n text: string;\n /** The parseable reason inside `> [review: ...]`, carried on the block\n * so more reasons can be added without changing the render format. */\n reason: string;\n words: Word[];\n };\n\n// Renders one EmitBlock to the exact markdown emitPage used to push\n// directly — shared so extract.ts and analyze.ts produce byte-identical text.\nexport function renderEmitBlock(b: EmitBlock): string {\n switch (b.kind) {\n case 'heading':\n return `${'#'.repeat(b.level)} ${b.text}`;\n case 'quote':\n return `> ${b.text}`;\n case 'review':\n return reviewLine(b.reason, b.text);\n default:\n return b.text;\n }\n}\n\nexport function emitPage(p: PageOCR, levelOf: Map<number, number>, tuning: Tuning): EmitBlock[] {\n const blocks: EmitBlock[] = [];\n let para: OcrLine[] = [];\n const flush = () => {\n if (!para.length) return;\n const texts = para.map((l) => l.text);\n const merged = texts.slice(1).reduce(hyphenJoin, texts[0]);\n const words = para.flatMap((l) => l.words);\n // Same review-marker rule as the heading path below: no word may be\n // dropped or reordered here, only how the block renders changes.\n blocks.push(isUnstructured(merged) ? { kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words } : { kind: 'paragraph', text: merged, words });\n para = [];\n };\n\n let i = 0;\n while (i < p.lines.length) {\n const line: OcrLine = p.lines[i];\n const prev = i > 0 ? p.lines[i - 1] : null;\n if (prev && (line.col !== prev.col || prev.midY - line.midY > p.pitch * tuning.paraGap)) {\n // Join across a column break (instead of flushing) only when the column\n // changed, the paragraph dangles (no terminal punct), and the new text starts lowercase; a miss stays visible via pageStats.danglingLong.\n const danglingEnd = para.length > 0 && !/[.!?:;\"'”’)\\]]$/.test(para[para.length - 1].text);\n const continuationStart = /^[a-z]/.test(line.text);\n const joinAcrossColumns = line.col !== prev.col && danglingEnd && continuationStart && !levelOf.has(bucket(line.h));\n if (!joinAcrossColumns) flush();\n }\n\n if (levelOf.has(bucket(line.h))) {\n // Gather the run of consecutive heading-size lines: a wrapped title is\n // one heading, a long run is a pull quote in a bigger face.\n flush();\n const run: OcrLine[] = [line];\n let maxB = bucket(line.h);\n while (i + 1 < p.lines.length && levelOf.has(bucket(p.lines[i + 1].h)) && p.lines[i + 1].col === p.lines[i].col && p.lines[i].midY - p.lines[i + 1].midY < Math.max(p.lines[i].h, p.lines[i + 1].h) * 2.2) {\n i++;\n run.push(p.lines[i]);\n maxB = Math.max(maxB, bucket(p.lines[i].h));\n }\n const runTexts = run.map((l) => l.text);\n const merged = runTexts.slice(1).reduce(hyphenJoin, runTexts[0]);\n const words = run.flatMap((l) => l.words);\n // Below UNSTRUCTURED_WORDLIKE_FRACTION = chart labels/stray numerals,\n // not a heading; checked FIRST so a long numeral run isn't asserted as a pull quote.\n const junk = isUnstructured(merged);\n if (junk) blocks.push({ kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words });\n else if (merged.length > 150) blocks.push({ kind: 'quote', text: merged, words });\n else blocks.push({ kind: 'heading', text: merged, level: levelOf.get(maxB) as number, words });\n } else {\n para.push(line);\n }\n i++;\n }\n flush();\n return blocks;\n}\n"],"names":["getDict","bucket","hyphenJoin","acc","next","endsWith","withoutDash","slice","pen","at","test","filtered","head","split","pop","firstTok","filter","c","join","tail","close","has","toLowerCase","REVIEW_REASON","unstructuredLabels","UNSTRUCTURED_WORDLIKE_FRACTION","isUnstructured","text","tokens","Boolean","length","wordLike","t","reviewLine","reason","renderEmitBlock","b","kind","repeat","level","emitPage","p","levelOf","tuning","blocks","para","flush","texts","map","l","merged","reduce","words","flatMap","push","i","lines","line","prev","col","midY","pitch","paraGap","danglingEnd","continuationStart","joinAcrossColumns","h","run","maxB","Math","max","runTexts","junk","get"],"mappings":"AAAA,uDAAuD;AACvD,SAASA,OAAO,QAAQ,YAAY;AAEpC,SAASC,MAAM,QAAQ,gBAAgB;AAEvC,2EAA2E;AAC3E,+DAA+D;AAC/D,SAASC,WAAWC,GAAW,EAAEC,IAAY;IAC3C,IAAID,IAAIE,QAAQ,CAAC,MAAM;QACrB,MAAMC,cAAcH,IAAII,KAAK,CAAC,GAAG,CAAC;QAClC,MAAMC,MAAMF,YAAYG,EAAE,CAAC,CAAC;QAC5B,IAAID,OAAO,cAAcE,IAAI,CAACF,MAAM;gBACrBF,wBACIF,cAEJO;YAHb,MAAMC,QAAON,yBAAAA,YAAYO,KAAK,CAAC,KAAKC,GAAG,gBAA1BR,oCAAAA,yBAAgC;YAC7C,MAAMS,YAAWX,eAAAA,KAAKS,KAAK,CAAC,IAAI,CAAC,EAAE,cAAlBT,0BAAAA,eAAsB;YACvC,MAAMO,WAAW;mBAAII;aAAS,CAACC,MAAM,CAAC,CAACC,IAAM,YAAYP,IAAI,CAACO,IAAIC,IAAI,CAAC;YACvE,MAAMC,QAAOR,mBAAAA,SAASE,KAAK,CAAC,IAAI,CAAC,EAAE,cAAtBF,8BAAAA,mBAA0B;YACvC,MAAMS,QAAQpB,UAAUqB,GAAG,CAAC,AAACT,CAAAA,OAAOO,IAAG,EAAGG,WAAW;YACrD,OAAO,AAACF,CAAAA,QAAQd,cAAcH,GAAE,IAAKC;QACvC;IACF;IACA,OAAO,GAAGD,IAAI,CAAC,EAAEC,MAAM;AACzB;AAEA,8DAA8D;AAC9D,qGAAqG;AACrG,MAAMmB,gBAAgB;IACpBC,oBAAoB;AACtB;AAEA;qGACqG,GACrG,OAAO,MAAMC,iCAAiC,IAAI;AAClD,SAASC,eAAeC,IAAY;IAClC,MAAMC,SAASD,KAAKd,KAAK,CAAC,OAAOG,MAAM,CAACa;IACxC,IAAID,OAAOE,MAAM,KAAK,GAAG,OAAO,OAAO,4BAA4B;IACnE,MAAMC,WAAWH,OAAOZ,MAAM,CAAC,CAACgB,IAAM,cAActB,IAAI,CAACsB,IAAIF,MAAM;IACnE,OAAOC,WAAWH,OAAOE,MAAM,GAAGL;AACpC;AAEA,SAASQ,WAAWC,MAAc,EAAEP,IAAY;IAC9C,OAAO,CAAC,WAAW,EAAEO,OAAO,EAAE,EAAEP,MAAM;AACxC;AAiBA,oEAAoE;AACpE,8EAA8E;AAC9E,OAAO,SAASQ,gBAAgBC,CAAY;IAC1C,OAAQA,EAAEC,IAAI;QACZ,KAAK;YACH,OAAO,GAAG,IAAIC,MAAM,CAACF,EAAEG,KAAK,EAAE,CAAC,EAAEH,EAAET,IAAI,EAAE;QAC3C,KAAK;YACH,OAAO,CAAC,EAAE,EAAES,EAAET,IAAI,EAAE;QACtB,KAAK;YACH,OAAOM,WAAWG,EAAEF,MAAM,EAAEE,EAAET,IAAI;QACpC;YACE,OAAOS,EAAET,IAAI;IACjB;AACF;AAEA,OAAO,SAASa,SAASC,CAAU,EAAEC,OAA4B,EAAEC,MAAc;IAC/E,MAAMC,SAAsB,EAAE;IAC9B,IAAIC,OAAkB,EAAE;IACxB,MAAMC,QAAQ;QACZ,IAAI,CAACD,KAAKf,MAAM,EAAE;QAClB,MAAMiB,QAAQF,KAAKG,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;QACpC,MAAMuB,SAASH,MAAMxC,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAY6C,KAAK,CAAC,EAAE;QACzD,MAAMK,QAAQP,KAAKQ,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;QACzC,oEAAoE;QACpE,iEAAiE;QACjER,OAAOU,IAAI,CAAC5B,eAAewB,UAAU;YAAEb,MAAM;YAAUV,MAAMuB;YAAQhB,QAAQX,cAAcC,kBAAkB;YAAE4B;QAAM,IAAI;YAAEf,MAAM;YAAaV,MAAMuB;YAAQE;QAAM;QAClKP,OAAO,EAAE;IACX;IAEA,IAAIU,IAAI;IACR,MAAOA,IAAId,EAAEe,KAAK,CAAC1B,MAAM,CAAE;QACzB,MAAM2B,OAAgBhB,EAAEe,KAAK,CAACD,EAAE;QAChC,MAAMG,OAAOH,IAAI,IAAId,EAAEe,KAAK,CAACD,IAAI,EAAE,GAAG;QACtC,IAAIG,QAASD,CAAAA,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAID,KAAKE,IAAI,GAAGH,KAAKG,IAAI,GAAGnB,EAAEoB,KAAK,GAAGlB,OAAOmB,OAAO,AAAD,GAAI;YACvF,wEAAwE;YACxE,0IAA0I;YAC1I,MAAMC,cAAclB,KAAKf,MAAM,GAAG,KAAK,CAAC,kBAAkBpB,IAAI,CAACmC,IAAI,CAACA,KAAKf,MAAM,GAAG,EAAE,CAACH,IAAI;YACzF,MAAMqC,oBAAoB,SAAStD,IAAI,CAAC+C,KAAK9B,IAAI;YACjD,MAAMsC,oBAAoBR,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAII,eAAeC,qBAAqB,CAACtB,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC;YACjH,IAAI,CAACD,mBAAmBnB;QAC1B;QAEA,IAAIJ,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC,IAAI;YAC/B,uEAAuE;YACvE,4DAA4D;YAC5DpB;YACA,MAAMqB,MAAiB;gBAACV;aAAK;YAC7B,IAAIW,OAAOnE,OAAOwD,KAAKS,CAAC;YACxB,MAAOX,IAAI,IAAId,EAAEe,KAAK,CAAC1B,MAAM,IAAIY,QAAQrB,GAAG,CAACpB,OAAOwC,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,MAAMzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACI,GAAG,KAAKlB,EAAEe,KAAK,CAACD,EAAE,CAACI,GAAG,IAAIlB,EAAEe,KAAK,CAACD,EAAE,CAACK,IAAI,GAAGnB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACK,IAAI,GAAGS,KAAKC,GAAG,CAAC7B,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC,EAAEzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,IAAI,IAAK;gBACzMX;gBACAY,IAAIb,IAAI,CAACb,EAAEe,KAAK,CAACD,EAAE;gBACnBa,OAAOC,KAAKC,GAAG,CAACF,MAAMnE,OAAOwC,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC;YAC3C;YACA,MAAMK,WAAWJ,IAAInB,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;YACtC,MAAMuB,SAASqB,SAAShE,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAYqE,QAAQ,CAAC,EAAE;YAC/D,MAAMnB,QAAQe,IAAId,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;YACxC,sEAAsE;YACtE,qFAAqF;YACrF,MAAMoB,OAAO9C,eAAewB;YAC5B,IAAIsB,MAAM5B,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAUV,MAAMuB;gBAAQhB,QAAQX,cAAcC,kBAAkB;gBAAE4B;YAAM;iBACjG,IAAIF,OAAOpB,MAAM,GAAG,KAAKc,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAASV,MAAMuB;gBAAQE;YAAM;iBAC1ER,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAWV,MAAMuB;gBAAQX,OAAOG,QAAQ+B,GAAG,CAACL;gBAAiBhB;YAAM;QAC9F,OAAO;YACLP,KAAKS,IAAI,CAACG;QACZ;QACAF;IACF;IACAT;IACA,OAAOF;AACT"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/open-source/formats/pdf-to-md/src/emit.ts"],"sourcesContent":["// emit.ts — OCR-path pass 3: lines -> markdown blocks.\nimport { getDict } from './dict.ts';\nimport type { OcrLine, PageOCR, Tuning, Word } from './geometry.ts';\nimport { bucket } from './geometry.ts';\n\n// Join a wrapped line onto the open paragraph, closing a line-final hyphen\n// only when the dictionary confirms the joined form is a word.\nfunction hyphenJoin(acc: string, next: string): string {\n if (acc.endsWith('-')) {\n const withoutDash = acc.slice(0, -1);\n const pen = withoutDash.at(-1);\n if (pen && /[A-Za-z0-9]/.test(pen)) {\n const head = withoutDash.split(' ').pop() ?? '';\n const firstTok = next.split(' ')[0] ?? '';\n const filtered = [...firstTok].filter((c) => /[A-Za-z-]/.test(c)).join('');\n const tail = filtered.split('-')[0] ?? '';\n const close = getDict().has((head + tail).toLowerCase());\n return (close ? withoutDash : acc) + next;\n }\n }\n return `${acc} ${next}`;\n}\n\n// Blocks with too few word-like tokens to trust as prose (see\n// UNSTRUCTURED_WORDLIKE_FRACTION below) are labeled `> [review: reason] text` instead of guessed at.\nconst REVIEW_REASON = {\n unstructuredLabels: 'unstructured labels',\n} as const;\n\n/** Share of word-like tokens (3+ letter run) below which a block counts as\n * unstructured. Measured (`review-blocks.test.ts`): junk 0.000-0.100 vs prose 0.743-0.857, cut 0.4. */\nexport const UNSTRUCTURED_WORDLIKE_FRACTION = 0.4;\nfunction isUnstructured(text: string): boolean {\n const tokens = text.split(/\\s+/).filter(Boolean);\n if (tokens.length === 0) return false; // never mark an empty block\n const wordLike = tokens.filter((t) => /[A-Za-z]{3}/.test(t)).length;\n return wordLike / tokens.length < UNSTRUCTURED_WORDLIKE_FRACTION;\n}\n\nfunction reviewLine(reason: string, text: string): string {\n return `> [review: ${reason}] ${text}`;\n}\n\n// One classified block, carrying constituent words for analyze()'s\n// block->word linkage; discriminated on `kind` so `reason`/`level` are compile-time guaranteed present on their variants.\nexport type EmitBlock =\n | { kind: 'heading'; text: string; level: number; words: Word[] }\n | { kind: 'paragraph'; text: string; words: Word[] }\n | { kind: 'quote'; text: string; words: Word[] }\n | {\n kind: 'review';\n text: string;\n /** The parseable reason inside `> [review: ...]`, carried on the block\n * so more reasons can be added without changing the render format. */\n reason: string;\n words: Word[];\n };\n\n// Renders one EmitBlock to the exact markdown emitPage used to push\n// directly — shared so extract.ts and analyze.ts produce byte-identical text.\nexport function renderEmitBlock(b: EmitBlock): string {\n switch (b.kind) {\n case 'heading':\n return `${'#'.repeat(b.level)} ${b.text}`;\n case 'quote':\n return `> ${b.text}`;\n case 'review':\n return reviewLine(b.reason, b.text);\n default:\n return b.text;\n }\n}\n\nexport function emitPage(p: PageOCR, levelOf: Map<number, number>, tuning: Tuning): EmitBlock[] {\n const blocks: EmitBlock[] = [];\n let para: OcrLine[] = [];\n const flush = () => {\n if (!para.length) return;\n const texts = para.map((l) => l.text);\n const merged = texts.slice(1).reduce(hyphenJoin, texts[0]);\n const words = para.flatMap((l) => l.words);\n // Same review-marker rule as the heading path below: no word may be\n // dropped or reordered here, only how the block renders changes.\n blocks.push(isUnstructured(merged) ? { kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words } : { kind: 'paragraph', text: merged, words });\n para = [];\n };\n\n let i = 0;\n while (i < p.lines.length) {\n const line: OcrLine = p.lines[i];\n const prev = i > 0 ? p.lines[i - 1] : null;\n if (prev && (line.col !== prev.col || prev.midY - line.midY > p.pitch * tuning.paraGap)) {\n // Join across a column break (instead of flushing) only when the column\n // changed, the paragraph dangles (no terminal punct), and the new text starts lowercase; a miss stays visible via pageStats.danglingLong.\n const danglingEnd = para.length > 0 && !/[.!?:;\"'”’)\\]]$/.test(para[para.length - 1].text);\n const continuationStart = /^[a-z]/.test(line.text);\n const joinAcrossColumns = line.col !== prev.col && danglingEnd && continuationStart && !levelOf.has(bucket(line.h));\n if (!joinAcrossColumns) flush();\n }\n\n if (levelOf.has(bucket(line.h))) {\n // Gather the run of consecutive heading-size lines: a wrapped title is\n // one heading, a long run is a pull quote in a bigger face.\n flush();\n const run: OcrLine[] = [line];\n let maxB = bucket(line.h);\n while (i + 1 < p.lines.length && levelOf.has(bucket(p.lines[i + 1].h)) && p.lines[i + 1].col === p.lines[i].col && p.lines[i].midY - p.lines[i + 1].midY < Math.max(p.lines[i].h, p.lines[i + 1].h) * 2.2) {\n i++;\n run.push(p.lines[i]);\n maxB = Math.max(maxB, bucket(p.lines[i].h));\n }\n const runTexts = run.map((l) => l.text);\n const merged = runTexts.slice(1).reduce(hyphenJoin, runTexts[0]);\n const words = run.flatMap((l) => l.words);\n // Below UNSTRUCTURED_WORDLIKE_FRACTION = chart labels/stray numerals,\n // not a heading; checked FIRST so a long numeral run isn't asserted as a pull quote.\n const junk = isUnstructured(merged);\n if (junk) blocks.push({ kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words });\n else if (merged.length > 150) blocks.push({ kind: 'quote', text: merged, words });\n else blocks.push({ kind: 'heading', text: merged, level: levelOf.get(maxB) as number, words });\n } else {\n para.push(line);\n }\n i++;\n }\n flush();\n return blocks;\n}\n"],"names":["getDict","bucket","hyphenJoin","acc","next","endsWith","withoutDash","slice","pen","at","test","filtered","head","split","pop","firstTok","filter","c","join","tail","close","has","toLowerCase","REVIEW_REASON","unstructuredLabels","UNSTRUCTURED_WORDLIKE_FRACTION","isUnstructured","text","tokens","Boolean","length","wordLike","t","reviewLine","reason","renderEmitBlock","b","kind","repeat","level","emitPage","p","levelOf","tuning","blocks","para","flush","texts","map","l","merged","reduce","words","flatMap","push","i","lines","line","prev","col","midY","pitch","paraGap","danglingEnd","continuationStart","joinAcrossColumns","h","run","maxB","Math","max","runTexts","junk","get"],"mappings":"AAAA,uDAAuD;AACvD,SAASA,OAAO,QAAQ,YAAY;AAEpC,SAASC,MAAM,QAAQ,gBAAgB;AAEvC,2EAA2E;AAC3E,+DAA+D;AAC/D,SAASC,WAAWC,GAAW,EAAEC,IAAY;IAC3C,IAAID,IAAIE,QAAQ,CAAC,MAAM;QACrB,MAAMC,cAAcH,IAAII,KAAK,CAAC,GAAG,CAAC;QAClC,MAAMC,MAAMF,YAAYG,EAAE,CAAC,CAAC;QAC5B,IAAID,OAAO,cAAcE,IAAI,CAACF,MAAM;gBACrBF,wBACIF,cAEJO;YAHb,MAAMC,QAAON,yBAAAA,YAAYO,KAAK,CAAC,KAAKC,GAAG,gBAA1BR,oCAAAA,yBAAgC;YAC7C,MAAMS,YAAWX,eAAAA,KAAKS,KAAK,CAAC,IAAI,CAAC,EAAE,cAAlBT,0BAAAA,eAAsB;YACvC,MAAMO,WAAW;mBAAII;aAAS,CAACC,MAAM,CAAC,CAACC,IAAM,YAAYP,IAAI,CAACO,IAAIC,IAAI,CAAC;YACvE,MAAMC,QAAOR,mBAAAA,SAASE,KAAK,CAAC,IAAI,CAAC,EAAE,cAAtBF,8BAAAA,mBAA0B;YACvC,MAAMS,QAAQpB,UAAUqB,GAAG,CAAC,AAACT,CAAAA,OAAOO,IAAG,EAAGG,WAAW;YACrD,OAAO,AAACF,CAAAA,QAAQd,cAAcH,GAAE,IAAKC;QACvC;IACF;IACA,OAAO,GAAGD,IAAI,CAAC,EAAEC,MAAM;AACzB;AAEA,8DAA8D;AAC9D,qGAAqG;AACrG,MAAMmB,gBAAgB;IACpBC,oBAAoB;AACtB;AAEA;qGACqG,GACrG,OAAO,MAAMC,iCAAiC,IAAI;AAClD,SAASC,eAAeC,IAAY;IAClC,MAAMC,SAASD,KAAKd,KAAK,CAAC,OAAOG,MAAM,CAACa;IACxC,IAAID,OAAOE,MAAM,KAAK,GAAG,OAAO,OAAO,4BAA4B;IACnE,MAAMC,WAAWH,OAAOZ,MAAM,CAAC,CAACgB,IAAM,cAActB,IAAI,CAACsB,IAAIF,MAAM;IACnE,OAAOC,WAAWH,OAAOE,MAAM,GAAGL;AACpC;AAEA,SAASQ,WAAWC,MAAc,EAAEP,IAAY;IAC9C,OAAO,CAAC,WAAW,EAAEO,OAAO,EAAE,EAAEP,MAAM;AACxC;AAiBA,oEAAoE;AACpE,8EAA8E;AAC9E,OAAO,SAASQ,gBAAgBC,CAAY;IAC1C,OAAQA,EAAEC,IAAI;QACZ,KAAK;YACH,OAAO,GAAG,IAAIC,MAAM,CAACF,EAAEG,KAAK,EAAE,CAAC,EAAEH,EAAET,IAAI,EAAE;QAC3C,KAAK;YACH,OAAO,CAAC,EAAE,EAAES,EAAET,IAAI,EAAE;QACtB,KAAK;YACH,OAAOM,WAAWG,EAAEF,MAAM,EAAEE,EAAET,IAAI;QACpC;YACE,OAAOS,EAAET,IAAI;IACjB;AACF;AAEA,OAAO,SAASa,SAASC,CAAU,EAAEC,OAA4B,EAAEC,MAAc;IAC/E,MAAMC,SAAsB,EAAE;IAC9B,IAAIC,OAAkB,EAAE;IACxB,MAAMC,QAAQ;QACZ,IAAI,CAACD,KAAKf,MAAM,EAAE;QAClB,MAAMiB,QAAQF,KAAKG,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;QACpC,MAAMuB,SAASH,MAAMxC,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAY6C,KAAK,CAAC,EAAE;QACzD,MAAMK,QAAQP,KAAKQ,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;QACzC,oEAAoE;QACpE,iEAAiE;QACjER,OAAOU,IAAI,CAAC5B,eAAewB,UAAU;YAAEb,MAAM;YAAUV,MAAMuB;YAAQhB,QAAQX,cAAcC,kBAAkB;YAAE4B;QAAM,IAAI;YAAEf,MAAM;YAAaV,MAAMuB;YAAQE;QAAM;QAClKP,OAAO,EAAE;IACX;IAEA,IAAIU,IAAI;IACR,MAAOA,IAAId,EAAEe,KAAK,CAAC1B,MAAM,CAAE;QACzB,MAAM2B,OAAgBhB,EAAEe,KAAK,CAACD,EAAE;QAChC,MAAMG,OAAOH,IAAI,IAAId,EAAEe,KAAK,CAACD,IAAI,EAAE,GAAG;QACtC,IAAIG,QAASD,CAAAA,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAID,KAAKE,IAAI,GAAGH,KAAKG,IAAI,GAAGnB,EAAEoB,KAAK,GAAGlB,OAAOmB,OAAO,AAAD,GAAI;YACvF,wEAAwE;YACxE,0IAA0I;YAC1I,MAAMC,cAAclB,KAAKf,MAAM,GAAG,KAAK,CAAC,kBAAkBpB,IAAI,CAACmC,IAAI,CAACA,KAAKf,MAAM,GAAG,EAAE,CAACH,IAAI;YACzF,MAAMqC,oBAAoB,SAAStD,IAAI,CAAC+C,KAAK9B,IAAI;YACjD,MAAMsC,oBAAoBR,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAII,eAAeC,qBAAqB,CAACtB,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC;YACjH,IAAI,CAACD,mBAAmBnB;QAC1B;QAEA,IAAIJ,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC,IAAI;YAC/B,uEAAuE;YACvE,4DAA4D;YAC5DpB;YACA,MAAMqB,MAAiB;gBAACV;aAAK;YAC7B,IAAIW,OAAOnE,OAAOwD,KAAKS,CAAC;YACxB,MAAOX,IAAI,IAAId,EAAEe,KAAK,CAAC1B,MAAM,IAAIY,QAAQrB,GAAG,CAACpB,OAAOwC,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,MAAMzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACI,GAAG,KAAKlB,EAAEe,KAAK,CAACD,EAAE,CAACI,GAAG,IAAIlB,EAAEe,KAAK,CAACD,EAAE,CAACK,IAAI,GAAGnB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACK,IAAI,GAAGS,KAAKC,GAAG,CAAC7B,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC,EAAEzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,IAAI,IAAK;gBACzMX;gBACAY,IAAIb,IAAI,CAACb,EAAEe,KAAK,CAACD,EAAE;gBACnBa,OAAOC,KAAKC,GAAG,CAACF,MAAMnE,OAAOwC,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC;YAC3C;YACA,MAAMK,WAAWJ,IAAInB,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;YACtC,MAAMuB,SAASqB,SAAShE,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAYqE,QAAQ,CAAC,EAAE;YAC/D,MAAMnB,QAAQe,IAAId,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;YACxC,sEAAsE;YACtE,qFAAqF;YACrF,MAAMoB,OAAO9C,eAAewB;YAC5B,IAAIsB,MAAM5B,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAUV,MAAMuB;gBAAQhB,QAAQX,cAAcC,kBAAkB;gBAAE4B;YAAM;iBACjG,IAAIF,OAAOpB,MAAM,GAAG,KAAKc,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAASV,MAAMuB;gBAAQE;YAAM;iBAC1ER,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAWV,MAAMuB;gBAAQX,OAAOG,QAAQ+B,GAAG,CAACL;gBAAiBhB;YAAM;QAC9F,OAAO;YACLP,KAAKS,IAAI,CAACG;QACZ;QACAF;IACF;IACAT;IACA,OAAOF;AACT"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/engines/tesseract.ts"],"sourcesContent":["// engines/tesseract.ts — THE OCR engine: tesseract.js on every platform.\n//\n// Seam contract: words come back in normalized 0-1 page coordinates, origin\n// bottom-left, y increasing upward — the y-flip happens here, and nowhere else.\n//\n// Pure WASM, in-process, no native dependency: no crash containment needed.\nimport { mkdirSync } from 'node:fs';\nimport os from 'node:os';\nimport path from 'node:path';\nimport type { Page as TesseractPage } from 'tesseract.js';\nimport { createWorker, PSM } from 'tesseract.js';\nimport type { PageWord, Word } from '../geometry.ts';\nimport { foldTypographic } from '../lines.ts';\nimport { awaitPdfOpen, openPdfForRender } from '../pdf-open.ts';\nimport { renderLoadedPage } from '../raster.ts';\nimport { resolvePackagePath } from '../resolve.ts';\n\n// 288 dpi: the render resolution the engine's word-overlap numbers were\n// measured at (test/fixtures/ocr-engine-poc/) — the known-good operating point.\nexport const OCR_RENDER_DPI = 288;\n\n// Chart pages starve PSM.AUTO (unrecognized text can't be marked for review).\n// Prose-fraction retry trigger, measured (sparse-retry.test.ts): chart 0.442 vs prose 0.767-0.938, cut 0.6.\nexport const SPARSE_RETRY_PROSE_FRACTION = 0.6;\n\n/** Share of `words` whose text contains a run of 3+ letters (`src/emit.ts`'s isUnstructured canary, applied per-page instead of per-block). */\nfunction proseFraction(words: { text: string }[]): number {\n if (words.length === 0) return 0;\n return words.filter((w) => /[A-Za-z]{3}/.test(w.text)).length / words.length;\n}\n\n/** Does this page's PSM.AUTO pass look chart-like enough to re-recognize\n * with PSM.SPARSE_TEXT? A page with 0 words never retries (nothing to take a prose fraction of). */\nexport function shouldRetrySparse(words: { text: string }[]): boolean {\n return words.length > 0 && proseFraction(words) < SPARSE_RETRY_PROSE_FRACTION;\n}\n\n// The locally installed traineddata (npm package @tesseract.js-data/eng),\n// resolved dynamically so depcheck won't flag it \"unused\" — it's a genuine runtime dependency, listed in .depcheckrc's `ignores`.\nconst TESSDATA_DIR = path.join(path.dirname(resolvePackagePath('@tesseract.js-data/eng/package.json')), '4.0.0');\n\n/** Per-page recognition progress, delivered once per page as it finishes.\n * Engine chatter goes through `onWarning`/`warnings` instead — this is timing/count data only. */\nexport type OcrProgressEvent = { type: 'page'; page: number; words: number; ms: number };\n\nexport interface RecognizeResult {\n words: PageWord[];\n /** pages (1-based) where the engine returned zero words */\n emptyPages: number[];\n /** wall-clock ms per page (render + recognize), keyed by page number */\n msPerPage: Map<number, number>;\n}\n\n/** tesseract.js's nested blocks/paragraphs/lines/words -> this repo's flat Word contract (normalize + y-flip). Same shape regardless of which PSM produced `page`, so both the AUTO pass and the SPARSE_TEXT retry share it. */\nfunction wordsFromPage(page: TesseractPage, width: number, height: number): Word[] {\n const pageWords: Word[] = [];\n for (const block of page.blocks ?? []) {\n for (const para of block.paragraphs ?? []) {\n for (const line of para.lines ?? []) {\n for (const w of line.words ?? []) {\n const { x0, y0, x1, y1 } = w.bbox;\n pageWords.push({\n // Same typographic folding as the text path (src/lines.ts) — both\n // paths must agree on how a printed apostrophe is spelled.\n text: foldTypographic(w.text),\n x: x0 / width,\n y: (height - y1) / height, // y-flip: top-left px -> bottom-left normalized\n w: (x1 - x0) / width,\n h: (y1 - y0) / height,\n // Plumbed through unmodified — tesseract.js's own 0-100\n // per-word confidence. See src/report.ts's LOW_CONFIDENCE_THRESHOLD.\n confidence: w.confidence,\n });\n }\n }\n }\n }\n return pageWords;\n}\n\n// tesseract.js-core's WASM build writes native diagnostics straight to this\n// process's stdout/stderr, with no override to silence them at the source.\n//\n// Suppression must span the whole run: worker-thread stdio arrives\n// asynchronously, on later event-loop turns; the synchronous carve-out below (`withRealStdio`) is the only safe exit.\nlet suppressDepth = 0;\nlet savedStdoutWrite: typeof process.stdout.write | null = null;\nlet savedStderrWrite: typeof process.stderr.write | null = null;\nasync function withSuppressedStdio<T>(fn: () => Promise<T>): Promise<T> {\n if (suppressDepth === 0) {\n savedStdoutWrite = process.stdout.write.bind(process.stdout);\n savedStderrWrite = process.stderr.write.bind(process.stderr);\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n suppressDepth++;\n try {\n return await fn();\n } finally {\n suppressDepth--;\n if (suppressDepth === 0) {\n if (savedStdoutWrite) process.stdout.write = savedStdoutWrite;\n if (savedStderrWrite) process.stderr.write = savedStderrWrite;\n savedStdoutWrite = null;\n savedStderrWrite = null;\n }\n }\n}\n\n/** Pierces an active `withSuppressedStdio` region SYNCHRONOUSLY to let this\n * module's own onProgress/onWarning calls through. No-op outside any suppressed region. */\nfunction withRealStdio<T>(fn: () => T): T {\n if (suppressDepth === 0) return fn();\n const realOut = savedStdoutWrite;\n const realErr = savedStderrWrite;\n if (realOut) process.stdout.write = realOut;\n if (realErr) process.stderr.write = realErr;\n try {\n return fn();\n } finally {\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n}\n\n/** Recognize words on pages [first..last] of a PDF (renders at `dpi`,\n * default OCR_RENDER_DPI). Progress via `onProgress`; diagnostics via `onWarning` — this module never writes to stdout/stderr itself. */\nexport async function recognizePdf(pdfPath: string, first: number, last: number, { dpi = OCR_RENDER_DPI, onProgress, onWarning }: { dpi?: number; onProgress?: (e: OcrProgressEvent) => void; onWarning?: (message: string) => void } = {}): Promise<RecognizeResult> {\n // tesseract.js writes its decompressed traineddata cache here by default,\n // dropping a model file into the user's cwd — redirect it under tmpdir() instead.\n const cachePath = path.join(os.tmpdir(), 'pdf-to-md-engine');\n mkdirSync(cachePath, { recursive: true });\n\n const words: PageWord[] = [];\n const emptyPages: number[] = [];\n const msPerPage = new Map<number, number>();\n\n // ONE continuous suppressed region for the whole run; onProgress/onWarning\n // below go through `withRealStdio`, the only sanctioned way out of it.\n await withSuppressedStdio(async () => {\n // OEM 1 (LSTM only, tesseract.js's own default): measured 99.0% recall\n // on the hard benchmark pages (2026-08-16).\n const worker = await createWorker('eng', 1, {\n langPath: TESSDATA_DIR,\n cachePath,\n });\n try {\n // NOT optional: tesseract.js's own default PSM (SINGLE_BLOCK) measured\n // 3-8 percentage points worse than AUTO on two-column pages.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n const loadingTask = openPdfForRender(pdfPath);\n try {\n const pdf = await awaitPdfOpen(loadingTask, pdfPath);\n for (let pn = first; pn <= Math.min(last, pdf.numPages); pn++) {\n const t0 = Date.now();\n const { png, width, height } = await renderLoadedPage(pdf, pn, dpi);\n const { data } = await worker.recognize(png, {}, { blocks: true });\n let pageWords: Word[] = wordsFromPage(data, width, height);\n\n // Adaptive second pass: re-recognizes the SAME rendered PNG, so\n // this costs one extra recognize() call only on chart-like pages.\n if (shouldRetrySparse(pageWords)) {\n const autoFraction = proseFraction(pageWords);\n const autoWords = pageWords;\n // setParameters on the shared worker is reversible, so one\n // in-process worker can alternate PSM per page instead of paying for a second worker.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.SPARSE_TEXT });\n const sparse = await worker.recognize(png, {}, { blocks: true });\n const sparseWords = wordsFromPage(sparse.data, width, height);\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n if (sparseWords.length > autoWords.length) {\n pageWords = sparseWords;\n const autoCount = autoWords.length;\n const sparseCount = sparseWords.length;\n withRealStdio(() => onWarning?.(`p${pn}: chart-like page (prose fraction ${autoFraction.toFixed(2)}) — retried with sparse-text segmentation, ${autoCount} -> ${sparseCount} words`));\n }\n }\n\n const ms = Date.now() - t0;\n for (const w of pageWords) words.push({ page: pn, ...w });\n if (pageWords.length === 0) emptyPages.push(pn);\n msPerPage.set(pn, ms);\n const wordCount = pageWords.length;\n withRealStdio(() => onProgress?.({ type: 'page', page: pn, words: wordCount, ms }));\n }\n } finally {\n await loadingTask.destroy();\n }\n } finally {\n await worker.terminate();\n }\n });\n\n return { words, emptyPages, msPerPage };\n}\n"],"names":["mkdirSync","os","path","createWorker","PSM","foldTypographic","awaitPdfOpen","openPdfForRender","renderLoadedPage","resolvePackagePath","OCR_RENDER_DPI","SPARSE_RETRY_PROSE_FRACTION","proseFraction","words","length","filter","w","test","text","shouldRetrySparse","TESSDATA_DIR","join","dirname","wordsFromPage","page","width","height","pageWords","block","blocks","para","paragraphs","line","lines","x0","y0","x1","y1","bbox","push","x","y","h","confidence","suppressDepth","savedStdoutWrite","savedStderrWrite","withSuppressedStdio","fn","process","stdout","write","bind","stderr","withRealStdio","realOut","realErr","recognizePdf","pdfPath","first","last","dpi","onProgress","onWarning","cachePath","tmpdir","recursive","emptyPages","msPerPage","Map","worker","langPath","setParameters","tessedit_pageseg_mode","AUTO","loadingTask","pdf","pn","Math","min","numPages","t0","Date","now","png","data","recognize","autoFraction","autoWords","SPARSE_TEXT","sparse","sparseWords","autoCount","sparseCount","toFixed","ms","set","wordCount","type","destroy","terminate"],"mappings":"AAAA,yEAAyE;AACzE,EAAE;AACF,4EAA4E;AAC5E,gFAAgF;AAChF,EAAE;AACF,4EAA4E;AAC5E,SAASA,SAAS,QAAQ,UAAU;AACpC,OAAOC,QAAQ,UAAU;AACzB,OAAOC,UAAU,YAAY;AAE7B,SAASC,YAAY,EAAEC,GAAG,QAAQ,eAAe;AAEjD,SAASC,eAAe,QAAQ,cAAc;AAC9C,SAASC,YAAY,EAAEC,gBAAgB,QAAQ,iBAAiB;AAChE,SAASC,gBAAgB,QAAQ,eAAe;AAChD,SAASC,kBAAkB,QAAQ,gBAAgB;AAEnD,wEAAwE;AACxE,gFAAgF;AAChF,OAAO,MAAMC,iBAAiB,IAAI;AAElC,8EAA8E;AAC9E,4GAA4G;AAC5G,OAAO,MAAMC,8BAA8B,IAAI;AAE/C,6IAA6I,GAC7I,SAASC,cAAcC,KAAyB;IAC9C,IAAIA,MAAMC,MAAM,KAAK,GAAG,OAAO;IAC/B,OAAOD,MAAME,MAAM,CAAC,CAACC,IAAM,cAAcC,IAAI,CAACD,EAAEE,IAAI,GAAGJ,MAAM,GAAGD,MAAMC,MAAM;AAC9E;AAEA;kGACkG,GAClG,OAAO,SAASK,kBAAkBN,KAAyB;IACzD,OAAOA,MAAMC,MAAM,GAAG,KAAKF,cAAcC,SAASF;AACpD;AAEA,0EAA0E;AAC1E,kIAAkI;AAClI,MAAMS,eAAelB,KAAKmB,IAAI,CAACnB,KAAKoB,OAAO,CAACb,mBAAmB,yCAAyC;AAcxG,8NAA8N,GAC9N,SAASc,cAAcC,IAAmB,EAAEC,KAAa,EAAEC,MAAc;QAEnDF;IADpB,MAAMG,YAAoB,EAAE;IAC5B,KAAK,MAAMC,UAASJ,eAAAA,KAAKK,MAAM,cAAXL,0BAAAA,eAAe,EAAE,CAAE;YAClBI;QAAnB,KAAK,MAAME,SAAQF,oBAAAA,MAAMG,UAAU,cAAhBH,+BAAAA,oBAAoB,EAAE,CAAE;gBACtBE;YAAnB,KAAK,MAAME,SAAQF,cAAAA,KAAKG,KAAK,cAAVH,yBAAAA,cAAc,EAAE,CAAE;oBACnBE;gBAAhB,KAAK,MAAMhB,MAAKgB,cAAAA,KAAKnB,KAAK,cAAVmB,yBAAAA,cAAc,EAAE,CAAE;oBAChC,MAAM,EAAEE,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAE,GAAGrB,EAAEsB,IAAI;oBACjCX,UAAUY,IAAI,CAAC;wBACb,kEAAkE;wBAClE,2DAA2D;wBAC3DrB,MAAMb,gBAAgBW,EAAEE,IAAI;wBAC5BsB,GAAGN,KAAKT;wBACRgB,GAAG,AAACf,CAAAA,SAASW,EAAC,IAAKX;wBACnBV,GAAG,AAACoB,CAAAA,KAAKF,EAAC,IAAKT;wBACfiB,GAAG,AAACL,CAAAA,KAAKF,EAAC,IAAKT;wBACf,wDAAwD;wBACxD,qEAAqE;wBACrEiB,YAAY3B,EAAE2B,UAAU;oBAC1B;gBACF;YACF;QACF;IACF;IACA,OAAOhB;AACT;AAEA,4EAA4E;AAC5E,2EAA2E;AAC3E,EAAE;AACF,mEAAmE;AACnE,sHAAsH;AACtH,IAAIiB,gBAAgB;AACpB,IAAIC,mBAAuD;AAC3D,IAAIC,mBAAuD;AAC3D,eAAeC,oBAAuBC,EAAoB;IACxD,IAAIJ,kBAAkB,GAAG;QACvBC,mBAAmBI,QAAQC,MAAM,CAACC,KAAK,CAACC,IAAI,CAACH,QAAQC,MAAM;QAC3DJ,mBAAmBG,QAAQI,MAAM,CAACF,KAAK,CAACC,IAAI,CAACH,QAAQI,MAAM;QAC3DJ,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;IACAP;IACA,IAAI;QACF,OAAO,MAAMI;IACf,SAAU;QACRJ;QACA,IAAIA,kBAAkB,GAAG;YACvB,IAAIC,kBAAkBI,QAAQC,MAAM,CAACC,KAAK,GAAGN;YAC7C,IAAIC,kBAAkBG,QAAQI,MAAM,CAACF,KAAK,GAAGL;YAC7CD,mBAAmB;YACnBC,mBAAmB;QACrB;IACF;AACF;AAEA;yFACyF,GACzF,SAASQ,cAAiBN,EAAW;IACnC,IAAIJ,kBAAkB,GAAG,OAAOI;IAChC,MAAMO,UAAUV;IAChB,MAAMW,UAAUV;IAChB,IAAIS,SAASN,QAAQC,MAAM,CAACC,KAAK,GAAGI;IACpC,IAAIC,SAASP,QAAQI,MAAM,CAACF,KAAK,GAAGK;IACpC,IAAI;QACF,OAAOR;IACT,SAAU;QACRC,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;AACF;AAEA;uIACuI,GACvI,OAAO,eAAeM,aAAaC,OAAe,EAAEC,KAAa,EAAEC,IAAY,EAAE,EAAEC,MAAMnD,cAAc,EAAEoD,UAAU,EAAEC,SAAS,EAAuG,GAAG,CAAC,CAAC;IACxO,0EAA0E;IAC1E,kFAAkF;IAClF,MAAMC,YAAY9D,KAAKmB,IAAI,CAACpB,GAAGgE,MAAM,IAAI;IACzCjE,UAAUgE,WAAW;QAAEE,WAAW;IAAK;IAEvC,MAAMrD,QAAoB,EAAE;IAC5B,MAAMsD,aAAuB,EAAE;IAC/B,MAAMC,YAAY,IAAIC;IAEtB,2EAA2E;IAC3E,uEAAuE;IACvE,MAAMtB,oBAAoB;QACxB,uEAAuE;QACvE,4CAA4C;QAC5C,MAAMuB,SAAS,MAAMnE,aAAa,OAAO,GAAG;YAC1CoE,UAAUnD;YACV4C;QACF;QACA,IAAI;YACF,uEAAuE;YACvE,6DAA6D;YAC7D,MAAMM,OAAOE,aAAa,CAAC;gBAAEC,uBAAuBrE,IAAIsE,IAAI;YAAC;YAE7D,MAAMC,cAAcpE,iBAAiBmD;YACrC,IAAI;gBACF,MAAMkB,MAAM,MAAMtE,aAAaqE,aAAajB;gBAC5C,IAAK,IAAImB,KAAKlB,OAAOkB,MAAMC,KAAKC,GAAG,CAACnB,MAAMgB,IAAII,QAAQ,GAAGH,KAAM;oBAC7D,MAAMI,KAAKC,KAAKC,GAAG;oBACnB,MAAM,EAAEC,GAAG,EAAE3D,KAAK,EAAEC,MAAM,EAAE,GAAG,MAAMlB,iBAAiBoE,KAAKC,IAAIhB;oBAC/D,MAAM,EAAEwB,IAAI,EAAE,GAAG,MAAMf,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;wBAAEvD,QAAQ;oBAAK;oBAChE,IAAIF,YAAoBJ,cAAc8D,MAAM5D,OAAOC;oBAEnD,gEAAgE;oBAChE,kEAAkE;oBAClE,IAAIP,kBAAkBQ,YAAY;wBAChC,MAAM4D,eAAe3E,cAAce;wBACnC,MAAM6D,YAAY7D;wBAClB,2DAA2D;wBAC3D,sFAAsF;wBACtF,MAAM2C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIqF,WAAW;wBAAC;wBACpE,MAAMC,SAAS,MAAMpB,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;4BAAEvD,QAAQ;wBAAK;wBAC9D,MAAM8D,cAAcpE,cAAcmE,OAAOL,IAAI,EAAE5D,OAAOC;wBACtD,MAAM4C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIsE,IAAI;wBAAC;wBAE7D,IAAIiB,YAAY7E,MAAM,GAAG0E,UAAU1E,MAAM,EAAE;4BACzCa,YAAYgE;4BACZ,MAAMC,YAAYJ,UAAU1E,MAAM;4BAClC,MAAM+E,cAAcF,YAAY7E,MAAM;4BACtCwC,cAAc,IAAMS,sBAAAA,gCAAAA,UAAY,CAAC,CAAC,EAAEc,GAAG,kCAAkC,EAAEU,aAAaO,OAAO,CAAC,GAAG,2CAA2C,EAAEF,UAAU,IAAI,EAAEC,YAAY,MAAM,CAAC;wBACrL;oBACF;oBAEA,MAAME,KAAKb,KAAKC,GAAG,KAAKF;oBACxB,KAAK,MAAMjE,KAAKW,UAAWd,MAAM0B,IAAI,CAAC;wBAAEf,MAAMqD;wBAAI,GAAG7D,CAAC;oBAAC;oBACvD,IAAIW,UAAUb,MAAM,KAAK,GAAGqD,WAAW5B,IAAI,CAACsC;oBAC5CT,UAAU4B,GAAG,CAACnB,IAAIkB;oBAClB,MAAME,YAAYtE,UAAUb,MAAM;oBAClCwC,cAAc,IAAMQ,uBAAAA,iCAAAA,WAAa;4BAAEoC,MAAM;4BAAQ1E,MAAMqD;4BAAIhE,OAAOoF;4BAAWF;wBAAG;gBAClF;YACF,SAAU;gBACR,MAAMpB,YAAYwB,OAAO;YAC3B;QACF,SAAU;YACR,MAAM7B,OAAO8B,SAAS;QACxB;IACF;IAEA,OAAO;QAAEvF;QAAOsD;QAAYC;IAAU;AACxC"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/open-source/formats/pdf-to-md/src/engines/tesseract.ts"],"sourcesContent":["// engines/tesseract.ts — THE OCR engine: tesseract.js on every platform.\n//\n// Seam contract: words come back in normalized 0-1 page coordinates, origin\n// bottom-left, y increasing upward — the y-flip happens here, and nowhere else.\n//\n// Pure WASM, in-process, no native dependency: no crash containment needed.\nimport { mkdirSync } from 'node:fs';\nimport os from 'node:os';\nimport path from 'node:path';\nimport type { Page as TesseractPage } from 'tesseract.js';\nimport { createWorker, PSM } from 'tesseract.js';\nimport type { PageWord, Word } from '../geometry.ts';\nimport { foldTypographic } from '../lines.ts';\nimport { awaitPdfOpen, openPdfForRender } from '../pdf-open.ts';\nimport { renderLoadedPage } from '../raster.ts';\nimport { resolvePackagePath } from '../resolve.ts';\n\n// 288 dpi: the render resolution the engine's word-overlap numbers were\n// measured at (test/fixtures/ocr-engine-poc/) — the known-good operating point.\nexport const OCR_RENDER_DPI = 288;\n\n// Chart pages starve PSM.AUTO (unrecognized text can't be marked for review).\n// Prose-fraction retry trigger, measured (sparse-retry.test.ts): chart 0.442 vs prose 0.767-0.938, cut 0.6.\nexport const SPARSE_RETRY_PROSE_FRACTION = 0.6;\n\n/** Share of `words` whose text contains a run of 3+ letters (`src/emit.ts`'s isUnstructured canary, applied per-page instead of per-block). */\nfunction proseFraction(words: { text: string }[]): number {\n if (words.length === 0) return 0;\n return words.filter((w) => /[A-Za-z]{3}/.test(w.text)).length / words.length;\n}\n\n/** Does this page's PSM.AUTO pass look chart-like enough to re-recognize\n * with PSM.SPARSE_TEXT? A page with 0 words never retries (nothing to take a prose fraction of). */\nexport function shouldRetrySparse(words: { text: string }[]): boolean {\n return words.length > 0 && proseFraction(words) < SPARSE_RETRY_PROSE_FRACTION;\n}\n\n// The locally installed traineddata (npm package @tesseract.js-data/eng),\n// resolved dynamically so depcheck won't flag it \"unused\" — it's a genuine runtime dependency, listed in .depcheckrc's `ignores`.\nconst TESSDATA_DIR = path.join(path.dirname(resolvePackagePath('@tesseract.js-data/eng/package.json')), '4.0.0');\n\n/** Per-page recognition progress, delivered once per page as it finishes.\n * Engine chatter goes through `onWarning`/`warnings` instead — this is timing/count data only. */\nexport type OcrProgressEvent = { type: 'page'; page: number; words: number; ms: number };\n\nexport interface RecognizeResult {\n words: PageWord[];\n /** pages (1-based) where the engine returned zero words */\n emptyPages: number[];\n /** wall-clock ms per page (render + recognize), keyed by page number */\n msPerPage: Map<number, number>;\n}\n\n/** tesseract.js's nested blocks/paragraphs/lines/words -> this repo's flat Word contract (normalize + y-flip). Same shape regardless of which PSM produced `page`, so both the AUTO pass and the SPARSE_TEXT retry share it. */\nfunction wordsFromPage(page: TesseractPage, width: number, height: number): Word[] {\n const pageWords: Word[] = [];\n for (const block of page.blocks ?? []) {\n for (const para of block.paragraphs ?? []) {\n for (const line of para.lines ?? []) {\n for (const w of line.words ?? []) {\n const { x0, y0, x1, y1 } = w.bbox;\n pageWords.push({\n // Same typographic folding as the text path (src/lines.ts) — both\n // paths must agree on how a printed apostrophe is spelled.\n text: foldTypographic(w.text),\n x: x0 / width,\n y: (height - y1) / height, // y-flip: top-left px -> bottom-left normalized\n w: (x1 - x0) / width,\n h: (y1 - y0) / height,\n // Plumbed through unmodified — tesseract.js's own 0-100\n // per-word confidence. See src/report.ts's LOW_CONFIDENCE_THRESHOLD.\n confidence: w.confidence,\n });\n }\n }\n }\n }\n return pageWords;\n}\n\n// tesseract.js-core's WASM build writes native diagnostics straight to this\n// process's stdout/stderr, with no override to silence them at the source.\n//\n// Suppression must span the whole run: worker-thread stdio arrives\n// asynchronously, on later event-loop turns; the synchronous carve-out below (`withRealStdio`) is the only safe exit.\nlet suppressDepth = 0;\nlet savedStdoutWrite: typeof process.stdout.write | null = null;\nlet savedStderrWrite: typeof process.stderr.write | null = null;\nasync function withSuppressedStdio<T>(fn: () => Promise<T>): Promise<T> {\n if (suppressDepth === 0) {\n savedStdoutWrite = process.stdout.write.bind(process.stdout);\n savedStderrWrite = process.stderr.write.bind(process.stderr);\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n suppressDepth++;\n try {\n return await fn();\n } finally {\n suppressDepth--;\n if (suppressDepth === 0) {\n if (savedStdoutWrite) process.stdout.write = savedStdoutWrite;\n if (savedStderrWrite) process.stderr.write = savedStderrWrite;\n savedStdoutWrite = null;\n savedStderrWrite = null;\n }\n }\n}\n\n/** Pierces an active `withSuppressedStdio` region SYNCHRONOUSLY to let this\n * module's own onProgress/onWarning calls through. No-op outside any suppressed region. */\nfunction withRealStdio<T>(fn: () => T): T {\n if (suppressDepth === 0) return fn();\n const realOut = savedStdoutWrite;\n const realErr = savedStderrWrite;\n if (realOut) process.stdout.write = realOut;\n if (realErr) process.stderr.write = realErr;\n try {\n return fn();\n } finally {\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n}\n\n/** Recognize words on pages [first..last] of a PDF (renders at `dpi`,\n * default OCR_RENDER_DPI). Progress via `onProgress`; diagnostics via `onWarning` — this module never writes to stdout/stderr itself. */\nexport async function recognizePdf(pdfPath: string, first: number, last: number, { dpi = OCR_RENDER_DPI, onProgress, onWarning }: { dpi?: number; onProgress?: (e: OcrProgressEvent) => void; onWarning?: (message: string) => void } = {}): Promise<RecognizeResult> {\n // tesseract.js writes its decompressed traineddata cache here by default,\n // dropping a model file into the user's cwd — redirect it under tmpdir() instead.\n const cachePath = path.join(os.tmpdir(), 'pdf-to-md-engine');\n mkdirSync(cachePath, { recursive: true });\n\n const words: PageWord[] = [];\n const emptyPages: number[] = [];\n const msPerPage = new Map<number, number>();\n\n // ONE continuous suppressed region for the whole run; onProgress/onWarning\n // below go through `withRealStdio`, the only sanctioned way out of it.\n await withSuppressedStdio(async () => {\n // OEM 1 (LSTM only, tesseract.js's own default): measured 99.0% recall\n // on the hard benchmark pages (2026-08-16).\n const worker = await createWorker('eng', 1, {\n langPath: TESSDATA_DIR,\n cachePath,\n });\n try {\n // NOT optional: tesseract.js's own default PSM (SINGLE_BLOCK) measured\n // 3-8 percentage points worse than AUTO on two-column pages.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n const loadingTask = openPdfForRender(pdfPath);\n try {\n const pdf = await awaitPdfOpen(loadingTask, pdfPath);\n for (let pn = first; pn <= Math.min(last, pdf.numPages); pn++) {\n const t0 = Date.now();\n const { png, width, height } = await renderLoadedPage(pdf, pn, dpi);\n const { data } = await worker.recognize(png, {}, { blocks: true });\n let pageWords: Word[] = wordsFromPage(data, width, height);\n\n // Adaptive second pass: re-recognizes the SAME rendered PNG, so\n // this costs one extra recognize() call only on chart-like pages.\n if (shouldRetrySparse(pageWords)) {\n const autoFraction = proseFraction(pageWords);\n const autoWords = pageWords;\n // setParameters on the shared worker is reversible, so one\n // in-process worker can alternate PSM per page instead of paying for a second worker.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.SPARSE_TEXT });\n const sparse = await worker.recognize(png, {}, { blocks: true });\n const sparseWords = wordsFromPage(sparse.data, width, height);\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n if (sparseWords.length > autoWords.length) {\n pageWords = sparseWords;\n const autoCount = autoWords.length;\n const sparseCount = sparseWords.length;\n withRealStdio(() => onWarning?.(`p${pn}: chart-like page (prose fraction ${autoFraction.toFixed(2)}) — retried with sparse-text segmentation, ${autoCount} -> ${sparseCount} words`));\n }\n }\n\n const ms = Date.now() - t0;\n for (const w of pageWords) words.push({ page: pn, ...w });\n if (pageWords.length === 0) emptyPages.push(pn);\n msPerPage.set(pn, ms);\n const wordCount = pageWords.length;\n withRealStdio(() => onProgress?.({ type: 'page', page: pn, words: wordCount, ms }));\n }\n } finally {\n await loadingTask.destroy();\n }\n } finally {\n await worker.terminate();\n }\n });\n\n return { words, emptyPages, msPerPage };\n}\n"],"names":["mkdirSync","os","path","createWorker","PSM","foldTypographic","awaitPdfOpen","openPdfForRender","renderLoadedPage","resolvePackagePath","OCR_RENDER_DPI","SPARSE_RETRY_PROSE_FRACTION","proseFraction","words","length","filter","w","test","text","shouldRetrySparse","TESSDATA_DIR","join","dirname","wordsFromPage","page","width","height","pageWords","block","blocks","para","paragraphs","line","lines","x0","y0","x1","y1","bbox","push","x","y","h","confidence","suppressDepth","savedStdoutWrite","savedStderrWrite","withSuppressedStdio","fn","process","stdout","write","bind","stderr","withRealStdio","realOut","realErr","recognizePdf","pdfPath","first","last","dpi","onProgress","onWarning","cachePath","tmpdir","recursive","emptyPages","msPerPage","Map","worker","langPath","setParameters","tessedit_pageseg_mode","AUTO","loadingTask","pdf","pn","Math","min","numPages","t0","Date","now","png","data","recognize","autoFraction","autoWords","SPARSE_TEXT","sparse","sparseWords","autoCount","sparseCount","toFixed","ms","set","wordCount","type","destroy","terminate"],"mappings":"AAAA,yEAAyE;AACzE,EAAE;AACF,4EAA4E;AAC5E,gFAAgF;AAChF,EAAE;AACF,4EAA4E;AAC5E,SAASA,SAAS,QAAQ,UAAU;AACpC,OAAOC,QAAQ,UAAU;AACzB,OAAOC,UAAU,YAAY;AAE7B,SAASC,YAAY,EAAEC,GAAG,QAAQ,eAAe;AAEjD,SAASC,eAAe,QAAQ,cAAc;AAC9C,SAASC,YAAY,EAAEC,gBAAgB,QAAQ,iBAAiB;AAChE,SAASC,gBAAgB,QAAQ,eAAe;AAChD,SAASC,kBAAkB,QAAQ,gBAAgB;AAEnD,wEAAwE;AACxE,gFAAgF;AAChF,OAAO,MAAMC,iBAAiB,IAAI;AAElC,8EAA8E;AAC9E,4GAA4G;AAC5G,OAAO,MAAMC,8BAA8B,IAAI;AAE/C,6IAA6I,GAC7I,SAASC,cAAcC,KAAyB;IAC9C,IAAIA,MAAMC,MAAM,KAAK,GAAG,OAAO;IAC/B,OAAOD,MAAME,MAAM,CAAC,CAACC,IAAM,cAAcC,IAAI,CAACD,EAAEE,IAAI,GAAGJ,MAAM,GAAGD,MAAMC,MAAM;AAC9E;AAEA;kGACkG,GAClG,OAAO,SAASK,kBAAkBN,KAAyB;IACzD,OAAOA,MAAMC,MAAM,GAAG,KAAKF,cAAcC,SAASF;AACpD;AAEA,0EAA0E;AAC1E,kIAAkI;AAClI,MAAMS,eAAelB,KAAKmB,IAAI,CAACnB,KAAKoB,OAAO,CAACb,mBAAmB,yCAAyC;AAcxG,8NAA8N,GAC9N,SAASc,cAAcC,IAAmB,EAAEC,KAAa,EAAEC,MAAc;QAEnDF;IADpB,MAAMG,YAAoB,EAAE;IAC5B,KAAK,MAAMC,UAASJ,eAAAA,KAAKK,MAAM,cAAXL,0BAAAA,eAAe,EAAE,CAAE;YAClBI;QAAnB,KAAK,MAAME,SAAQF,oBAAAA,MAAMG,UAAU,cAAhBH,+BAAAA,oBAAoB,EAAE,CAAE;gBACtBE;YAAnB,KAAK,MAAME,SAAQF,cAAAA,KAAKG,KAAK,cAAVH,yBAAAA,cAAc,EAAE,CAAE;oBACnBE;gBAAhB,KAAK,MAAMhB,MAAKgB,cAAAA,KAAKnB,KAAK,cAAVmB,yBAAAA,cAAc,EAAE,CAAE;oBAChC,MAAM,EAAEE,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAE,GAAGrB,EAAEsB,IAAI;oBACjCX,UAAUY,IAAI,CAAC;wBACb,kEAAkE;wBAClE,2DAA2D;wBAC3DrB,MAAMb,gBAAgBW,EAAEE,IAAI;wBAC5BsB,GAAGN,KAAKT;wBACRgB,GAAG,AAACf,CAAAA,SAASW,EAAC,IAAKX;wBACnBV,GAAG,AAACoB,CAAAA,KAAKF,EAAC,IAAKT;wBACfiB,GAAG,AAACL,CAAAA,KAAKF,EAAC,IAAKT;wBACf,wDAAwD;wBACxD,qEAAqE;wBACrEiB,YAAY3B,EAAE2B,UAAU;oBAC1B;gBACF;YACF;QACF;IACF;IACA,OAAOhB;AACT;AAEA,4EAA4E;AAC5E,2EAA2E;AAC3E,EAAE;AACF,mEAAmE;AACnE,sHAAsH;AACtH,IAAIiB,gBAAgB;AACpB,IAAIC,mBAAuD;AAC3D,IAAIC,mBAAuD;AAC3D,eAAeC,oBAAuBC,EAAoB;IACxD,IAAIJ,kBAAkB,GAAG;QACvBC,mBAAmBI,QAAQC,MAAM,CAACC,KAAK,CAACC,IAAI,CAACH,QAAQC,MAAM;QAC3DJ,mBAAmBG,QAAQI,MAAM,CAACF,KAAK,CAACC,IAAI,CAACH,QAAQI,MAAM;QAC3DJ,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;IACAP;IACA,IAAI;QACF,OAAO,MAAMI;IACf,SAAU;QACRJ;QACA,IAAIA,kBAAkB,GAAG;YACvB,IAAIC,kBAAkBI,QAAQC,MAAM,CAACC,KAAK,GAAGN;YAC7C,IAAIC,kBAAkBG,QAAQI,MAAM,CAACF,KAAK,GAAGL;YAC7CD,mBAAmB;YACnBC,mBAAmB;QACrB;IACF;AACF;AAEA;yFACyF,GACzF,SAASQ,cAAiBN,EAAW;IACnC,IAAIJ,kBAAkB,GAAG,OAAOI;IAChC,MAAMO,UAAUV;IAChB,MAAMW,UAAUV;IAChB,IAAIS,SAASN,QAAQC,MAAM,CAACC,KAAK,GAAGI;IACpC,IAAIC,SAASP,QAAQI,MAAM,CAACF,KAAK,GAAGK;IACpC,IAAI;QACF,OAAOR;IACT,SAAU;QACRC,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;AACF;AAEA;uIACuI,GACvI,OAAO,eAAeM,aAAaC,OAAe,EAAEC,KAAa,EAAEC,IAAY,EAAE,EAAEC,MAAMnD,cAAc,EAAEoD,UAAU,EAAEC,SAAS,EAAuG,GAAG,CAAC,CAAC;IACxO,0EAA0E;IAC1E,kFAAkF;IAClF,MAAMC,YAAY9D,KAAKmB,IAAI,CAACpB,GAAGgE,MAAM,IAAI;IACzCjE,UAAUgE,WAAW;QAAEE,WAAW;IAAK;IAEvC,MAAMrD,QAAoB,EAAE;IAC5B,MAAMsD,aAAuB,EAAE;IAC/B,MAAMC,YAAY,IAAIC;IAEtB,2EAA2E;IAC3E,uEAAuE;IACvE,MAAMtB,oBAAoB;QACxB,uEAAuE;QACvE,4CAA4C;QAC5C,MAAMuB,SAAS,MAAMnE,aAAa,OAAO,GAAG;YAC1CoE,UAAUnD;YACV4C;QACF;QACA,IAAI;YACF,uEAAuE;YACvE,6DAA6D;YAC7D,MAAMM,OAAOE,aAAa,CAAC;gBAAEC,uBAAuBrE,IAAIsE,IAAI;YAAC;YAE7D,MAAMC,cAAcpE,iBAAiBmD;YACrC,IAAI;gBACF,MAAMkB,MAAM,MAAMtE,aAAaqE,aAAajB;gBAC5C,IAAK,IAAImB,KAAKlB,OAAOkB,MAAMC,KAAKC,GAAG,CAACnB,MAAMgB,IAAII,QAAQ,GAAGH,KAAM;oBAC7D,MAAMI,KAAKC,KAAKC,GAAG;oBACnB,MAAM,EAAEC,GAAG,EAAE3D,KAAK,EAAEC,MAAM,EAAE,GAAG,MAAMlB,iBAAiBoE,KAAKC,IAAIhB;oBAC/D,MAAM,EAAEwB,IAAI,EAAE,GAAG,MAAMf,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;wBAAEvD,QAAQ;oBAAK;oBAChE,IAAIF,YAAoBJ,cAAc8D,MAAM5D,OAAOC;oBAEnD,gEAAgE;oBAChE,kEAAkE;oBAClE,IAAIP,kBAAkBQ,YAAY;wBAChC,MAAM4D,eAAe3E,cAAce;wBACnC,MAAM6D,YAAY7D;wBAClB,2DAA2D;wBAC3D,sFAAsF;wBACtF,MAAM2C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIqF,WAAW;wBAAC;wBACpE,MAAMC,SAAS,MAAMpB,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;4BAAEvD,QAAQ;wBAAK;wBAC9D,MAAM8D,cAAcpE,cAAcmE,OAAOL,IAAI,EAAE5D,OAAOC;wBACtD,MAAM4C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIsE,IAAI;wBAAC;wBAE7D,IAAIiB,YAAY7E,MAAM,GAAG0E,UAAU1E,MAAM,EAAE;4BACzCa,YAAYgE;4BACZ,MAAMC,YAAYJ,UAAU1E,MAAM;4BAClC,MAAM+E,cAAcF,YAAY7E,MAAM;4BACtCwC,cAAc,IAAMS,sBAAAA,gCAAAA,UAAY,CAAC,CAAC,EAAEc,GAAG,kCAAkC,EAAEU,aAAaO,OAAO,CAAC,GAAG,2CAA2C,EAAEF,UAAU,IAAI,EAAEC,YAAY,MAAM,CAAC;wBACrL;oBACF;oBAEA,MAAME,KAAKb,KAAKC,GAAG,KAAKF;oBACxB,KAAK,MAAMjE,KAAKW,UAAWd,MAAM0B,IAAI,CAAC;wBAAEf,MAAMqD;wBAAI,GAAG7D,CAAC;oBAAC;oBACvD,IAAIW,UAAUb,MAAM,KAAK,GAAGqD,WAAW5B,IAAI,CAACsC;oBAC5CT,UAAU4B,GAAG,CAACnB,IAAIkB;oBAClB,MAAME,YAAYtE,UAAUb,MAAM;oBAClCwC,cAAc,IAAMQ,uBAAAA,iCAAAA,WAAa;4BAAEoC,MAAM;4BAAQ1E,MAAMqD;4BAAIhE,OAAOoF;4BAAWF;wBAAG;gBAClF;YACF,SAAU;gBACR,MAAMpB,YAAYwB,OAAO;YAC3B;QACF,SAAU;YACR,MAAM7B,OAAO8B,SAAS;QACxB;IACF;IAEA,OAAO;QAAEvF;QAAOsD;QAAYC;IAAU;AACxC"}
@@ -11,6 +11,8 @@ export type PdfToMdErrorCode =
11
11
  | 'OCR_PAGE_FAILED'
12
12
  /** malformed word-dump input (--words-json / extractOcr words) */
13
13
  | 'WORDS_INPUT'
14
+ /** malformed or insufficient evidence passed to auditWords */
15
+ | 'AUDIT_INPUT'
14
16
  /** malformed Analysis handed to toMarkdown/toText (e.g. an out-of-range
15
17
  * wordIndexes entry) — validated fail-fast, whole object checked before any output. */
16
18
  | 'ANALYSIS_INPUT';
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/errors.ts"],"sourcesContent":["// Library code throws; only src/cli.ts prints and exits.\n\n/** The closed set of error codes every `PdfToMdError` carries — switch on\n * `code` instead of parsing `message`. */\nexport type PdfToMdErrorCode =\n /** input missing/unreadable/not a PDF */\n | 'PDF_OPEN'\n /** malformed page selection (validated before any work) */\n | 'PAGE_RANGE'\n /** rendering an opened, in-range page to PNG failed */\n | 'PAGE_RENDER'\n /** the OCR engine crashed twice on one page (message names it) */\n | 'OCR_PAGE_FAILED'\n /** malformed word-dump input (--words-json / extractOcr words) */\n | 'WORDS_INPUT'\n /** malformed Analysis handed to toMarkdown/toText (e.g. an out-of-range\n * wordIndexes entry) — validated fail-fast, whole object checked before any output. */\n | 'ANALYSIS_INPUT';\n\n/** The error every entry point throws — never a bare `Error`. */\nexport class PdfToMdError extends Error {\n code: PdfToMdErrorCode;\n\n constructor(code: PdfToMdErrorCode, message: string) {\n super(message);\n this.name = 'PdfToMdError';\n this.code = code;\n }\n}\n"],"names":["PdfToMdError","Error","code","message","name"],"mappings":"AAAA,yDAAyD;AAEzD;wCACwC,GAgBxC,+DAA+D,GAC/D,OAAO,MAAMA,qBAAqBC;IAGhC,YAAYC,IAAsB,EAAEC,OAAe,CAAE;QACnD,KAAK,CAACA;QACN,IAAI,CAACC,IAAI,GAAG;QACZ,IAAI,CAACF,IAAI,GAAGA;IACd;AACF"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/open-source/formats/pdf-to-md/src/errors.ts"],"sourcesContent":["// Library code throws; only src/cli.ts prints and exits.\n\n/** The closed set of error codes every `PdfToMdError` carries — switch on\n * `code` instead of parsing `message`. */\nexport type PdfToMdErrorCode =\n /** input missing/unreadable/not a PDF */\n | 'PDF_OPEN'\n /** malformed page selection (validated before any work) */\n | 'PAGE_RANGE'\n /** rendering an opened, in-range page to PNG failed */\n | 'PAGE_RENDER'\n /** the OCR engine crashed twice on one page (message names it) */\n | 'OCR_PAGE_FAILED'\n /** malformed word-dump input (--words-json / extractOcr words) */\n | 'WORDS_INPUT'\n /** malformed or insufficient evidence passed to auditWords */\n | 'AUDIT_INPUT'\n /** malformed Analysis handed to toMarkdown/toText (e.g. an out-of-range\n * wordIndexes entry) — validated fail-fast, whole object checked before any output. */\n | 'ANALYSIS_INPUT';\n\n/** The error every entry point throws — never a bare `Error`. */\nexport class PdfToMdError extends Error {\n code: PdfToMdErrorCode;\n\n constructor(code: PdfToMdErrorCode, message: string) {\n super(message);\n this.name = 'PdfToMdError';\n this.code = code;\n }\n}\n"],"names":["PdfToMdError","Error","code","message","name"],"mappings":"AAAA,yDAAyD;AAEzD;wCACwC,GAkBxC,+DAA+D,GAC/D,OAAO,MAAMA,qBAAqBC;IAGhC,YAAYC,IAAsB,EAAEC,OAAe,CAAE;QACnD,KAAK,CAACA;QACN,IAAI,CAACC,IAAI,GAAG;QACZ,IAAI,CAACF,IAAI,GAAGA;IACd;AACF"}
@@ -1,6 +1,6 @@
1
1
  import type { OcrProgressEvent } from './engines/tesseract.js';
2
2
  import type { PageWord, Tuning } from './geometry.js';
3
- import type { Report } from './types.js';
3
+ import type { Analysis, Report } from './types.js';
4
4
  export type { OcrProgressEvent } from './engines/tesseract.js';
5
5
  /** One recognized word with its page — the OCR replay input contract (same
6
6
  * shape as `geometry.ts`'s `PageWord`, re-exported under this module's public name). */
@@ -34,6 +34,9 @@ export interface ExtractTextOptions extends CommonOptions {
34
34
  /** emit a `### pN` marker before each page's content (output formatting
35
35
  * only — `report.pageStats` is always populated regardless) */
36
36
  pageMarkers?: boolean;
37
+ /** source PDF text runs entering layout, delivered once for the selected
38
+ * final result. Runs retain measured bounds and are not fabricated words. */
39
+ onWords?: (words: OcrWordInput[]) => void;
37
40
  }
38
41
  /** Options for {@link extractOcr}. */
39
42
  export interface ExtractOcrOptions extends CommonOptions {
@@ -59,6 +62,10 @@ export interface PdfToMarkdownOptions extends ExtractTextOptions, Omit<ExtractOc
59
62
  /** never fall back to OCR, even for image-only pages */
60
63
  noOcr?: boolean;
61
64
  }
65
+ /** Extract embedded text with a QA report and warnings; never runs OCR.
66
+ * @param pdfPath path to the PDF file
67
+ * @param opts see {@link ExtractTextOptions}
68
+ */
62
69
  export declare function extractText(pdfPath: string, opts?: ExtractTextOptions): Promise<ExtractResult>;
63
70
  /**
64
71
  * OCR extraction. Input is either a PDF (pages render at `dpi` and the
@@ -72,15 +79,19 @@ export declare function extractOcr(input: {
72
79
  } | {
73
80
  words: OcrWordInput[];
74
81
  }, opts?: ExtractOcrOptions): Promise<ExtractResult>;
75
- /** pdfToMarkdown's auto-OCR fallback predicate, shared with the CLI's
76
- * --format=raw|txt path so the two call sites can never drift apart. Module-internal: not re-exported by `src/index.ts`. */
77
- export declare function needsOcrFallback(charsPerPage: number): boolean;
78
- /** The exact stderr/warnings notice printed when needsOcrFallback fires —
79
- * one copy of the string for both call sites (asserted verbatim by the suite). */
80
- export declare function ocrFallbackNotice(charsPerPage: number): string;
82
+ interface AutoAnalysisResult {
83
+ analysis: Analysis;
84
+ markdown: string;
85
+ warnings: string[];
86
+ }
87
+ /** Shared automatic route for the library's Markdown result and the CLI's
88
+ * raw/plain-text formats. A page changes producer only when it has image
89
+ * evidence and an unusable text layer, so a blank or vector-only page never
90
+ * becomes speculative OCR input. */
91
+ export declare function analyzePdfAuto(pdfPath: string, opts?: PdfToMarkdownOptions): Promise<AutoAnalysisResult>;
81
92
  /**
82
- * The headline API: extract a PDF to markdown, text layer first, falling
83
- * back to OCR when the text layer is empty enough the pages are certainly images of text (< `AUTO_OCR_CHARS_PER_PAGE` chars/page) — the same decision the CLI makes.
93
+ * Extract each page from embedded text, using OCR only when the page has
94
+ * image content and fewer than `AUTO_OCR_CHARS_PER_PAGE` text characters.
84
95
  *
85
96
  * @param pdfPath path to the PDF file
86
97
  * @param opts see {@link PdfToMarkdownOptions}
@@ -6,18 +6,11 @@
6
6
  import { buildOcrAnalysisPages } from './analyze.js';
7
7
  import { collect } from './collect.js';
8
8
  import { dictAvailable, dictUnavailableWarning } from './dict.js';
9
- import { toLines } from './lines.js';
10
- import { buildMarkdown } from './markdown.js';
11
9
  import { renderPagesToMarkdown } from './render-analysis.js';
12
- import { AUTO_OCR_CHARS_PER_PAGE, buildOcrReport, buildReport } from './report.js';
10
+ import { AUTO_OCR_CHARS_PER_PAGE, buildOcrReport } from './report.js';
11
+ import { buildTextAnalysis, selectCollectedPages } from './text-analysis.js';
13
12
  import { validatePages } from './validate.js';
14
- /**
15
- * Text-layer extraction. Returns markdown + QA report + warnings (hybrid
16
- * pages carrying large images whose baked-in text the text layer can never see). Makes no OCR decision — see pdfToMarkdown for the fallback.
17
- *
18
- * @param pdfPath path to the PDF file
19
- * @param opts see {@link ExtractTextOptions}
20
- */ // One collector per entry point: collects into `warnings`, streams via
13
+ // One collector per entry point: collects into `warnings`, streams via
21
14
  // onWarning, and emits the dictionary-unavailable notice exactly once up front.
22
15
  function warnCollector(onWarning) {
23
16
  const warnings = [];
@@ -31,47 +24,25 @@ function warnCollector(onWarning) {
31
24
  warn
32
25
  };
33
26
  }
34
- export async function extractText(pdfPath, opts = {}) {
27
+ /** Extract embedded text with a QA report and warnings; never runs OCR.
28
+ * @param pdfPath path to the PDF file
29
+ * @param opts see {@link ExtractTextOptions}
30
+ */ export async function extractText(pdfPath, opts = {}) {
35
31
  var _opts_pages_last, _opts_pageMarkers;
32
+ var _opts_onWords;
36
33
  validatePages(opts.pages);
37
34
  const { warnings, warn } = warnCollector(opts.onWarning);
38
35
  const range = opts.pages ? {
39
36
  first: opts.pages.first,
40
37
  last: (_opts_pages_last = opts.pages.last) !== null && _opts_pages_last !== void 0 ? _opts_pages_last : opts.pages.first
41
38
  } : {};
42
- const { numPages, pages, heightChars } = await collect(pdfPath, range);
43
- const pageLines = pages.map((p)=>({
44
- page: p.page,
45
- lines: toLines(p)
46
- }));
47
- const pageMarkers = (_opts_pageMarkers = opts.pageMarkers) !== null && _opts_pageMarkers !== void 0 ? _opts_pageMarkers : false;
48
- const { md, stats, bodyH, headingSizes } = buildMarkdown(pageLines, heightChars, {
49
- pageMarkers
50
- });
51
- // pageStats needs "### pN" markers to segment content; when the caller
52
- // didn't request them, build a throwaway marked variant (same heightChars) purely to measure stats from, so pageStats is always populated.
53
- const statsMd = pageMarkers ? md : buildMarkdown(pageLines, heightChars, {
54
- pageMarkers: true
55
- }).md;
56
- // Advisory, not prescriptive: `hasLargeImage` fires on any large image,
57
- // even one with no baked-in text, where `--ocr` recovers nothing and can
58
- // actively degrade a correct result (measured: "I" -> "|", stray em-dashes) — so this tells the caller to COMPARE the OCR rerun, never blindly replace.
59
- const hybridPages = pages.filter((p)=>p.largeImage).map((p)=>p.page);
60
- for (const n of hybridPages){
61
- warn(`p${n}: text layer used but page carries large image(s); if it has baked-in text this output won't have it — compare (don't blindly replace) against --pages ${n} --ocr, which can itself introduce errors`);
62
- }
63
- const report = buildReport({
64
- numPages,
65
- md,
66
- statsMd,
67
- stats,
68
- bodyH,
69
- headingSizes,
70
- hybridPages
71
- });
39
+ const collected = await collect(pdfPath, range);
40
+ const text = buildTextAnalysis(collected, (_opts_pageMarkers = opts.pageMarkers) !== null && _opts_pageMarkers !== void 0 ? _opts_pageMarkers : false);
41
+ textNotices(collected, new Set(), warn);
42
+ (_opts_onWords = opts.onWords) === null || _opts_onWords === void 0 ? void 0 : _opts_onWords.call(opts, text.words);
72
43
  return {
73
- markdown: md,
74
- report,
44
+ markdown: text.markdown,
45
+ report: text.analysis.report,
75
46
  warnings
76
47
  };
77
48
  }
@@ -104,44 +75,164 @@ export async function extractText(pdfPath, opts = {}) {
104
75
  warnings
105
76
  };
106
77
  }
107
- /** pdfToMarkdown's auto-OCR fallback predicate, shared with the CLI's
108
- * --format=raw|txt path so the two call sites can never drift apart. Module-internal: not re-exported by `src/index.ts`. */ export function needsOcrFallback(charsPerPage) {
109
- return charsPerPage < AUTO_OCR_CHARS_PER_PAGE;
78
+ /** A page needs OCR only when its image signal accompanies an unusable text
79
+ * layer. The caller supplies the image part of that decision. */ function needsPageOcr(chars) {
80
+ return chars < AUTO_OCR_CHARS_PER_PAGE;
81
+ }
82
+ function textNotices(collected, ocrPages, warn) {
83
+ for (const page of collected.pages){
84
+ if (page.textChars !== 0 || ocrPages.has(page.page)) continue;
85
+ if (page.hasImage) warn(`p${page.page}: no extractable text; page has image content and OCR was not selected`);
86
+ else warn(`p${page.page}: no extractable text; the page may be blank or contain only vector content`);
87
+ }
88
+ for (const page of collected.pages){
89
+ if (!page.largeImage || ocrPages.has(page.page)) continue;
90
+ warn(`p${page.page}: text layer used but page carries large image(s); if it has baked-in text this output won't have it — compare (don't blindly replace) against --pages ${page.page} --ocr, which can itself introduce errors`);
91
+ }
110
92
  }
111
- /** The exact stderr/warnings notice printed when needsOcrFallback fires —
112
- * one copy of the string for both call sites (asserted verbatim by the suite). */ export function ocrFallbackNotice(charsPerPage) {
113
- return `text layer yields ~${charsPerPage} chars/page (< ${AUTO_OCR_CHARS_PER_PAGE}) — pages are images of text; falling back to OCR (suppress with --no-ocr)`;
93
+ function analysisWords(pages) {
94
+ return pages.flatMap((page)=>page.words.map((word)=>({
95
+ page: page.page,
96
+ text: word.text,
97
+ ...word.box,
98
+ confidence: word.confidence
99
+ })));
114
100
  }
115
- /**
116
- * The headline API: extract a PDF to markdown, text layer first, falling
117
- * back to OCR when the text layer is empty enough the pages are certainly images of text (< `AUTO_OCR_CHARS_PER_PAGE` chars/page) — the same decision the CLI makes.
118
- *
119
- * @param pdfPath path to the PDF file
120
- * @param opts see {@link PdfToMarkdownOptions}
121
- */ export async function pdfToMarkdown(pdfPath, opts = {}) {
122
- var _opts_onWarning;
123
- if (opts.ocr) return extractOcr({
124
- pdfPath
125
- }, opts);
126
- // Warning DELIVERY is deferred on the text pass: if the fallback fires,
127
- // those hybrid warnings describe output being discarded, so only a kept text result replays them.
128
- const text = await extractText(pdfPath, {
129
- ...opts,
130
- onWarning: undefined
131
- });
132
- if (needsOcrFallback(text.report.charsPerPage) && !opts.noOcr) {
133
- var _opts_onWarning1;
134
- const notice = ocrFallbackNotice(text.report.charsPerPage);
135
- (_opts_onWarning1 = opts.onWarning) === null || _opts_onWarning1 === void 0 ? void 0 : _opts_onWarning1.call(opts, notice);
136
- const ocr = await extractOcr({
101
+ /** Shared automatic route for the library's Markdown result and the CLI's
102
+ * raw/plain-text formats. A page changes producer only when it has image
103
+ * evidence and an unusable text layer, so a blank or vector-only page never
104
+ * becomes speculative OCR input. */ export async function analyzePdfAuto(pdfPath, opts = {}) {
105
+ var _opts_pages_last;
106
+ var _opts_onWords;
107
+ validatePages(opts.pages);
108
+ const { warnings, warn } = warnCollector(opts.onWarning);
109
+ const range = opts.pages ? {
110
+ first: opts.pages.first,
111
+ last: (_opts_pages_last = opts.pages.last) !== null && _opts_pages_last !== void 0 ? _opts_pages_last : opts.pages.first
112
+ } : {};
113
+ if (opts.ocr) {
114
+ var _opts_onWords1;
115
+ const { pages, words } = await buildOcrAnalysisPages({
137
116
  pdfPath
138
117
  }, {
139
118
  ...opts,
140
- fallback: true
119
+ onWords: undefined
120
+ }, warn);
121
+ const markdown = renderPagesToMarkdown(pages, {
122
+ forcePageMarkers: true,
123
+ pageMarkers: true
141
124
  });
142
- ocr.warnings.unshift(notice);
143
- return ocr;
125
+ const report = buildOcrReport({
126
+ md: markdown,
127
+ words
128
+ });
129
+ const analysis = {
130
+ outputVersion: report.outputVersion,
131
+ report,
132
+ pages
133
+ };
134
+ (_opts_onWords1 = opts.onWords) === null || _opts_onWords1 === void 0 ? void 0 : _opts_onWords1.call(opts, words);
135
+ return {
136
+ analysis,
137
+ markdown,
138
+ warnings
139
+ };
140
+ }
141
+ const collected = await collect(pdfPath, range);
142
+ const ocrPageNumbers = new Set(!opts.noOcr ? collected.pages.filter((page)=>page.hasImage && needsPageOcr(page.textChars)).map((page)=>page.page) : []);
143
+ if (!ocrPageNumbers.size) {
144
+ var _opts_pageMarkers;
145
+ var _opts_onWords2;
146
+ const text = buildTextAnalysis(collected, (_opts_pageMarkers = opts.pageMarkers) !== null && _opts_pageMarkers !== void 0 ? _opts_pageMarkers : false);
147
+ textNotices(collected, ocrPageNumbers, warn);
148
+ (_opts_onWords2 = opts.onWords) === null || _opts_onWords2 === void 0 ? void 0 : _opts_onWords2.call(opts, text.words);
149
+ return {
150
+ analysis: text.analysis,
151
+ markdown: text.markdown,
152
+ warnings
153
+ };
154
+ }
155
+ const textPageNumbers = new Set(collected.pages.filter((page)=>!ocrPageNumbers.has(page.page)).map((page)=>page.page));
156
+ const text = textPageNumbers.size ? buildTextAnalysis(selectCollectedPages(collected, textPageNumbers), true) : undefined;
157
+ const textByPage = new Map(text === null || text === void 0 ? void 0 : text.analysis.pages.map((page)=>[
158
+ page.page,
159
+ page
160
+ ]));
161
+ const ocrByPage = new Map();
162
+ const ocrWords = [];
163
+ const ocrRanges = [];
164
+ for (const page of collected.pages){
165
+ if (!ocrPageNumbers.has(page.page)) continue;
166
+ warn(`p${page.page}: sparse text layer with image content; using OCR for this page`);
167
+ const lastRange = ocrRanges.at(-1);
168
+ if (lastRange && lastRange.last + 1 === page.page) lastRange.last = page.page;
169
+ else ocrRanges.push({
170
+ first: page.page,
171
+ last: page.page
172
+ });
173
+ }
174
+ for (const range of ocrRanges){
175
+ const result = await buildOcrAnalysisPages({
176
+ pdfPath
177
+ }, {
178
+ ...opts,
179
+ pages: range,
180
+ onWords: undefined
181
+ }, warn);
182
+ for (const page of result.pages)ocrByPage.set(page.page, page);
183
+ ocrWords.push(...result.words);
184
+ }
185
+ textNotices(collected, ocrPageNumbers, warn);
186
+ const pages = collected.pages.map((page)=>{
187
+ var _ocrByPage_get;
188
+ return (_ocrByPage_get = ocrByPage.get(page.page)) !== null && _ocrByPage_get !== void 0 ? _ocrByPage_get : textByPage.get(page.page);
189
+ });
190
+ const markdown = renderPagesToMarkdown(pages, {
191
+ forcePageMarkers: true,
192
+ pageMarkers: true
193
+ });
194
+ const pagePaths = new Map(pages.map((page)=>[
195
+ page.page,
196
+ page.path
197
+ ]));
198
+ const report = buildOcrReport({
199
+ md: markdown,
200
+ fallback: true,
201
+ words: ocrWords,
202
+ pagePaths,
203
+ hybridPages: text === null || text === void 0 ? void 0 : text.analysis.report.hybridPages
204
+ });
205
+ if (text) {
206
+ report.bodyFontHeight = text.analysis.report.bodyFontHeight;
207
+ report.headingSizes = text.analysis.report.headingSizes;
208
+ report.listItems = text.analysis.report.listItems;
209
+ report.joinedPairs = text.analysis.report.joinedPairs;
210
+ report.closedHyphens = text.analysis.report.closedHyphens;
211
+ report.hybridPages = text.analysis.report.hybridPages;
144
212
  }
145
- for (const w of text.warnings)(_opts_onWarning = opts.onWarning) === null || _opts_onWarning === void 0 ? void 0 : _opts_onWarning.call(opts, w);
146
- return text;
213
+ const analysis = {
214
+ outputVersion: report.outputVersion,
215
+ report,
216
+ pages
217
+ };
218
+ (_opts_onWords = opts.onWords) === null || _opts_onWords === void 0 ? void 0 : _opts_onWords.call(opts, analysisWords(pages));
219
+ return {
220
+ analysis,
221
+ markdown,
222
+ warnings
223
+ };
224
+ }
225
+ /**
226
+ * Extract each page from embedded text, using OCR only when the page has
227
+ * image content and fewer than `AUTO_OCR_CHARS_PER_PAGE` text characters.
228
+ *
229
+ * @param pdfPath path to the PDF file
230
+ * @param opts see {@link PdfToMarkdownOptions}
231
+ */ export async function pdfToMarkdown(pdfPath, opts = {}) {
232
+ const result = await analyzePdfAuto(pdfPath, opts);
233
+ return {
234
+ markdown: result.markdown,
235
+ report: result.analysis.report,
236
+ warnings: result.warnings
237
+ };
147
238
  }
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/extract.ts"],"sourcesContent":["// The library layer for PDF -> markdown conversion: extractText/extractOcr/\n// pdfToMarkdown each take plain args, return plain data, and throw PdfToMdError — no process.exit, no console.*, no file writing.\n//\n// Warnings are DATA: every notice is returned in `warnings` verbatim and also\n// delivered incrementally via `onWarning` — the exact strings are load-bearing (the test suite asserts them).\nimport { buildOcrAnalysisPages } from './analyze.ts';\nimport { collect } from './collect.ts';\nimport { dictAvailable, dictUnavailableWarning } from './dict.ts';\nimport type { OcrProgressEvent } from './engines/tesseract.ts';\nimport type { PageWord, Tuning } from './geometry.ts';\nimport { toLines } from './lines.ts';\nimport { buildMarkdown } from './markdown.ts';\nimport { renderPagesToMarkdown } from './render-analysis.ts';\nimport { AUTO_OCR_CHARS_PER_PAGE, buildOcrReport, buildReport } from './report.ts';\nimport type { PageLines, Report } from './types.ts';\nimport { validatePages } from './validate.ts';\n\nexport type { OcrProgressEvent } from './engines/tesseract.ts';\n\n// A plain-JS caller (or `.ts` with `as any`) gets no compile-time check on\n// `pages`, so it's validated once at runtime — see src/validate.ts's `validatePages`.\n\n/** One recognized word with its page — the OCR replay input contract (same\n * shape as `geometry.ts`'s `PageWord`, re-exported under this module's public name). */\nexport type OcrWordInput = PageWord;\n\n/** A 1-based, inclusive page range — the shape every entry point's `pages`\n * option takes (the CLI's own `--pages N[-M]` parses into this). */\nexport interface PageSelection {\n first: number;\n /** inclusive; defaults to `first` */\n last?: number;\n}\n\n/** extractText/extractOcr/pdfToMarkdown's common return shape. */\nexport interface ExtractResult {\n markdown: string;\n /** per-document and per-page QA metrics — see `Report` (`src/types.ts`) */\n report: Report;\n /** every notice also delivered incrementally via `onWarning`, collected\n * here in full (hybrid pages, auto-OCR-fallback, empty OCR pages, sparse-retry) */\n warnings: string[];\n}\n\ninterface CommonOptions {\n /** convert only this page range (1-based, inclusive); defaults to the\n * whole document */\n pages?: PageSelection;\n /** delivered as each warning happens (also returned, in full, as\n * `warnings` on the resolved result) */\n onWarning?: (message: string) => void;\n}\n\n/** Options for {@link extractText}. */\nexport interface ExtractTextOptions extends CommonOptions {\n /** emit a `### pN` marker before each page's content (output formatting\n * only — `report.pageStats` is always populated regardless) */\n pageMarkers?: boolean;\n}\n\n/** Options for {@link extractOcr}. */\nexport interface ExtractOcrOptions extends CommonOptions {\n /** override any subset of the OCR-path geometry defaults — see\n * DEFAULT_TUNING (`src/geometry.ts`) for each knob's measured default */\n tuning?: Partial<Tuning>;\n /** page-render resolution for engine recognition (default 288, measured;\n * >~400 measured WORSE on ordinary pages — not a \"bigger is safer\" knob) */\n dpi?: number;\n /** marks report.ocrFallback — set by pdfToMarkdown's auto-fallback */\n fallback?: boolean;\n /** one event per page as recognition finishes (timing/count only — engine\n * chatter and the sparse-retry notice arrive via `onWarning` instead) */\n onProgress?: (e: OcrProgressEvent) => void;\n /** every recognized word entering the geometry, delivered once after\n * recognition — the severity-1 audit instrument (--debug-words) */\n onWords?: (words: OcrWordInput[]) => void;\n}\n\n/** Options for {@link pdfToMarkdown}. */\nexport interface PdfToMarkdownOptions extends ExtractTextOptions, Omit<ExtractOcrOptions, 'fallback'> {\n /** force the OCR path even when a text layer exists */\n ocr?: boolean;\n /** never fall back to OCR, even for image-only pages */\n noOcr?: boolean;\n}\n\n/**\n * Text-layer extraction. Returns markdown + QA report + warnings (hybrid\n * pages carrying large images whose baked-in text the text layer can never see). Makes no OCR decision — see pdfToMarkdown for the fallback.\n *\n * @param pdfPath path to the PDF file\n * @param opts see {@link ExtractTextOptions}\n */\n// One collector per entry point: collects into `warnings`, streams via\n// onWarning, and emits the dictionary-unavailable notice exactly once up front.\nfunction warnCollector(onWarning?: (message: string) => void): { warnings: string[]; warn: (m: string) => void } {\n const warnings: string[] = [];\n const warn = (m: string) => {\n warnings.push(m);\n onWarning?.(m);\n };\n if (!dictAvailable()) warn(dictUnavailableWarning());\n return { warnings, warn };\n}\n\nexport async function extractText(pdfPath: string, opts: ExtractTextOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n\n const range = opts.pages ? { first: opts.pages.first, last: opts.pages.last ?? opts.pages.first } : {};\n const { numPages, pages, heightChars } = await collect(pdfPath, range);\n const pageLines: PageLines[] = pages.map((p) => ({\n page: p.page,\n lines: toLines(p),\n }));\n const pageMarkers = opts.pageMarkers ?? false;\n const { md, stats, bodyH, headingSizes } = buildMarkdown(pageLines, heightChars, { pageMarkers });\n // pageStats needs \"### pN\" markers to segment content; when the caller\n // didn't request them, build a throwaway marked variant (same heightChars) purely to measure stats from, so pageStats is always populated.\n const statsMd = pageMarkers ? md : buildMarkdown(pageLines, heightChars, { pageMarkers: true }).md;\n\n // Advisory, not prescriptive: `hasLargeImage` fires on any large image,\n // even one with no baked-in text, where `--ocr` recovers nothing and can\n // actively degrade a correct result (measured: \"I\" -> \"|\", stray em-dashes) — so this tells the caller to COMPARE the OCR rerun, never blindly replace.\n const hybridPages = pages.filter((p) => p.largeImage).map((p) => p.page);\n for (const n of hybridPages) {\n warn(`p${n}: text layer used but page carries large image(s); if it has baked-in text this output won't have it — compare (don't blindly replace) against --pages ${n} --ocr, which can itself introduce errors`);\n }\n\n const report = buildReport({\n numPages,\n md,\n statsMd,\n stats,\n bodyH,\n headingSizes,\n hybridPages,\n });\n return { markdown: md, report, warnings };\n}\n\n/**\n * OCR extraction. Input is either a PDF (pages render at `dpi` and the\n * engine recognizes them, imported lazily) or an already-recognized word dump (the offline replay input — see `--words-json`).\n *\n * @param input `{ pdfPath }` to recognize, or `{ words }` to replay a dump\n * @param opts see {@link ExtractOcrOptions}\n */\nexport async function extractOcr(input: { pdfPath: string } | { words: OcrWordInput[] }, opts: ExtractOcrOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n\n // Shared with analyze(pdf, {path:'ocr'}) — see that module's header for\n // why this makes the two byte-identical by construction, not by hand.\n const { pages, words } = await buildOcrAnalysisPages(input, opts, warn);\n const md = renderPagesToMarkdown(pages, { forcePageMarkers: true, pageMarkers: true });\n\n const report = buildOcrReport({\n md,\n fallback: opts.fallback,\n // Confidence is on `words` — see report.ts's low-confidence detector for\n // how it handles a confidence-less replay honestly.\n words,\n });\n return { markdown: md, report, warnings };\n}\n\n/** pdfToMarkdown's auto-OCR fallback predicate, shared with the CLI's\n * --format=raw|txt path so the two call sites can never drift apart. Module-internal: not re-exported by `src/index.ts`. */\nexport function needsOcrFallback(charsPerPage: number): boolean {\n return charsPerPage < AUTO_OCR_CHARS_PER_PAGE;\n}\n\n/** The exact stderr/warnings notice printed when needsOcrFallback fires —\n * one copy of the string for both call sites (asserted verbatim by the suite). */\nexport function ocrFallbackNotice(charsPerPage: number): string {\n return `text layer yields ~${charsPerPage} chars/page (< ${AUTO_OCR_CHARS_PER_PAGE}) — pages are images of text; falling back to OCR (suppress with --no-ocr)`;\n}\n\n/**\n * The headline API: extract a PDF to markdown, text layer first, falling\n * back to OCR when the text layer is empty enough the pages are certainly images of text (< `AUTO_OCR_CHARS_PER_PAGE` chars/page) — the same decision the CLI makes.\n *\n * @param pdfPath path to the PDF file\n * @param opts see {@link PdfToMarkdownOptions}\n */\nexport async function pdfToMarkdown(pdfPath: string, opts: PdfToMarkdownOptions = {}): Promise<ExtractResult> {\n if (opts.ocr) return extractOcr({ pdfPath }, opts);\n\n // Warning DELIVERY is deferred on the text pass: if the fallback fires,\n // those hybrid warnings describe output being discarded, so only a kept text result replays them.\n const text = await extractText(pdfPath, { ...opts, onWarning: undefined });\n if (needsOcrFallback(text.report.charsPerPage) && !opts.noOcr) {\n const notice = ocrFallbackNotice(text.report.charsPerPage);\n opts.onWarning?.(notice);\n const ocr = await extractOcr({ pdfPath }, { ...opts, fallback: true });\n ocr.warnings.unshift(notice);\n return ocr;\n }\n for (const w of text.warnings) opts.onWarning?.(w);\n return text;\n}\n"],"names":["buildOcrAnalysisPages","collect","dictAvailable","dictUnavailableWarning","toLines","buildMarkdown","renderPagesToMarkdown","AUTO_OCR_CHARS_PER_PAGE","buildOcrReport","buildReport","validatePages","warnCollector","onWarning","warnings","warn","m","push","extractText","pdfPath","opts","pages","range","first","last","numPages","heightChars","pageLines","map","p","page","lines","pageMarkers","md","stats","bodyH","headingSizes","statsMd","hybridPages","filter","largeImage","n","report","markdown","extractOcr","input","words","forcePageMarkers","fallback","needsOcrFallback","charsPerPage","ocrFallbackNotice","pdfToMarkdown","ocr","text","undefined","noOcr","notice","unshift","w"],"mappings":"AAAA,4EAA4E;AAC5E,kIAAkI;AAClI,EAAE;AACF,8EAA8E;AAC9E,8GAA8G;AAC9G,SAASA,qBAAqB,QAAQ,eAAe;AACrD,SAASC,OAAO,QAAQ,eAAe;AACvC,SAASC,aAAa,EAAEC,sBAAsB,QAAQ,YAAY;AAGlE,SAASC,OAAO,QAAQ,aAAa;AACrC,SAASC,aAAa,QAAQ,gBAAgB;AAC9C,SAASC,qBAAqB,QAAQ,uBAAuB;AAC7D,SAASC,uBAAuB,EAAEC,cAAc,EAAEC,WAAW,QAAQ,cAAc;AAEnF,SAASC,aAAa,QAAQ,gBAAgB;AAuE9C;;;;;;CAMC,GACD,uEAAuE;AACvE,gFAAgF;AAChF,SAASC,cAAcC,SAAqC;IAC1D,MAAMC,WAAqB,EAAE;IAC7B,MAAMC,OAAO,CAACC;QACZF,SAASG,IAAI,CAACD;QACdH,sBAAAA,gCAAAA,UAAYG;IACd;IACA,IAAI,CAACb,iBAAiBY,KAAKX;IAC3B,OAAO;QAAEU;QAAUC;IAAK;AAC1B;AAEA,OAAO,eAAeG,YAAYC,OAAe,EAAEC,OAA2B,CAAC,CAAC;QAIlBA,kBAMxCA;IATpBT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IAEvD,MAAMS,QAAQF,KAAKC,KAAK,GAAG;QAAEE,OAAOH,KAAKC,KAAK,CAACE,KAAK;QAAEC,IAAI,GAAEJ,mBAAAA,KAAKC,KAAK,CAACG,IAAI,cAAfJ,8BAAAA,mBAAmBA,KAAKC,KAAK,CAACE,KAAK;IAAC,IAAI,CAAC;IACrG,MAAM,EAAEE,QAAQ,EAAEJ,KAAK,EAAEK,WAAW,EAAE,GAAG,MAAMxB,QAAQiB,SAASG;IAChE,MAAMK,YAAyBN,MAAMO,GAAG,CAAC,CAACC,IAAO,CAAA;YAC/CC,MAAMD,EAAEC,IAAI;YACZC,OAAO1B,QAAQwB;QACjB,CAAA;IACA,MAAMG,eAAcZ,oBAAAA,KAAKY,WAAW,cAAhBZ,+BAAAA,oBAAoB;IACxC,MAAM,EAAEa,EAAE,EAAEC,KAAK,EAAEC,KAAK,EAAEC,YAAY,EAAE,GAAG9B,cAAcqB,WAAWD,aAAa;QAAEM;IAAY;IAC/F,uEAAuE;IACvE,2IAA2I;IAC3I,MAAMK,UAAUL,cAAcC,KAAK3B,cAAcqB,WAAWD,aAAa;QAAEM,aAAa;IAAK,GAAGC,EAAE;IAElG,wEAAwE;IACxE,yEAAyE;IACzE,wJAAwJ;IACxJ,MAAMK,cAAcjB,MAAMkB,MAAM,CAAC,CAACV,IAAMA,EAAEW,UAAU,EAAEZ,GAAG,CAAC,CAACC,IAAMA,EAAEC,IAAI;IACvE,KAAK,MAAMW,KAAKH,YAAa;QAC3BvB,KAAK,CAAC,CAAC,EAAE0B,EAAE,uJAAuJ,EAAEA,EAAE,yCAAyC,CAAC;IAClN;IAEA,MAAMC,SAAShC,YAAY;QACzBe;QACAQ;QACAI;QACAH;QACAC;QACAC;QACAE;IACF;IACA,OAAO;QAAEK,UAAUV;QAAIS;QAAQ5B;IAAS;AAC1C;AAEA;;;;;;CAMC,GACD,OAAO,eAAe8B,WAAWC,KAAsD,EAAEzB,OAA0B,CAAC,CAAC;IACnHT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IAEvD,wEAAwE;IACxE,sEAAsE;IACtE,MAAM,EAAEQ,KAAK,EAAEyB,KAAK,EAAE,GAAG,MAAM7C,sBAAsB4C,OAAOzB,MAAML;IAClE,MAAMkB,KAAK1B,sBAAsBc,OAAO;QAAE0B,kBAAkB;QAAMf,aAAa;IAAK;IAEpF,MAAMU,SAASjC,eAAe;QAC5BwB;QACAe,UAAU5B,KAAK4B,QAAQ;QACvB,yEAAyE;QACzE,oDAAoD;QACpDF;IACF;IACA,OAAO;QAAEH,UAAUV;QAAIS;QAAQ5B;IAAS;AAC1C;AAEA;0HAC0H,GAC1H,OAAO,SAASmC,iBAAiBC,YAAoB;IACnD,OAAOA,eAAe1C;AACxB;AAEA;gFACgF,GAChF,OAAO,SAAS2C,kBAAkBD,YAAoB;IACpD,OAAO,CAAC,mBAAmB,EAAEA,aAAa,eAAe,EAAE1C,wBAAwB,0EAA0E,CAAC;AAChK;AAEA;;;;;;CAMC,GACD,OAAO,eAAe4C,cAAcjC,OAAe,EAAEC,OAA6B,CAAC,CAAC;QAanDA;IAZ/B,IAAIA,KAAKiC,GAAG,EAAE,OAAOT,WAAW;QAAEzB;IAAQ,GAAGC;IAE7C,wEAAwE;IACxE,kGAAkG;IAClG,MAAMkC,OAAO,MAAMpC,YAAYC,SAAS;QAAE,GAAGC,IAAI;QAAEP,WAAW0C;IAAU;IACxE,IAAIN,iBAAiBK,KAAKZ,MAAM,CAACQ,YAAY,KAAK,CAAC9B,KAAKoC,KAAK,EAAE;YAE7DpC;QADA,MAAMqC,SAASN,kBAAkBG,KAAKZ,MAAM,CAACQ,YAAY;SACzD9B,mBAAAA,KAAKP,SAAS,cAAdO,uCAAAA,sBAAAA,MAAiBqC;QACjB,MAAMJ,MAAM,MAAMT,WAAW;YAAEzB;QAAQ,GAAG;YAAE,GAAGC,IAAI;YAAE4B,UAAU;QAAK;QACpEK,IAAIvC,QAAQ,CAAC4C,OAAO,CAACD;QACrB,OAAOJ;IACT;IACA,KAAK,MAAMM,KAAKL,KAAKxC,QAAQ,EAAEM,kBAAAA,KAAKP,SAAS,cAAdO,sCAAAA,qBAAAA,MAAiBuC;IAChD,OAAOL;AACT"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/open-source/formats/pdf-to-md/src/extract.ts"],"sourcesContent":["// The library layer for PDF -> markdown conversion: extractText/extractOcr/\n// pdfToMarkdown each take plain args, return plain data, and throw PdfToMdError — no process.exit, no console.*, no file writing.\n//\n// Warnings are DATA: every notice is returned in `warnings` verbatim and also\n// delivered incrementally via `onWarning` — the exact strings are load-bearing (the test suite asserts them).\nimport { buildOcrAnalysisPages } from './analyze.ts';\nimport { collect } from './collect.ts';\nimport { dictAvailable, dictUnavailableWarning } from './dict.ts';\nimport type { OcrProgressEvent } from './engines/tesseract.ts';\nimport type { PageWord, Tuning } from './geometry.ts';\nimport { renderPagesToMarkdown } from './render-analysis.ts';\nimport { AUTO_OCR_CHARS_PER_PAGE, buildOcrReport } from './report.ts';\nimport { buildTextAnalysis, selectCollectedPages } from './text-analysis.ts';\nimport type { Analysis, AnalysisPage, CollectResult, Report } from './types.ts';\nimport { validatePages } from './validate.ts';\n\nexport type { OcrProgressEvent } from './engines/tesseract.ts';\n\n// A plain-JS caller (or `.ts` with `as any`) gets no compile-time check on\n// `pages`, so it's validated once at runtime — see src/validate.ts's `validatePages`.\n\n/** One recognized word with its page — the OCR replay input contract (same\n * shape as `geometry.ts`'s `PageWord`, re-exported under this module's public name). */\nexport type OcrWordInput = PageWord;\n\n/** A 1-based, inclusive page range — the shape every entry point's `pages`\n * option takes (the CLI's own `--pages N[-M]` parses into this). */\nexport interface PageSelection {\n first: number;\n /** inclusive; defaults to `first` */\n last?: number;\n}\n\n/** extractText/extractOcr/pdfToMarkdown's common return shape. */\nexport interface ExtractResult {\n markdown: string;\n /** per-document and per-page QA metrics — see `Report` (`src/types.ts`) */\n report: Report;\n /** every notice also delivered incrementally via `onWarning`, collected\n * here in full (hybrid pages, auto-OCR-fallback, empty OCR pages, sparse-retry) */\n warnings: string[];\n}\n\ninterface CommonOptions {\n /** convert only this page range (1-based, inclusive); defaults to the\n * whole document */\n pages?: PageSelection;\n /** delivered as each warning happens (also returned, in full, as\n * `warnings` on the resolved result) */\n onWarning?: (message: string) => void;\n}\n\n/** Options for {@link extractText}. */\nexport interface ExtractTextOptions extends CommonOptions {\n /** emit a `### pN` marker before each page's content (output formatting\n * only — `report.pageStats` is always populated regardless) */\n pageMarkers?: boolean;\n /** source PDF text runs entering layout, delivered once for the selected\n * final result. Runs retain measured bounds and are not fabricated words. */\n onWords?: (words: OcrWordInput[]) => void;\n}\n\n/** Options for {@link extractOcr}. */\nexport interface ExtractOcrOptions extends CommonOptions {\n /** override any subset of the OCR-path geometry defaults — see\n * DEFAULT_TUNING (`src/geometry.ts`) for each knob's measured default */\n tuning?: Partial<Tuning>;\n /** page-render resolution for engine recognition (default 288, measured;\n * >~400 measured WORSE on ordinary pages — not a \"bigger is safer\" knob) */\n dpi?: number;\n /** marks report.ocrFallback — set by pdfToMarkdown's auto-fallback */\n fallback?: boolean;\n /** one event per page as recognition finishes (timing/count only — engine\n * chatter and the sparse-retry notice arrive via `onWarning` instead) */\n onProgress?: (e: OcrProgressEvent) => void;\n /** every recognized word entering the geometry, delivered once after\n * recognition — the severity-1 audit instrument (--debug-words) */\n onWords?: (words: OcrWordInput[]) => void;\n}\n\n/** Options for {@link pdfToMarkdown}. */\nexport interface PdfToMarkdownOptions extends ExtractTextOptions, Omit<ExtractOcrOptions, 'fallback'> {\n /** force the OCR path even when a text layer exists */\n ocr?: boolean;\n /** never fall back to OCR, even for image-only pages */\n noOcr?: boolean;\n}\n\n// One collector per entry point: collects into `warnings`, streams via\n// onWarning, and emits the dictionary-unavailable notice exactly once up front.\nfunction warnCollector(onWarning?: (message: string) => void): { warnings: string[]; warn: (m: string) => void } {\n const warnings: string[] = [];\n const warn = (m: string) => {\n warnings.push(m);\n onWarning?.(m);\n };\n if (!dictAvailable()) warn(dictUnavailableWarning());\n return { warnings, warn };\n}\n\n/** Extract embedded text with a QA report and warnings; never runs OCR.\n * @param pdfPath path to the PDF file\n * @param opts see {@link ExtractTextOptions}\n */\nexport async function extractText(pdfPath: string, opts: ExtractTextOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n const range = opts.pages ? { first: opts.pages.first, last: opts.pages.last ?? opts.pages.first } : {};\n const collected = await collect(pdfPath, range);\n const text = buildTextAnalysis(collected, opts.pageMarkers ?? false);\n textNotices(collected, new Set(), warn);\n opts.onWords?.(text.words);\n return { markdown: text.markdown, report: text.analysis.report, warnings };\n}\n\n/**\n * OCR extraction. Input is either a PDF (pages render at `dpi` and the\n * engine recognizes them, imported lazily) or an already-recognized word dump (the offline replay input — see `--words-json`).\n *\n * @param input `{ pdfPath }` to recognize, or `{ words }` to replay a dump\n * @param opts see {@link ExtractOcrOptions}\n */\nexport async function extractOcr(input: { pdfPath: string } | { words: OcrWordInput[] }, opts: ExtractOcrOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n\n // Shared with analyze(pdf, {path:'ocr'}) — see that module's header for\n // why this makes the two byte-identical by construction, not by hand.\n const { pages, words } = await buildOcrAnalysisPages(input, opts, warn);\n const md = renderPagesToMarkdown(pages, { forcePageMarkers: true, pageMarkers: true });\n\n const report = buildOcrReport({\n md,\n fallback: opts.fallback,\n // Confidence is on `words` — see report.ts's low-confidence detector for\n // how it handles a confidence-less replay honestly.\n words,\n });\n return { markdown: md, report, warnings };\n}\n\n/** A page needs OCR only when its image signal accompanies an unusable text\n * layer. The caller supplies the image part of that decision. */\nfunction needsPageOcr(chars: number): boolean {\n return chars < AUTO_OCR_CHARS_PER_PAGE;\n}\n\nfunction textNotices(collected: CollectResult, ocrPages: Set<number>, warn: (message: string) => void): void {\n for (const page of collected.pages) {\n if (page.textChars !== 0 || ocrPages.has(page.page)) continue;\n if (page.hasImage) warn(`p${page.page}: no extractable text; page has image content and OCR was not selected`);\n else warn(`p${page.page}: no extractable text; the page may be blank or contain only vector content`);\n }\n for (const page of collected.pages) {\n if (!page.largeImage || ocrPages.has(page.page)) continue;\n warn(`p${page.page}: text layer used but page carries large image(s); if it has baked-in text this output won't have it — compare (don't blindly replace) against --pages ${page.page} --ocr, which can itself introduce errors`);\n }\n}\n\nfunction analysisWords(pages: AnalysisPage[]): OcrWordInput[] {\n return pages.flatMap((page) => page.words.map((word) => ({ page: page.page, text: word.text, ...word.box, confidence: word.confidence })));\n}\n\ninterface AutoAnalysisResult {\n analysis: Analysis;\n markdown: string;\n warnings: string[];\n}\n\n/** Shared automatic route for the library's Markdown result and the CLI's\n * raw/plain-text formats. A page changes producer only when it has image\n * evidence and an unusable text layer, so a blank or vector-only page never\n * becomes speculative OCR input. */\nexport async function analyzePdfAuto(pdfPath: string, opts: PdfToMarkdownOptions = {}): Promise<AutoAnalysisResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n const range = opts.pages ? { first: opts.pages.first, last: opts.pages.last ?? opts.pages.first } : {};\n\n if (opts.ocr) {\n const { pages, words } = await buildOcrAnalysisPages({ pdfPath }, { ...opts, onWords: undefined }, warn);\n const markdown = renderPagesToMarkdown(pages, { forcePageMarkers: true, pageMarkers: true });\n const report = buildOcrReport({ md: markdown, words });\n const analysis: Analysis = { outputVersion: report.outputVersion, report, pages };\n opts.onWords?.(words);\n return { analysis, markdown, warnings };\n }\n\n const collected = await collect(pdfPath, range);\n const ocrPageNumbers = new Set(!opts.noOcr ? collected.pages.filter((page) => page.hasImage && needsPageOcr(page.textChars)).map((page) => page.page) : []);\n if (!ocrPageNumbers.size) {\n const text = buildTextAnalysis(collected, opts.pageMarkers ?? false);\n textNotices(collected, ocrPageNumbers, warn);\n opts.onWords?.(text.words);\n return { analysis: text.analysis, markdown: text.markdown, warnings };\n }\n\n const textPageNumbers = new Set(collected.pages.filter((page) => !ocrPageNumbers.has(page.page)).map((page) => page.page));\n const text = textPageNumbers.size ? buildTextAnalysis(selectCollectedPages(collected, textPageNumbers), true) : undefined;\n const textByPage = new Map(text?.analysis.pages.map((page) => [page.page, page]));\n const ocrByPage = new Map<number, AnalysisPage>();\n const ocrWords: OcrWordInput[] = [];\n const ocrRanges: Array<{ first: number; last: number }> = [];\n for (const page of collected.pages) {\n if (!ocrPageNumbers.has(page.page)) continue;\n warn(`p${page.page}: sparse text layer with image content; using OCR for this page`);\n const lastRange = ocrRanges.at(-1);\n if (lastRange && lastRange.last + 1 === page.page) lastRange.last = page.page;\n else ocrRanges.push({ first: page.page, last: page.page });\n }\n for (const range of ocrRanges) {\n const result = await buildOcrAnalysisPages({ pdfPath }, { ...opts, pages: range, onWords: undefined }, warn);\n for (const page of result.pages) ocrByPage.set(page.page, page);\n ocrWords.push(...result.words);\n }\n textNotices(collected, ocrPageNumbers, warn);\n const pages = collected.pages.map((page) => ocrByPage.get(page.page) ?? (textByPage.get(page.page) as AnalysisPage));\n const markdown = renderPagesToMarkdown(pages, { forcePageMarkers: true, pageMarkers: true });\n const pagePaths = new Map(pages.map((page) => [page.page, page.path]));\n const report = buildOcrReport({\n md: markdown,\n fallback: true,\n words: ocrWords,\n pagePaths,\n hybridPages: text?.analysis.report.hybridPages,\n });\n if (text) {\n report.bodyFontHeight = text.analysis.report.bodyFontHeight;\n report.headingSizes = text.analysis.report.headingSizes;\n report.listItems = text.analysis.report.listItems;\n report.joinedPairs = text.analysis.report.joinedPairs;\n report.closedHyphens = text.analysis.report.closedHyphens;\n report.hybridPages = text.analysis.report.hybridPages;\n }\n const analysis: Analysis = { outputVersion: report.outputVersion, report, pages };\n opts.onWords?.(analysisWords(pages));\n return { analysis, markdown, warnings };\n}\n\n/**\n * Extract each page from embedded text, using OCR only when the page has\n * image content and fewer than `AUTO_OCR_CHARS_PER_PAGE` text characters.\n *\n * @param pdfPath path to the PDF file\n * @param opts see {@link PdfToMarkdownOptions}\n */\nexport async function pdfToMarkdown(pdfPath: string, opts: PdfToMarkdownOptions = {}): Promise<ExtractResult> {\n const result = await analyzePdfAuto(pdfPath, opts);\n return { markdown: result.markdown, report: result.analysis.report, warnings: result.warnings };\n}\n"],"names":["buildOcrAnalysisPages","collect","dictAvailable","dictUnavailableWarning","renderPagesToMarkdown","AUTO_OCR_CHARS_PER_PAGE","buildOcrReport","buildTextAnalysis","selectCollectedPages","validatePages","warnCollector","onWarning","warnings","warn","m","push","extractText","pdfPath","opts","pages","range","first","last","collected","text","pageMarkers","textNotices","Set","onWords","words","markdown","report","analysis","extractOcr","input","md","forcePageMarkers","fallback","needsPageOcr","chars","ocrPages","page","textChars","has","hasImage","largeImage","analysisWords","flatMap","map","word","box","confidence","analyzePdfAuto","ocr","undefined","outputVersion","ocrPageNumbers","noOcr","filter","size","textPageNumbers","textByPage","Map","ocrByPage","ocrWords","ocrRanges","lastRange","at","result","set","get","pagePaths","path","hybridPages","bodyFontHeight","headingSizes","listItems","joinedPairs","closedHyphens","pdfToMarkdown"],"mappings":"AAAA,4EAA4E;AAC5E,kIAAkI;AAClI,EAAE;AACF,8EAA8E;AAC9E,8GAA8G;AAC9G,SAASA,qBAAqB,QAAQ,eAAe;AACrD,SAASC,OAAO,QAAQ,eAAe;AACvC,SAASC,aAAa,EAAEC,sBAAsB,QAAQ,YAAY;AAGlE,SAASC,qBAAqB,QAAQ,uBAAuB;AAC7D,SAASC,uBAAuB,EAAEC,cAAc,QAAQ,cAAc;AACtE,SAASC,iBAAiB,EAAEC,oBAAoB,QAAQ,qBAAqB;AAE7E,SAASC,aAAa,QAAQ,gBAAgB;AA0E9C,uEAAuE;AACvE,gFAAgF;AAChF,SAASC,cAAcC,SAAqC;IAC1D,MAAMC,WAAqB,EAAE;IAC7B,MAAMC,OAAO,CAACC;QACZF,SAASG,IAAI,CAACD;QACdH,sBAAAA,gCAAAA,UAAYG;IACd;IACA,IAAI,CAACZ,iBAAiBW,KAAKV;IAC3B,OAAO;QAAES;QAAUC;IAAK;AAC1B;AAEA;;;CAGC,GACD,OAAO,eAAeG,YAAYC,OAAe,EAAEC,OAA2B,CAAC,CAAC;QAGlBA,kBAElBA;QAE1CA;IANAT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IACvD,MAAMS,QAAQF,KAAKC,KAAK,GAAG;QAAEE,OAAOH,KAAKC,KAAK,CAACE,KAAK;QAAEC,IAAI,GAAEJ,mBAAAA,KAAKC,KAAK,CAACG,IAAI,cAAfJ,8BAAAA,mBAAmBA,KAAKC,KAAK,CAACE,KAAK;IAAC,IAAI,CAAC;IACrG,MAAME,YAAY,MAAMtB,QAAQgB,SAASG;IACzC,MAAMI,OAAOjB,kBAAkBgB,YAAWL,oBAAAA,KAAKO,WAAW,cAAhBP,+BAAAA,oBAAoB;IAC9DQ,YAAYH,WAAW,IAAII,OAAOd;KAClCK,gBAAAA,KAAKU,OAAO,cAAZV,oCAAAA,mBAAAA,MAAeM,KAAKK,KAAK;IACzB,OAAO;QAAEC,UAAUN,KAAKM,QAAQ;QAAEC,QAAQP,KAAKQ,QAAQ,CAACD,MAAM;QAAEnB;IAAS;AAC3E;AAEA;;;;;;CAMC,GACD,OAAO,eAAeqB,WAAWC,KAAsD,EAAEhB,OAA0B,CAAC,CAAC;IACnHT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IAEvD,wEAAwE;IACxE,sEAAsE;IACtE,MAAM,EAAEQ,KAAK,EAAEU,KAAK,EAAE,GAAG,MAAM7B,sBAAsBkC,OAAOhB,MAAML;IAClE,MAAMsB,KAAK/B,sBAAsBe,OAAO;QAAEiB,kBAAkB;QAAMX,aAAa;IAAK;IAEpF,MAAMM,SAASzB,eAAe;QAC5B6B;QACAE,UAAUnB,KAAKmB,QAAQ;QACvB,yEAAyE;QACzE,oDAAoD;QACpDR;IACF;IACA,OAAO;QAAEC,UAAUK;QAAIJ;QAAQnB;IAAS;AAC1C;AAEA;+DAC+D,GAC/D,SAAS0B,aAAaC,KAAa;IACjC,OAAOA,QAAQlC;AACjB;AAEA,SAASqB,YAAYH,SAAwB,EAAEiB,QAAqB,EAAE3B,IAA+B;IACnG,KAAK,MAAM4B,QAAQlB,UAAUJ,KAAK,CAAE;QAClC,IAAIsB,KAAKC,SAAS,KAAK,KAAKF,SAASG,GAAG,CAACF,KAAKA,IAAI,GAAG;QACrD,IAAIA,KAAKG,QAAQ,EAAE/B,KAAK,CAAC,CAAC,EAAE4B,KAAKA,IAAI,CAAC,sEAAsE,CAAC;aACxG5B,KAAK,CAAC,CAAC,EAAE4B,KAAKA,IAAI,CAAC,2EAA2E,CAAC;IACtG;IACA,KAAK,MAAMA,QAAQlB,UAAUJ,KAAK,CAAE;QAClC,IAAI,CAACsB,KAAKI,UAAU,IAAIL,SAASG,GAAG,CAACF,KAAKA,IAAI,GAAG;QACjD5B,KAAK,CAAC,CAAC,EAAE4B,KAAKA,IAAI,CAAC,uJAAuJ,EAAEA,KAAKA,IAAI,CAAC,yCAAyC,CAAC;IAClO;AACF;AAEA,SAASK,cAAc3B,KAAqB;IAC1C,OAAOA,MAAM4B,OAAO,CAAC,CAACN,OAASA,KAAKZ,KAAK,CAACmB,GAAG,CAAC,CAACC,OAAU,CAAA;gBAAER,MAAMA,KAAKA,IAAI;gBAAEjB,MAAMyB,KAAKzB,IAAI;gBAAE,GAAGyB,KAAKC,GAAG;gBAAEC,YAAYF,KAAKE,UAAU;YAAC,CAAA;AACxI;AAQA;;;kCAGkC,GAClC,OAAO,eAAeC,eAAenC,OAAe,EAAEC,OAA6B,CAAC,CAAC;QAGvBA;QA0D5DA;IA5DAT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IACvD,MAAMS,QAAQF,KAAKC,KAAK,GAAG;QAAEE,OAAOH,KAAKC,KAAK,CAACE,KAAK;QAAEC,IAAI,GAAEJ,mBAAAA,KAAKC,KAAK,CAACG,IAAI,cAAfJ,8BAAAA,mBAAmBA,KAAKC,KAAK,CAACE,KAAK;IAAC,IAAI,CAAC;IAErG,IAAIH,KAAKmC,GAAG,EAAE;YAKZnC;QAJA,MAAM,EAAEC,KAAK,EAAEU,KAAK,EAAE,GAAG,MAAM7B,sBAAsB;YAAEiB;QAAQ,GAAG;YAAE,GAAGC,IAAI;YAAEU,SAAS0B;QAAU,GAAGzC;QACnG,MAAMiB,WAAW1B,sBAAsBe,OAAO;YAAEiB,kBAAkB;YAAMX,aAAa;QAAK;QAC1F,MAAMM,SAASzB,eAAe;YAAE6B,IAAIL;YAAUD;QAAM;QACpD,MAAMG,WAAqB;YAAEuB,eAAexB,OAAOwB,aAAa;YAAExB;YAAQZ;QAAM;SAChFD,iBAAAA,KAAKU,OAAO,cAAZV,qCAAAA,oBAAAA,MAAeW;QACf,OAAO;YAAEG;YAAUF;YAAUlB;QAAS;IACxC;IAEA,MAAMW,YAAY,MAAMtB,QAAQgB,SAASG;IACzC,MAAMoC,iBAAiB,IAAI7B,IAAI,CAACT,KAAKuC,KAAK,GAAGlC,UAAUJ,KAAK,CAACuC,MAAM,CAAC,CAACjB,OAASA,KAAKG,QAAQ,IAAIN,aAAaG,KAAKC,SAAS,GAAGM,GAAG,CAAC,CAACP,OAASA,KAAKA,IAAI,IAAI,EAAE;IAC1J,IAAI,CAACe,eAAeG,IAAI,EAAE;YACkBzC;YAE1CA;QAFA,MAAMM,OAAOjB,kBAAkBgB,YAAWL,oBAAAA,KAAKO,WAAW,cAAhBP,+BAAAA,oBAAoB;QAC9DQ,YAAYH,WAAWiC,gBAAgB3C;SACvCK,iBAAAA,KAAKU,OAAO,cAAZV,qCAAAA,oBAAAA,MAAeM,KAAKK,KAAK;QACzB,OAAO;YAAEG,UAAUR,KAAKQ,QAAQ;YAAEF,UAAUN,KAAKM,QAAQ;YAAElB;QAAS;IACtE;IAEA,MAAMgD,kBAAkB,IAAIjC,IAAIJ,UAAUJ,KAAK,CAACuC,MAAM,CAAC,CAACjB,OAAS,CAACe,eAAeb,GAAG,CAACF,KAAKA,IAAI,GAAGO,GAAG,CAAC,CAACP,OAASA,KAAKA,IAAI;IACxH,MAAMjB,OAAOoC,gBAAgBD,IAAI,GAAGpD,kBAAkBC,qBAAqBe,WAAWqC,kBAAkB,QAAQN;IAChH,MAAMO,aAAa,IAAIC,IAAItC,iBAAAA,2BAAAA,KAAMQ,QAAQ,CAACb,KAAK,CAAC6B,GAAG,CAAC,CAACP,OAAS;YAACA,KAAKA,IAAI;YAAEA;SAAK;IAC/E,MAAMsB,YAAY,IAAID;IACtB,MAAME,WAA2B,EAAE;IACnC,MAAMC,YAAoD,EAAE;IAC5D,KAAK,MAAMxB,QAAQlB,UAAUJ,KAAK,CAAE;QAClC,IAAI,CAACqC,eAAeb,GAAG,CAACF,KAAKA,IAAI,GAAG;QACpC5B,KAAK,CAAC,CAAC,EAAE4B,KAAKA,IAAI,CAAC,+DAA+D,CAAC;QACnF,MAAMyB,YAAYD,UAAUE,EAAE,CAAC,CAAC;QAChC,IAAID,aAAaA,UAAU5C,IAAI,GAAG,MAAMmB,KAAKA,IAAI,EAAEyB,UAAU5C,IAAI,GAAGmB,KAAKA,IAAI;aACxEwB,UAAUlD,IAAI,CAAC;YAAEM,OAAOoB,KAAKA,IAAI;YAAEnB,MAAMmB,KAAKA,IAAI;QAAC;IAC1D;IACA,KAAK,MAAMrB,SAAS6C,UAAW;QAC7B,MAAMG,SAAS,MAAMpE,sBAAsB;YAAEiB;QAAQ,GAAG;YAAE,GAAGC,IAAI;YAAEC,OAAOC;YAAOQ,SAAS0B;QAAU,GAAGzC;QACvG,KAAK,MAAM4B,QAAQ2B,OAAOjD,KAAK,CAAE4C,UAAUM,GAAG,CAAC5B,KAAKA,IAAI,EAAEA;QAC1DuB,SAASjD,IAAI,IAAIqD,OAAOvC,KAAK;IAC/B;IACAH,YAAYH,WAAWiC,gBAAgB3C;IACvC,MAAMM,QAAQI,UAAUJ,KAAK,CAAC6B,GAAG,CAAC,CAACP;YAASsB;gBAAAA,iBAAAA,UAAUO,GAAG,CAAC7B,KAAKA,IAAI,eAAvBsB,4BAAAA,iBAA6BF,WAAWS,GAAG,CAAC7B,KAAKA,IAAI;;IACjG,MAAMX,WAAW1B,sBAAsBe,OAAO;QAAEiB,kBAAkB;QAAMX,aAAa;IAAK;IAC1F,MAAM8C,YAAY,IAAIT,IAAI3C,MAAM6B,GAAG,CAAC,CAACP,OAAS;YAACA,KAAKA,IAAI;YAAEA,KAAK+B,IAAI;SAAC;IACpE,MAAMzC,SAASzB,eAAe;QAC5B6B,IAAIL;QACJO,UAAU;QACVR,OAAOmC;QACPO;QACAE,WAAW,EAAEjD,iBAAAA,2BAAAA,KAAMQ,QAAQ,CAACD,MAAM,CAAC0C,WAAW;IAChD;IACA,IAAIjD,MAAM;QACRO,OAAO2C,cAAc,GAAGlD,KAAKQ,QAAQ,CAACD,MAAM,CAAC2C,cAAc;QAC3D3C,OAAO4C,YAAY,GAAGnD,KAAKQ,QAAQ,CAACD,MAAM,CAAC4C,YAAY;QACvD5C,OAAO6C,SAAS,GAAGpD,KAAKQ,QAAQ,CAACD,MAAM,CAAC6C,SAAS;QACjD7C,OAAO8C,WAAW,GAAGrD,KAAKQ,QAAQ,CAACD,MAAM,CAAC8C,WAAW;QACrD9C,OAAO+C,aAAa,GAAGtD,KAAKQ,QAAQ,CAACD,MAAM,CAAC+C,aAAa;QACzD/C,OAAO0C,WAAW,GAAGjD,KAAKQ,QAAQ,CAACD,MAAM,CAAC0C,WAAW;IACvD;IACA,MAAMzC,WAAqB;QAAEuB,eAAexB,OAAOwB,aAAa;QAAExB;QAAQZ;IAAM;KAChFD,gBAAAA,KAAKU,OAAO,cAAZV,oCAAAA,mBAAAA,MAAe4B,cAAc3B;IAC7B,OAAO;QAAEa;QAAUF;QAAUlB;IAAS;AACxC;AAEA;;;;;;CAMC,GACD,OAAO,eAAemE,cAAc9D,OAAe,EAAEC,OAA6B,CAAC,CAAC;IAClF,MAAMkD,SAAS,MAAMhB,eAAenC,SAASC;IAC7C,OAAO;QAAEY,UAAUsC,OAAOtC,QAAQ;QAAEC,QAAQqC,OAAOpC,QAAQ,CAACD,MAAM;QAAEnB,UAAUwD,OAAOxD,QAAQ;IAAC;AAChG"}