@effortlessmotion/pdf-to-md 0.1.0

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 (189) hide show
  1. package/LICENSE +21 -0
  2. package/README.md +157 -0
  3. package/bin/cli.js +9 -0
  4. package/dist/cjs/analyze.d.cts +35 -0
  5. package/dist/cjs/analyze.d.ts +35 -0
  6. package/dist/cjs/analyze.js +620 -0
  7. package/dist/cjs/analyze.js.map +1 -0
  8. package/dist/cjs/audit.d.cts +38 -0
  9. package/dist/cjs/audit.d.ts +38 -0
  10. package/dist/cjs/audit.js +252 -0
  11. package/dist/cjs/audit.js.map +1 -0
  12. package/dist/cjs/cli/audit.d.cts +3 -0
  13. package/dist/cjs/cli/audit.d.ts +3 -0
  14. package/dist/cjs/cli/audit.js +269 -0
  15. package/dist/cjs/cli/audit.js.map +1 -0
  16. package/dist/cjs/cli/extract.d.cts +3 -0
  17. package/dist/cjs/cli/extract.d.ts +3 -0
  18. package/dist/cjs/cli/extract.js +704 -0
  19. package/dist/cjs/cli/extract.js.map +1 -0
  20. package/dist/cjs/cli/index.d.cts +5 -0
  21. package/dist/cjs/cli/index.d.ts +5 -0
  22. package/dist/cjs/cli/index.js +58 -0
  23. package/dist/cjs/cli/index.js.map +1 -0
  24. package/dist/cjs/cli/render.d.cts +3 -0
  25. package/dist/cjs/cli/render.d.ts +3 -0
  26. package/dist/cjs/cli/render.js +361 -0
  27. package/dist/cjs/cli/render.js.map +1 -0
  28. package/dist/cjs/cli/types.d.cts +10 -0
  29. package/dist/cjs/cli/types.d.ts +10 -0
  30. package/dist/cjs/cli/types.js +7 -0
  31. package/dist/cjs/cli/types.js.map +1 -0
  32. package/dist/cjs/cli.d.cts +2 -0
  33. package/dist/cjs/cli.d.ts +2 -0
  34. package/dist/cjs/cli.js +345 -0
  35. package/dist/cjs/cli.js.map +1 -0
  36. package/dist/cjs/collect.d.cts +6 -0
  37. package/dist/cjs/collect.d.ts +6 -0
  38. package/dist/cjs/collect.js +370 -0
  39. package/dist/cjs/collect.js.map +1 -0
  40. package/dist/cjs/dict.d.cts +8 -0
  41. package/dist/cjs/dict.d.ts +8 -0
  42. package/dist/cjs/dict.js +55 -0
  43. package/dist/cjs/dict.js.map +1 -0
  44. package/dist/cjs/emit.d.cts +27 -0
  45. package/dist/cjs/emit.d.ts +27 -0
  46. package/dist/cjs/emit.js +187 -0
  47. package/dist/cjs/emit.js.map +1 -0
  48. package/dist/cjs/engines/tesseract.d.cts +30 -0
  49. package/dist/cjs/engines/tesseract.d.ts +30 -0
  50. package/dist/cjs/engines/tesseract.js +650 -0
  51. package/dist/cjs/engines/tesseract.js.map +1 -0
  52. package/dist/cjs/errors.d.cts +21 -0
  53. package/dist/cjs/errors.d.ts +21 -0
  54. package/dist/cjs/errors.js +124 -0
  55. package/dist/cjs/errors.js.map +1 -0
  56. package/dist/cjs/extract.d.cts +88 -0
  57. package/dist/cjs/extract.d.ts +88 -0
  58. package/dist/cjs/extract.js +436 -0
  59. package/dist/cjs/extract.js.map +1 -0
  60. package/dist/cjs/geometry.d.cts +63 -0
  61. package/dist/cjs/geometry.d.ts +63 -0
  62. package/dist/cjs/geometry.js +521 -0
  63. package/dist/cjs/geometry.js.map +1 -0
  64. package/dist/cjs/index.d.cts +15 -0
  65. package/dist/cjs/index.d.ts +15 -0
  66. package/dist/cjs/index.js +56 -0
  67. package/dist/cjs/index.js.map +1 -0
  68. package/dist/cjs/lines.d.cts +3 -0
  69. package/dist/cjs/lines.d.ts +3 -0
  70. package/dist/cjs/lines.js +181 -0
  71. package/dist/cjs/lines.js.map +1 -0
  72. package/dist/cjs/markdown.d.cts +2 -0
  73. package/dist/cjs/markdown.d.ts +2 -0
  74. package/dist/cjs/markdown.js +222 -0
  75. package/dist/cjs/markdown.js.map +1 -0
  76. package/dist/cjs/package.json +1 -0
  77. package/dist/cjs/page-marker.d.cts +16 -0
  78. package/dist/cjs/page-marker.d.ts +16 -0
  79. package/dist/cjs/page-marker.js +85 -0
  80. package/dist/cjs/page-marker.js.map +1 -0
  81. package/dist/cjs/pdf-open.d.cts +12 -0
  82. package/dist/cjs/pdf-open.d.ts +12 -0
  83. package/dist/cjs/pdf-open.js +257 -0
  84. package/dist/cjs/pdf-open.js.map +1 -0
  85. package/dist/cjs/raster.d.cts +26 -0
  86. package/dist/cjs/raster.d.ts +26 -0
  87. package/dist/cjs/raster.js +262 -0
  88. package/dist/cjs/raster.js.map +1 -0
  89. package/dist/cjs/render-analysis.d.cts +25 -0
  90. package/dist/cjs/render-analysis.d.ts +25 -0
  91. package/dist/cjs/render-analysis.js +319 -0
  92. package/dist/cjs/render-analysis.js.map +1 -0
  93. package/dist/cjs/report.d.cts +12 -0
  94. package/dist/cjs/report.d.ts +12 -0
  95. package/dist/cjs/report.js +392 -0
  96. package/dist/cjs/report.js.map +1 -0
  97. package/dist/cjs/resolve.d.cts +1 -0
  98. package/dist/cjs/resolve.d.ts +1 -0
  99. package/dist/cjs/resolve.js +18 -0
  100. package/dist/cjs/resolve.js.map +1 -0
  101. package/dist/cjs/types.d.cts +171 -0
  102. package/dist/cjs/types.d.ts +171 -0
  103. package/dist/cjs/types.js +9 -0
  104. package/dist/cjs/types.js.map +1 -0
  105. package/dist/cjs/validate.d.cts +14 -0
  106. package/dist/cjs/validate.d.ts +14 -0
  107. package/dist/cjs/validate.js +96 -0
  108. package/dist/cjs/validate.js.map +1 -0
  109. package/dist/esm/analyze.d.ts +35 -0
  110. package/dist/esm/analyze.js +270 -0
  111. package/dist/esm/analyze.js.map +1 -0
  112. package/dist/esm/audit.d.ts +38 -0
  113. package/dist/esm/audit.js +113 -0
  114. package/dist/esm/audit.js.map +1 -0
  115. package/dist/esm/cli/audit.d.ts +3 -0
  116. package/dist/esm/cli/audit.js +20 -0
  117. package/dist/esm/cli/audit.js.map +1 -0
  118. package/dist/esm/cli/extract.d.ts +3 -0
  119. package/dist/esm/cli/extract.js +313 -0
  120. package/dist/esm/cli/extract.js.map +1 -0
  121. package/dist/esm/cli/index.d.ts +5 -0
  122. package/dist/esm/cli/index.js +7 -0
  123. package/dist/esm/cli/index.js.map +1 -0
  124. package/dist/esm/cli/render.d.ts +3 -0
  125. package/dist/esm/cli/render.js +77 -0
  126. package/dist/esm/cli/render.js.map +1 -0
  127. package/dist/esm/cli/types.d.ts +10 -0
  128. package/dist/esm/cli/types.js +3 -0
  129. package/dist/esm/cli/types.js.map +1 -0
  130. package/dist/esm/cli.d.ts +2 -0
  131. package/dist/esm/cli.js +90 -0
  132. package/dist/esm/cli.js.map +1 -0
  133. package/dist/esm/collect.d.ts +6 -0
  134. package/dist/esm/collect.js +80 -0
  135. package/dist/esm/collect.js.map +1 -0
  136. package/dist/esm/dict.d.ts +8 -0
  137. package/dist/esm/dict.js +28 -0
  138. package/dist/esm/dict.js.map +1 -0
  139. package/dist/esm/emit.d.ts +27 -0
  140. package/dist/esm/emit.js +131 -0
  141. package/dist/esm/emit.js.map +1 -0
  142. package/dist/esm/engines/tesseract.d.ts +30 -0
  143. package/dist/esm/engines/tesseract.js +190 -0
  144. package/dist/esm/engines/tesseract.js.map +1 -0
  145. package/dist/esm/errors.d.ts +21 -0
  146. package/dist/esm/errors.js +9 -0
  147. package/dist/esm/errors.js.map +1 -0
  148. package/dist/esm/extract.d.ts +88 -0
  149. package/dist/esm/extract.js +147 -0
  150. package/dist/esm/extract.js.map +1 -0
  151. package/dist/esm/geometry.d.ts +63 -0
  152. package/dist/esm/geometry.js +279 -0
  153. package/dist/esm/geometry.js.map +1 -0
  154. package/dist/esm/index.d.ts +15 -0
  155. package/dist/esm/index.js +10 -0
  156. package/dist/esm/index.js.map +1 -0
  157. package/dist/esm/lines.d.ts +3 -0
  158. package/dist/esm/lines.js +67 -0
  159. package/dist/esm/lines.js.map +1 -0
  160. package/dist/esm/markdown.d.ts +2 -0
  161. package/dist/esm/markdown.js +153 -0
  162. package/dist/esm/markdown.js.map +1 -0
  163. package/dist/esm/package.json +1 -0
  164. package/dist/esm/page-marker.d.ts +16 -0
  165. package/dist/esm/page-marker.js +40 -0
  166. package/dist/esm/page-marker.js.map +1 -0
  167. package/dist/esm/pdf-open.d.ts +12 -0
  168. package/dist/esm/pdf-open.js +61 -0
  169. package/dist/esm/pdf-open.js.map +1 -0
  170. package/dist/esm/raster.d.ts +26 -0
  171. package/dist/esm/raster.js +60 -0
  172. package/dist/esm/raster.js.map +1 -0
  173. package/dist/esm/render-analysis.d.ts +25 -0
  174. package/dist/esm/render-analysis.js +115 -0
  175. package/dist/esm/render-analysis.js.map +1 -0
  176. package/dist/esm/report.d.ts +12 -0
  177. package/dist/esm/report.js +182 -0
  178. package/dist/esm/report.js.map +1 -0
  179. package/dist/esm/resolve.d.ts +1 -0
  180. package/dist/esm/resolve.js +7 -0
  181. package/dist/esm/resolve.js.map +1 -0
  182. package/dist/esm/types.d.ts +171 -0
  183. package/dist/esm/types.js +6 -0
  184. package/dist/esm/types.js.map +1 -0
  185. package/dist/esm/validate.d.ts +14 -0
  186. package/dist/esm/validate.js +56 -0
  187. package/dist/esm/validate.js.map +1 -0
  188. package/package.json +67 -0
  189. package/skills/pdf-to-md/SKILL.md +102 -0
@@ -0,0 +1,171 @@
1
+ export interface Glyph {
2
+ s: string;
3
+ x: number;
4
+ y: number;
5
+ w: number;
6
+ h: number;
7
+ }
8
+ export interface CollectedPage {
9
+ page: number;
10
+ glyphs: Glyph[];
11
+ width: number;
12
+ height: number;
13
+ textChars: number;
14
+ largeImage: boolean;
15
+ }
16
+ export interface CollectResult {
17
+ numPages: number;
18
+ pages: CollectedPage[];
19
+ heightChars: Map<number, number>;
20
+ }
21
+ export interface Line {
22
+ text: string;
23
+ x: number;
24
+ y: number;
25
+ h: number;
26
+ col: number;
27
+ }
28
+ export interface PageLines {
29
+ page: number;
30
+ lines: Line[];
31
+ }
32
+ export interface BuildMarkdownOptions {
33
+ pageMarkers?: boolean;
34
+ }
35
+ export interface MarkdownStats {
36
+ headings: number;
37
+ listItems: number;
38
+ paragraphs: number;
39
+ joinedPairs: number;
40
+ reflowed: number;
41
+ callouts: number;
42
+ closedHyphens: number;
43
+ }
44
+ export interface MarkdownResult {
45
+ md: string;
46
+ stats: MarkdownStats;
47
+ bodyH: number;
48
+ headingSizes: number[];
49
+ }
50
+ /** One row of the per-page QA breakdown ("### pN" blocks) — used to decide
51
+ * which pages are worth rendering and re-checking (README's "Design contract"). */
52
+ export interface PageStatsRow {
53
+ page: number;
54
+ chars: number;
55
+ headings: number;
56
+ junkHeadings: number;
57
+ floats: number;
58
+ paragraphs: number;
59
+ danglingLong: number;
60
+ danglingShort: number;
61
+ /** count of `> [review: ...]` lines — blocks the geometry could not order
62
+ * and labeled instead of guessing (README's "Design contract"). */
63
+ reviewBlocks: number;
64
+ /** present only on the text path's rows; omitted (not `false`) on the OCR
65
+ * path, which has no per-page hybrid-image signal. */
66
+ largeImages?: boolean;
67
+ /** count of words below `report.ts`'s LOW_CONFIDENCE_THRESHOLD; omitted (not
68
+ * 0) when the page had no confidence data at all, so "measured 0" and "nothing to measure" stay distinct. */
69
+ lowConfidenceWords?: number;
70
+ /** the suspect tokens behind `lowConfidenceWords`, worst-confidence first
71
+ * and capped (`report.ts`'s LOW_CONFIDENCE_SAMPLE_CAP) so a chart-heavy page can't bloat the JSON. */
72
+ lowConfidenceSample?: string[];
73
+ }
74
+ export interface BuildReportInput {
75
+ numPages: number;
76
+ md: string;
77
+ /** identical to `md` but always carries "### pN" markers, so pageStats is
78
+ * measured independent of the caller's `--page-markers` choice; same string as `md` when markers were requested. */
79
+ statsMd: string;
80
+ stats: MarkdownStats;
81
+ bodyH: number;
82
+ headingSizes: number[];
83
+ hybridPages?: number[];
84
+ }
85
+ /** One recognized word's page/text/confidence — `report.ts`'s low-confidence
86
+ * detector input; a structural subset of OcrWordInput kept here so `types.ts` doesn't import the library-layer module that depends on it. */
87
+ export interface WordConfidenceInput {
88
+ page: number;
89
+ text: string;
90
+ confidence?: number;
91
+ }
92
+ export interface BuildOcrReportInput {
93
+ md: string;
94
+ fallback?: boolean;
95
+ /** every recognized word, for `report.ts`'s low-confidence detector.
96
+ * Defaults to [] (a pre-confidence --words-json replay), reporting the detector absent, never wrong. */
97
+ words?: WordConfidenceInput[];
98
+ }
99
+ /** Per-document and per-page QA metrics returned by every extraction entry
100
+ * point, discriminated by `path`; OCR-only and text-only fields are grouped below. */
101
+ export interface Report {
102
+ /** manually-bumped integer (not the npm package version) — see
103
+ * `report.ts`'s OUTPUT_VERSION for the bump rule; present on every Report so a frozen baseline's staleness is self-announcing. */
104
+ outputVersion: number;
105
+ path: 'text' | 'ocr';
106
+ pages: number;
107
+ chars: number;
108
+ charsPerPage: number;
109
+ sparse: boolean;
110
+ headings: number;
111
+ junkHeadings: number;
112
+ paragraphs: number;
113
+ callouts: number;
114
+ danglingLong: number;
115
+ danglingShort: number;
116
+ /** total `> [review: ...]` blocks across all pages (sum of pageStats'
117
+ * reviewBlocks); always 0 on the text path — only the OCR path emits review markers. */
118
+ reviewBlocks: number;
119
+ /** document-level sum of pageStats' `lowConfidenceWords` — see that
120
+ * field for why 0 can mean "measured 0" or "nothing to measure". */
121
+ lowConfidenceWords: number;
122
+ pageStats: PageStatsRow[];
123
+ floats?: number;
124
+ ocrFallback?: boolean;
125
+ bodyFontHeight?: number;
126
+ headingSizes?: number[];
127
+ listItems?: number;
128
+ joinedPairs?: number;
129
+ closedHyphens?: number;
130
+ hybridPages?: number[];
131
+ }
132
+ /** One recognized/collected word, page-relative; `box` is always present.
133
+ * See `AnalysisPage.words` for the coordinate contract and per-path granularity difference. */
134
+ export interface AnalysisWord {
135
+ text: string;
136
+ box: {
137
+ x: number;
138
+ y: number;
139
+ w: number;
140
+ h: number;
141
+ };
142
+ confidence?: number;
143
+ }
144
+ /** One structural block; `wordIndexes` links it to its page's `words` list (blocks -> words only).
145
+ * `text` is bare content for `'heading'` (level separate), else the LITERAL markdown line (decoration included, round-trippable) — a pull-quote is `'paragraph'`, not a 5th type. */
146
+ export interface AnalysisBlock {
147
+ type: 'heading' | 'paragraph' | 'float' | 'review';
148
+ text: string;
149
+ /** Heading level (integer 1-6); REQUIRED when `type` is `'heading'`,
150
+ * FORBIDDEN otherwise — toMarkdown()/toText() validate this fail-fast (throws PdfToMdError/'ANALYSIS_INPUT'), never silently ignored. */
151
+ level?: number;
152
+ wordIndexes: number[];
153
+ }
154
+ /** One page's flat words plus its structural blocks. */
155
+ export interface AnalysisPage {
156
+ page: number;
157
+ /** Flat, page-relative word list; `AnalysisBlock.wordIndexes` indexes into it. `box` is always normalized [0,1], origin bottom-left, y up, regardless of path.
158
+ * OCR-path is genuinely per-word; TEXT-path is per visual LINE (`box.w` always 0, `box.y` the line's baseline) — coarser but honest, not faked. */
159
+ words: AnalysisWord[];
160
+ blocks: AnalysisBlock[];
161
+ }
162
+ /** analyze()'s output: plain, JSON-serializable per-page words/blocks plus
163
+ * the same QA `report` the other entry points return. */
164
+ export interface Analysis {
165
+ /** manually-bumped integer — see `Report.outputVersion` / `report.ts`'s OUTPUT_VERSION. */
166
+ outputVersion: number;
167
+ /** the same QA report extractText/extractOcr/pdfToMarkdown return, always
168
+ * measured off real markdown (OCR: the shared core's output; TEXT: buildMarkdown's reflowed output, not this coarser IR) — never stale relative to the IR. */
169
+ report: Report;
170
+ pages: AnalysisPage[];
171
+ }
@@ -0,0 +1,9 @@
1
+ // Shared shapes threaded through the collect -> lines -> markdown -> report
2
+ // pipeline; kept in one place since every downstream stage uses at least one.
3
+ // A positioned glyph run from the PDF's text layer: `s` may be more than one
4
+ // character; `y` is flipped so ascending sorts top-down; `h` is font height rounded to 0.1pt.
5
+ "use strict";
6
+ Object.defineProperty(exports, "__esModule", {
7
+ value: true
8
+ });
9
+ /* CJS INTEROP */ if (exports.__esModule && exports.default) { try { Object.defineProperty(exports.default, '__esModule', { value: true }); for (var key in exports) { exports.default[key] = exports[key]; } } catch (_) {}; module.exports = exports.default; }
@@ -0,0 +1 @@
1
+ {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/types.ts"],"sourcesContent":["// Shared shapes threaded through the collect -> lines -> markdown -> report\n// pipeline; kept in one place since every downstream stage uses at least one.\n\n// A positioned glyph run from the PDF's text layer: `s` may be more than one\n// character; `y` is flipped so ascending sorts top-down; `h` is font height rounded to 0.1pt.\nexport interface Glyph {\n s: string;\n x: number;\n y: number;\n w: number;\n h: number;\n}\n\n// One page's worth of collected glyphs plus the geometry collect.ts needs\n// downstream (lines.ts's two-column detection, report.ts's hybrid-page flag).\nexport interface CollectedPage {\n page: number;\n glyphs: Glyph[];\n width: number;\n height: number;\n textChars: number;\n largeImage: boolean;\n}\n\n// collect.ts's return value: selected pages plus the char-weighted\n// font-height histogram, built ONLY from those pages (matters for a `--pages` re-extraction — see collect.ts).\nexport interface CollectResult {\n numPages: number;\n pages: CollectedPage[];\n heightChars: Map<number, number>;\n}\n\n// A visual line assembled from glyphs (lines.ts's `toLines`): `h` is the\n// char-weighted dominant height (not max); `col` is 0 (left/only) or 1 (right, de-braided page).\nexport interface Line {\n text: string;\n x: number;\n y: number;\n h: number;\n col: number;\n}\n\n// One page's lines, as src/extract.ts assembles them for buildMarkdown.\nexport interface PageLines {\n page: number;\n lines: Line[];\n}\n\nexport interface BuildMarkdownOptions {\n pageMarkers?: boolean;\n}\n\n// Running counters src/markdown.ts accumulates while building the output,\n// surfaced in the QA report (src/report.ts's buildReport).\nexport interface MarkdownStats {\n headings: number;\n listItems: number;\n paragraphs: number;\n joinedPairs: number;\n reflowed: number;\n callouts: number;\n closedHyphens: number;\n}\n\nexport interface MarkdownResult {\n md: string;\n stats: MarkdownStats;\n bodyH: number;\n headingSizes: number[];\n}\n\n/** One row of the per-page QA breakdown (\"### pN\" blocks) — used to decide\n * which pages are worth rendering and re-checking (README's \"Design contract\"). */\nexport interface PageStatsRow {\n page: number;\n chars: number;\n headings: number;\n junkHeadings: number;\n floats: number;\n paragraphs: number;\n danglingLong: number;\n danglingShort: number;\n /** count of `> [review: ...]` lines — blocks the geometry could not order\n * and labeled instead of guessing (README's \"Design contract\"). */\n reviewBlocks: number;\n /** present only on the text path's rows; omitted (not `false`) on the OCR\n * path, which has no per-page hybrid-image signal. */\n largeImages?: boolean;\n /** count of words below `report.ts`'s LOW_CONFIDENCE_THRESHOLD; omitted (not\n * 0) when the page had no confidence data at all, so \"measured 0\" and \"nothing to measure\" stay distinct. */\n lowConfidenceWords?: number;\n /** the suspect tokens behind `lowConfidenceWords`, worst-confidence first\n * and capped (`report.ts`'s LOW_CONFIDENCE_SAMPLE_CAP) so a chart-heavy page can't bloat the JSON. */\n lowConfidenceSample?: string[];\n}\n\n// buildReport's params (src/report.ts) — everything the text (default) path\n// hands over after running collect -> lines -> markdown.\nexport interface BuildReportInput {\n numPages: number;\n md: string;\n /** identical to `md` but always carries \"### pN\" markers, so pageStats is\n * measured independent of the caller's `--page-markers` choice; same string as `md` when markers were requested. */\n statsMd: string;\n stats: MarkdownStats;\n bodyH: number;\n headingSizes: number[];\n hybridPages?: number[];\n}\n\n/** One recognized word's page/text/confidence — `report.ts`'s low-confidence\n * detector input; a structural subset of OcrWordInput kept here so `types.ts` doesn't import the library-layer module that depends on it. */\nexport interface WordConfidenceInput {\n page: number;\n text: string;\n confidence?: number;\n}\n\nexport interface BuildOcrReportInput {\n md: string;\n fallback?: boolean;\n /** every recognized word, for `report.ts`'s low-confidence detector.\n * Defaults to [] (a pre-confidence --words-json replay), reporting the detector absent, never wrong. */\n words?: WordConfidenceInput[];\n}\n\n/** Per-document and per-page QA metrics returned by every extraction entry\n * point, discriminated by `path`; OCR-only and text-only fields are grouped below. */\nexport interface Report {\n /** manually-bumped integer (not the npm package version) — see\n * `report.ts`'s OUTPUT_VERSION for the bump rule; present on every Report so a frozen baseline's staleness is self-announcing. */\n outputVersion: number;\n path: 'text' | 'ocr';\n pages: number;\n chars: number;\n charsPerPage: number;\n sparse: boolean;\n headings: number;\n junkHeadings: number;\n paragraphs: number;\n callouts: number;\n danglingLong: number;\n danglingShort: number;\n /** total `> [review: ...]` blocks across all pages (sum of pageStats'\n * reviewBlocks); always 0 on the text path — only the OCR path emits review markers. */\n reviewBlocks: number;\n /** document-level sum of pageStats' `lowConfidenceWords` — see that\n * field for why 0 can mean \"measured 0\" or \"nothing to measure\". */\n lowConfidenceWords: number;\n pageStats: PageStatsRow[];\n\n // --- OCR-only ---\n floats?: number;\n ocrFallback?: boolean;\n\n // --- text-only ---\n bodyFontHeight?: number;\n headingSizes?: number[];\n listItems?: number;\n joinedPairs?: number;\n closedHyphens?: number;\n hybridPages?: number[];\n}\n\n// --- Analysis: the raw-mode IR (analyze() produces it; toMarkdown()/toText()\n// in render-analysis.ts render it — see toMarkdown's doc comment for the fail-fast validation contract) ---\n\n/** One recognized/collected word, page-relative; `box` is always present.\n * See `AnalysisPage.words` for the coordinate contract and per-path granularity difference. */\nexport interface AnalysisWord {\n text: string;\n box: { x: number; y: number; w: number; h: number };\n confidence?: number;\n}\n\n/** One structural block; `wordIndexes` links it to its page's `words` list (blocks -> words only).\n * `text` is bare content for `'heading'` (level separate), else the LITERAL markdown line (decoration included, round-trippable) — a pull-quote is `'paragraph'`, not a 5th type. */\nexport interface AnalysisBlock {\n type: 'heading' | 'paragraph' | 'float' | 'review';\n text: string;\n /** Heading level (integer 1-6); REQUIRED when `type` is `'heading'`,\n * FORBIDDEN otherwise — toMarkdown()/toText() validate this fail-fast (throws PdfToMdError/'ANALYSIS_INPUT'), never silently ignored. */\n level?: number;\n wordIndexes: number[];\n}\n\n/** One page's flat words plus its structural blocks. */\nexport interface AnalysisPage {\n page: number;\n /** Flat, page-relative word list; `AnalysisBlock.wordIndexes` indexes into it. `box` is always normalized [0,1], origin bottom-left, y up, regardless of path.\n * OCR-path is genuinely per-word; TEXT-path is per visual LINE (`box.w` always 0, `box.y` the line's baseline) — coarser but honest, not faked. */\n words: AnalysisWord[];\n blocks: AnalysisBlock[];\n}\n\n/** analyze()'s output: plain, JSON-serializable per-page words/blocks plus\n * the same QA `report` the other entry points return. */\nexport interface Analysis {\n /** manually-bumped integer — see `Report.outputVersion` / `report.ts`'s OUTPUT_VERSION. */\n outputVersion: number;\n /** the same QA report extractText/extractOcr/pdfToMarkdown return, always\n * measured off real markdown (OCR: the shared core's output; TEXT: buildMarkdown's reflowed output, not this coarser IR) — never stale relative to the IR. */\n report: Report;\n pages: AnalysisPage[];\n}\n"],"names":[],"mappings":"AAAA,4EAA4E;AAC5E,8EAA8E;AAE9E,6EAA6E;AAC7E,8FAA8F"}
@@ -0,0 +1,14 @@
1
+ import type { OcrWordInput } from './extract.js';
2
+ /** Validates the one option every entry point accepts before touching the
3
+ * PDF. Type-only import from `extract.ts` keeps this cycle-free. */
4
+ export declare function validatePages(pages: unknown): void;
5
+ /** Validates analyze()'s non-string `input` branch (`{ words }`) before
6
+ * `.words` is read, so a plain-JS mistake like `analyze(null)` throws PdfToMdError/WORDS_INPUT instead of a raw TypeError. */
7
+ export declare function validateWordsInput(input: unknown): OcrWordInput[];
8
+ /** Runtime shape guard for `buildOcrAnalysisPages`'s input: a plain-JS
9
+ * caller bypasses the TS union, so this turns a raw TypeError into PdfToMdError before any property is read off it. */
10
+ export declare function validateOcrAnalysisInput(input: unknown): {
11
+ pdfPath: string;
12
+ } | {
13
+ words: OcrWordInput[];
14
+ };
@@ -0,0 +1,14 @@
1
+ import type { OcrWordInput } from './extract.js';
2
+ /** Validates the one option every entry point accepts before touching the
3
+ * PDF. Type-only import from `extract.ts` keeps this cycle-free. */
4
+ export declare function validatePages(pages: unknown): void;
5
+ /** Validates analyze()'s non-string `input` branch (`{ words }`) before
6
+ * `.words` is read, so a plain-JS mistake like `analyze(null)` throws PdfToMdError/WORDS_INPUT instead of a raw TypeError. */
7
+ export declare function validateWordsInput(input: unknown): OcrWordInput[];
8
+ /** Runtime shape guard for `buildOcrAnalysisPages`'s input: a plain-JS
9
+ * caller bypasses the TS union, so this turns a raw TypeError into PdfToMdError before any property is read off it. */
10
+ export declare function validateOcrAnalysisInput(input: unknown): {
11
+ pdfPath: string;
12
+ } | {
13
+ words: OcrWordInput[];
14
+ };
@@ -0,0 +1,96 @@
1
+ // validate.ts — input-shape validation shared by analyze.ts and extract.ts,
2
+ // kept out of both to dodge their value-level import cycle (extract.ts imports analyze.ts's buildOcrAnalysisPages; both need this).
3
+ "use strict";
4
+ Object.defineProperty(exports, "__esModule", {
5
+ value: true
6
+ });
7
+ function _export(target, all) {
8
+ for(var name in all)Object.defineProperty(target, name, {
9
+ enumerable: true,
10
+ get: Object.getOwnPropertyDescriptor(all, name).get
11
+ });
12
+ }
13
+ _export(exports, {
14
+ get validateOcrAnalysisInput () {
15
+ return validateOcrAnalysisInput;
16
+ },
17
+ get validatePages () {
18
+ return validatePages;
19
+ },
20
+ get validateWordsInput () {
21
+ return validateWordsInput;
22
+ }
23
+ });
24
+ var _errorsts = require("./errors.js");
25
+ function _type_of(obj) {
26
+ "@swc/helpers - typeof";
27
+ return obj && typeof Symbol !== "undefined" && obj.constructor === Symbol ? "symbol" : typeof obj;
28
+ }
29
+ function validatePages(pages) {
30
+ if (pages === undefined) return;
31
+ var isRecord = (typeof pages === "undefined" ? "undefined" : _type_of(pages)) === 'object' && pages !== null;
32
+ var first = isRecord ? pages.first : undefined;
33
+ var last = isRecord ? pages.last : undefined;
34
+ var firstOk = typeof first === 'number' && Number.isFinite(first) && first >= 1;
35
+ var lastOk = last === undefined || typeof last === 'number' && Number.isFinite(last) && last >= first;
36
+ if (!isRecord || !firstOk || !lastOk) {
37
+ throw new _errorsts.PdfToMdError('PAGE_RANGE', "pages option must be { first: number, last?: number } with first >= 1 and last >= first (got ".concat(JSON.stringify(pages), ")"));
38
+ }
39
+ }
40
+ function validateWordsInput(input) {
41
+ var isRecord = (typeof input === "undefined" ? "undefined" : _type_of(input)) === 'object' && input !== null;
42
+ var words = isRecord ? input.words : undefined;
43
+ if (!Array.isArray(words)) {
44
+ throw new _errorsts.PdfToMdError('WORDS_INPUT', "analyze() input must be a pdf path (string) or { words: OcrWordInput[] } (got ".concat(isRecord ? JSON.stringify(input) : String(input), ")"));
45
+ }
46
+ return words;
47
+ }
48
+ function validateOcrAnalysisInput(input) {
49
+ if (typeof input === 'string') {
50
+ throw new _errorsts.PdfToMdError('PDF_OPEN', "extractOcr/analyze input must be { pdfPath: string } or { words: OcrWordInput[] }, not a bare path string (got ".concat(JSON.stringify(input), ") — did you mean { pdfPath: ").concat(JSON.stringify(input), " }?"));
51
+ }
52
+ var isRecord = (typeof input === "undefined" ? "undefined" : _type_of(input)) === 'object' && input !== null;
53
+ var rawPdfPath = isRecord ? input.pdfPath : undefined;
54
+ var rawWords = isRecord ? input.words : undefined;
55
+ if (!isRecord || rawPdfPath === undefined && rawWords === undefined) {
56
+ throw new _errorsts.PdfToMdError('WORDS_INPUT', "extractOcr/analyze input must be { pdfPath: string } or { words: OcrWordInput[] } (got ".concat(isRecord ? JSON.stringify(input) : String(input), ")"));
57
+ }
58
+ if (rawWords !== undefined && !Array.isArray(rawWords)) {
59
+ throw new _errorsts.PdfToMdError('WORDS_INPUT', "extractOcr/analyze { words } input must be an array (got ".concat(JSON.stringify(rawWords), ")"));
60
+ }
61
+ if (rawPdfPath !== undefined && typeof rawPdfPath !== 'string') {
62
+ throw new _errorsts.PdfToMdError('PDF_OPEN', "extractOcr/analyze { pdfPath } input must be a string (got ".concat(JSON.stringify(rawPdfPath), ")"));
63
+ }
64
+ if (rawWords !== undefined) {
65
+ var _iteratorNormalCompletion = true, _didIteratorError = false, _iteratorError = undefined;
66
+ try {
67
+ for(var _iterator = rawWords[Symbol.iterator](), _step; !(_iteratorNormalCompletion = (_step = _iterator.next()).done); _iteratorNormalCompletion = true){
68
+ var w = _step.value;
69
+ var rec = w;
70
+ if (typeof rec.page !== 'number' || typeof rec.text !== 'string' || typeof rec.x !== 'number' || typeof rec.y !== 'number' || typeof rec.w !== 'number' || typeof rec.h !== 'number') {
71
+ throw new _errorsts.PdfToMdError('WORDS_INPUT', "malformed word entry (need {page, text, x, y, w, h}): ".concat(JSON.stringify(w)));
72
+ }
73
+ }
74
+ } catch (err) {
75
+ _didIteratorError = true;
76
+ _iteratorError = err;
77
+ } finally{
78
+ try {
79
+ if (!_iteratorNormalCompletion && _iterator.return != null) {
80
+ _iterator.return();
81
+ }
82
+ } finally{
83
+ if (_didIteratorError) {
84
+ throw _iteratorError;
85
+ }
86
+ }
87
+ }
88
+ return {
89
+ words: rawWords
90
+ };
91
+ }
92
+ return {
93
+ pdfPath: rawPdfPath
94
+ };
95
+ }
96
+ /* CJS INTEROP */ if (exports.__esModule && exports.default) { try { Object.defineProperty(exports.default, '__esModule', { value: true }); for (var key in exports) { exports.default[key] = exports[key]; } } catch (_) {}; module.exports = exports.default; }
@@ -0,0 +1 @@
1
+ {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/validate.ts"],"sourcesContent":["// validate.ts — input-shape validation shared by analyze.ts and extract.ts,\n// kept out of both to dodge their value-level import cycle (extract.ts imports analyze.ts's buildOcrAnalysisPages; both need this).\nimport { PdfToMdError } from './errors.ts';\nimport type { OcrWordInput } from './extract.ts';\n\n/** Validates the one option every entry point accepts before touching the\n * PDF. Type-only import from `extract.ts` keeps this cycle-free. */\nexport function validatePages(pages: unknown): void {\n if (pages === undefined) return;\n const isRecord = typeof pages === 'object' && pages !== null;\n const first = isRecord ? (pages as Record<string, unknown>).first : undefined;\n const last = isRecord ? (pages as Record<string, unknown>).last : undefined;\n const firstOk = typeof first === 'number' && Number.isFinite(first) && first >= 1;\n const lastOk = last === undefined || (typeof last === 'number' && Number.isFinite(last) && last >= (first as number));\n if (!isRecord || !firstOk || !lastOk) {\n throw new PdfToMdError('PAGE_RANGE', `pages option must be { first: number, last?: number } with first >= 1 and last >= first (got ${JSON.stringify(pages)})`);\n }\n}\n\n/** Validates analyze()'s non-string `input` branch (`{ words }`) before\n * `.words` is read, so a plain-JS mistake like `analyze(null)` throws PdfToMdError/WORDS_INPUT instead of a raw TypeError. */\nexport function validateWordsInput(input: unknown): OcrWordInput[] {\n const isRecord = typeof input === 'object' && input !== null;\n const words = isRecord ? (input as Record<string, unknown>).words : undefined;\n if (!Array.isArray(words)) {\n throw new PdfToMdError('WORDS_INPUT', `analyze() input must be a pdf path (string) or { words: OcrWordInput[] } (got ${isRecord ? JSON.stringify(input) : String(input)})`);\n }\n return words as OcrWordInput[];\n}\n\n/** Runtime shape guard for `buildOcrAnalysisPages`'s input: a plain-JS\n * caller bypasses the TS union, so this turns a raw TypeError into PdfToMdError before any property is read off it. */\nexport function validateOcrAnalysisInput(input: unknown): { pdfPath: string } | { words: OcrWordInput[] } {\n if (typeof input === 'string') {\n throw new PdfToMdError('PDF_OPEN', `extractOcr/analyze input must be { pdfPath: string } or { words: OcrWordInput[] }, not a bare path string (got ${JSON.stringify(input)}) — did you mean { pdfPath: ${JSON.stringify(input)} }?`);\n }\n const isRecord = typeof input === 'object' && input !== null;\n const rawPdfPath = isRecord ? (input as Record<string, unknown>).pdfPath : undefined;\n const rawWords = isRecord ? (input as Record<string, unknown>).words : undefined;\n if (!isRecord || (rawPdfPath === undefined && rawWords === undefined)) {\n throw new PdfToMdError('WORDS_INPUT', `extractOcr/analyze input must be { pdfPath: string } or { words: OcrWordInput[] } (got ${isRecord ? JSON.stringify(input) : String(input)})`);\n }\n if (rawWords !== undefined && !Array.isArray(rawWords)) {\n throw new PdfToMdError('WORDS_INPUT', `extractOcr/analyze { words } input must be an array (got ${JSON.stringify(rawWords)})`);\n }\n if (rawPdfPath !== undefined && typeof rawPdfPath !== 'string') {\n throw new PdfToMdError('PDF_OPEN', `extractOcr/analyze { pdfPath } input must be a string (got ${JSON.stringify(rawPdfPath)})`);\n }\n if (rawWords !== undefined) {\n for (const w of rawWords as unknown[]) {\n const rec = w as Record<string, unknown>;\n if (typeof rec.page !== 'number' || typeof rec.text !== 'string' || typeof rec.x !== 'number' || typeof rec.y !== 'number' || typeof rec.w !== 'number' || typeof rec.h !== 'number') {\n throw new PdfToMdError('WORDS_INPUT', `malformed word entry (need {page, text, x, y, w, h}): ${JSON.stringify(w)}`);\n }\n }\n return { words: rawWords as OcrWordInput[] };\n }\n return { pdfPath: rawPdfPath as string };\n}\n"],"names":["validateOcrAnalysisInput","validatePages","validateWordsInput","pages","undefined","isRecord","first","last","firstOk","Number","isFinite","lastOk","PdfToMdError","JSON","stringify","input","words","Array","isArray","String","rawPdfPath","pdfPath","rawWords","w","rec","page","text","x","y","h"],"mappings":"AAAA,4EAA4E;AAC5E,oIAAoI;;;;;;;;;;;;QA+BpHA;eAAAA;;QAzBAC;eAAAA;;QAcAC;eAAAA;;;wBAnBa;;;;;AAKtB,SAASD,cAAcE,KAAc;IAC1C,IAAIA,UAAUC,WAAW;IACzB,IAAMC,WAAW,CAAA,OAAOF,sCAAP,SAAOA,MAAI,MAAM,YAAYA,UAAU;IACxD,IAAMG,QAAQD,WAAW,AAACF,MAAkCG,KAAK,GAAGF;IACpE,IAAMG,OAAOF,WAAW,AAACF,MAAkCI,IAAI,GAAGH;IAClE,IAAMI,UAAU,OAAOF,UAAU,YAAYG,OAAOC,QAAQ,CAACJ,UAAUA,SAAS;IAChF,IAAMK,SAASJ,SAASH,aAAc,OAAOG,SAAS,YAAYE,OAAOC,QAAQ,CAACH,SAASA,QAASD;IACpG,IAAI,CAACD,YAAY,CAACG,WAAW,CAACG,QAAQ;QACpC,MAAM,IAAIC,sBAAY,CAAC,cAAc,AAAC,gGAAqH,OAAtBC,KAAKC,SAAS,CAACX,QAAO;IAC7J;AACF;AAIO,SAASD,mBAAmBa,KAAc;IAC/C,IAAMV,WAAW,CAAA,OAAOU,sCAAP,SAAOA,MAAI,MAAM,YAAYA,UAAU;IACxD,IAAMC,QAAQX,WAAW,AAACU,MAAkCC,KAAK,GAAGZ;IACpE,IAAI,CAACa,MAAMC,OAAO,CAACF,QAAQ;QACzB,MAAM,IAAIJ,sBAAY,CAAC,eAAe,AAAC,iFAAiI,OAAjDP,WAAWQ,KAAKC,SAAS,CAACC,SAASI,OAAOJ,QAAO;IAC1K;IACA,OAAOC;AACT;AAIO,SAAShB,yBAAyBe,KAAc;IACrD,IAAI,OAAOA,UAAU,UAAU;QAC7B,MAAM,IAAIH,sBAAY,CAAC,YAAY,AAAC,kHAAqKC,OAApDA,KAAKC,SAAS,CAACC,QAAO,gCAAoD,OAAtBF,KAAKC,SAAS,CAACC,QAAO;IACjO;IACA,IAAMV,WAAW,CAAA,OAAOU,sCAAP,SAAOA,MAAI,MAAM,YAAYA,UAAU;IACxD,IAAMK,aAAaf,WAAW,AAACU,MAAkCM,OAAO,GAAGjB;IAC3E,IAAMkB,WAAWjB,WAAW,AAACU,MAAkCC,KAAK,GAAGZ;IACvE,IAAI,CAACC,YAAae,eAAehB,aAAakB,aAAalB,WAAY;QACrE,MAAM,IAAIQ,sBAAY,CAAC,eAAe,AAAC,0FAA0I,OAAjDP,WAAWQ,KAAKC,SAAS,CAACC,SAASI,OAAOJ,QAAO;IACnL;IACA,IAAIO,aAAalB,aAAa,CAACa,MAAMC,OAAO,CAACI,WAAW;QACtD,MAAM,IAAIV,sBAAY,CAAC,eAAe,AAAC,4DAAoF,OAAzBC,KAAKC,SAAS,CAACQ,WAAU;IAC7H;IACA,IAAIF,eAAehB,aAAa,OAAOgB,eAAe,UAAU;QAC9D,MAAM,IAAIR,sBAAY,CAAC,YAAY,AAAC,8DAAwF,OAA3BC,KAAKC,SAAS,CAACM,aAAY;IAC9H;IACA,IAAIE,aAAalB,WAAW;YACrB,kCAAA,2BAAA;;YAAL,QAAK,YAAWkB,6BAAX,SAAA,6BAAA,QAAA,yBAAA,iCAAkC;gBAAlC,IAAMC,IAAN;gBACH,IAAMC,MAAMD;gBACZ,IAAI,OAAOC,IAAIC,IAAI,KAAK,YAAY,OAAOD,IAAIE,IAAI,KAAK,YAAY,OAAOF,IAAIG,CAAC,KAAK,YAAY,OAAOH,IAAII,CAAC,KAAK,YAAY,OAAOJ,IAAID,CAAC,KAAK,YAAY,OAAOC,IAAIK,CAAC,KAAK,UAAU;oBACpL,MAAM,IAAIjB,sBAAY,CAAC,eAAe,AAAC,yDAA0E,OAAlBC,KAAKC,SAAS,CAACS;gBAChH;YACF;;YALK;YAAA;;;qBAAA,6BAAA;oBAAA;;;oBAAA;0BAAA;;;;QAML,OAAO;YAAEP,OAAOM;QAA2B;IAC7C;IACA,OAAO;QAAED,SAASD;IAAqB;AACzC"}
@@ -0,0 +1,35 @@
1
+ import type { ExtractOcrOptions, ExtractTextOptions, OcrWordInput } from './extract.js';
2
+ import type { Analysis, AnalysisPage } from './types.js';
3
+ /** analyze(pdf) with the default/'text' path — OCR-only fields don't exist
4
+ * on this branch, so passing e.g. `tuning` without `path: 'ocr'` is a compile error. */
5
+ export type AnalyzeTextOptions = ExtractTextOptions & {
6
+ path?: 'text';
7
+ };
8
+ /** analyze(pdf, {path:'ocr'}) — same fields as {@link ExtractOcrOptions},
9
+ * unlocked only once `path: 'ocr'` is set. */
10
+ export type AnalyzeOcrOptions = ExtractOcrOptions & {
11
+ path: 'ocr';
12
+ };
13
+ /** Options for {@link analyze}: which fields type-check depends on `path`
14
+ * (ignored, forced 'ocr', when `input` is a `{ words }` replay). */
15
+ export type AnalyzeOptions = AnalyzeTextOptions | AnalyzeOcrOptions;
16
+ /** The OCR-path core shared by extractOcr and analyze(..., {path:'ocr'}):
17
+ * recognizes (or replays) words and runs the unchanged buildPageOCR/emitPage pipeline. Word IDENTITY (object reference), not value, carries indices through into the emitted blocks — see `geometry.ts`'s OcrLine.words. */
18
+ export declare function buildOcrAnalysisPages(input: {
19
+ pdfPath: string;
20
+ } | {
21
+ words: OcrWordInput[];
22
+ }, opts: ExtractOcrOptions, warn: (m: string) => void): Promise<{
23
+ pages: AnalysisPage[];
24
+ words: OcrWordInput[];
25
+ }>;
26
+ /**
27
+ * Analyze a PDF (or replay an already-recognized OCR word dump) into
28
+ * `Analysis`: plain, JSON-serializable per-page `words`/`blocks` plus the same QA `report` the other entry points return. Render with toMarkdown()/toText(), or hand-edit/splice it (no merge API — see toMarkdown's doc comment).
29
+ *
30
+ * @param input a bare pdf path, or `{ words }` to replay a dump (always analyzed as OCR) — unlike extractOcr, never `{ pdfPath }`
31
+ * @param opts see {@link AnalyzeOptions}; for a pdf path, `opts.path` picks 'text' (default) or 'ocr' — analyze() makes no auto-fallback decision (that's pdfToMarkdown's job)
32
+ */
33
+ export declare function analyze(input: string | {
34
+ words: OcrWordInput[];
35
+ }, opts?: AnalyzeOptions): Promise<Analysis>;
@@ -0,0 +1,270 @@
1
+ // analyze() plus the OCR-path core (buildOcrAnalysisPages) that extractOcr
2
+ // shares with it — same buildPageOCR -> emitPage -> renderPagesToMarkdown pipeline, so OCR-path markdown is byte-identical by construction, not two implementations kept in sync by hand. The TEXT path is NOT shared this way (see analyzeText).
3
+ import { collect, pageCount } from './collect.js';
4
+ import { dictAvailable, dictUnavailableWarning } from './dict.js';
5
+ import { emitPage, renderEmitBlock } from './emit.js';
6
+ import { buildPageOCR, DEFAULT_TUNING, headingLevels } from './geometry.js';
7
+ import { toLines } from './lines.js';
8
+ import { buildMarkdown } from './markdown.js';
9
+ import { renderPagesToMarkdown } from './render-analysis.js';
10
+ import { buildOcrReport, buildReport, OUTPUT_VERSION } from './report.js';
11
+ import { validateOcrAnalysisInput, validatePages, validateWordsInput } from './validate.js';
12
+ function toAnalysisWord(w) {
13
+ const box = {
14
+ x: w.x,
15
+ y: w.y,
16
+ w: w.w,
17
+ h: w.h
18
+ };
19
+ return w.confidence === undefined ? {
20
+ text: w.text,
21
+ box
22
+ } : {
23
+ text: w.text,
24
+ box,
25
+ confidence: w.confidence
26
+ };
27
+ }
28
+ // EmitBlock -> AnalysisBlock: collapses emit.ts's 'quote' kind (a pull-quote)
29
+ // into Analysis's 'paragraph' type, `text` keeping its literal '> ' prefix — see AnalysisBlock's doc comment (src/types.ts) for why.
30
+ function toAnalysisBlock(b, wordIndex) {
31
+ const wordIndexes = b.words.map((w)=>wordIndex.get(w));
32
+ if (b.kind === 'heading') return {
33
+ type: 'heading',
34
+ text: b.text,
35
+ level: b.level,
36
+ wordIndexes
37
+ };
38
+ if (b.kind === 'review') return {
39
+ type: 'review',
40
+ text: renderEmitBlock(b),
41
+ wordIndexes
42
+ };
43
+ return {
44
+ type: 'paragraph',
45
+ text: renderEmitBlock(b),
46
+ wordIndexes
47
+ };
48
+ }
49
+ /** The OCR-path core shared by extractOcr and analyze(..., {path:'ocr'}):
50
+ * recognizes (or replays) words and runs the unchanged buildPageOCR/emitPage pipeline. Word IDENTITY (object reference), not value, carries indices through into the emitted blocks — see `geometry.ts`'s OcrLine.words. */ export async function buildOcrAnalysisPages(input, opts, warn) {
51
+ var _byPage_get;
52
+ var _opts_onWords;
53
+ const tuning = {
54
+ ...DEFAULT_TUNING,
55
+ ...opts.tuning
56
+ };
57
+ const validated = validateOcrAnalysisInput(input);
58
+ let words;
59
+ let first;
60
+ let last;
61
+ if ('words' in validated) {
62
+ var _ref, _ref1, _ref2;
63
+ var _opts_pages, _opts_pages1, _opts_pages2;
64
+ // Replay: page range from --pages if given, else every page up to the highest one the dump mentions.
65
+ words = validated.words;
66
+ first = (_ref = (_opts_pages = opts.pages) === null || _opts_pages === void 0 ? void 0 : _opts_pages.first) !== null && _ref !== void 0 ? _ref : 1;
67
+ last = (_ref1 = (_ref2 = (_opts_pages1 = opts.pages) === null || _opts_pages1 === void 0 ? void 0 : _opts_pages1.last) !== null && _ref2 !== void 0 ? _ref2 : (_opts_pages2 = opts.pages) === null || _opts_pages2 === void 0 ? void 0 : _opts_pages2.first) !== null && _ref1 !== void 0 ? _ref1 : Math.max(1, ...words.map((w)=>w.page));
68
+ } else {
69
+ var _ref3, _ref4;
70
+ var _opts_pages3, _opts_pages4;
71
+ const { recognizePdf } = await import('./engines/tesseract.js');
72
+ const totalPages = await pageCount(validated.pdfPath);
73
+ const rawFirst = (_ref3 = (_opts_pages3 = opts.pages) === null || _opts_pages3 === void 0 ? void 0 : _opts_pages3.first) !== null && _ref3 !== void 0 ? _ref3 : 1;
74
+ const rawLast = (_ref4 = (_opts_pages4 = opts.pages) === null || _opts_pages4 === void 0 ? void 0 : _opts_pages4.last) !== null && _ref4 !== void 0 ? _ref4 : opts.pages ? rawFirst : totalPages;
75
+ first = Math.min(Math.max(1, rawFirst), Math.max(1, totalPages));
76
+ last = Math.min(Math.max(first, rawLast), totalPages);
77
+ const result = await recognizePdf(validated.pdfPath, first, last, {
78
+ dpi: opts.dpi,
79
+ onProgress: opts.onProgress,
80
+ onWarning: warn
81
+ });
82
+ words = result.words;
83
+ }
84
+ (_opts_onWords = opts.onWords) === null || _opts_onWords === void 0 ? void 0 : _opts_onWords.call(opts, words);
85
+ // Stable per-page index, assigned in recognition/replay order — this IS
86
+ // the page's flat `words` list order; `wordIndex` recovers it later since geometry.ts never clones a Word (object identity survives row-building).
87
+ const byPage = new Map();
88
+ const wordIndex = new Map();
89
+ for (const w of words){
90
+ const list = byPage.get(w.page);
91
+ if (list) {
92
+ wordIndex.set(w, list.length);
93
+ list.push(w);
94
+ } else {
95
+ wordIndex.set(w, 0);
96
+ byPage.set(w.page, [
97
+ w
98
+ ]);
99
+ }
100
+ }
101
+ const pageOcrs = new Map();
102
+ for(let pn = first; pn <= last; pn++)pageOcrs.set(pn, buildPageOCR((_byPage_get = byPage.get(pn)) !== null && _byPage_get !== void 0 ? _byPage_get : [], tuning));
103
+ const { levelOf } = headingLevels([
104
+ ...pageOcrs.values()
105
+ ], tuning);
106
+ const pages = [];
107
+ for(let pn = first; pn <= last; pn++){
108
+ var _byPage_get1;
109
+ const p = pageOcrs.get(pn);
110
+ // An in-range page with no recognized words stays visible (marker +
111
+ // warning), never silently absent — same rule as extractOcr.
112
+ if (!byPage.has(pn)) warn(`p${pn}: no recognized words, page emitted with no OCR text`);
113
+ const blocks = emitPage(p, levelOf, tuning).map((b)=>toAnalysisBlock(b, wordIndex));
114
+ // Decorative fragments always render as one visible "> [floats]" aside
115
+ // per page (matching the one printed line) — see geometry.ts's PageOCR.floats.
116
+ if (p.floats.length) {
117
+ blocks.push({
118
+ type: 'float',
119
+ text: `> [floats] ${p.floats.map((f)=>f.text).join(' · ')}`,
120
+ wordIndexes: p.floats.flatMap((f)=>f.words).map((w)=>wordIndex.get(w))
121
+ });
122
+ }
123
+ pages.push({
124
+ page: pn,
125
+ words: ((_byPage_get1 = byPage.get(pn)) !== null && _byPage_get1 !== void 0 ? _byPage_get1 : []).map(toAnalysisWord),
126
+ blocks
127
+ });
128
+ }
129
+ return {
130
+ pages,
131
+ words
132
+ };
133
+ }
134
+ async function analyzeOcr(input, opts) {
135
+ const warn = (m)=>{
136
+ var _opts_onWarning;
137
+ return (_opts_onWarning = opts.onWarning) === null || _opts_onWarning === void 0 ? void 0 : _opts_onWarning.call(opts, m);
138
+ };
139
+ if (!dictAvailable()) warn(dictUnavailableWarning());
140
+ const { pages, words } = await buildOcrAnalysisPages(input, opts, warn);
141
+ const md = renderPagesToMarkdown(pages, {
142
+ forcePageMarkers: true,
143
+ pageMarkers: true
144
+ });
145
+ const report = buildOcrReport({
146
+ md,
147
+ fallback: opts.fallback,
148
+ words
149
+ });
150
+ return {
151
+ outputVersion: OUTPUT_VERSION,
152
+ report,
153
+ pages
154
+ };
155
+ }
156
+ // TEXT path: deliberately NOT routed through the OCR shared-core pattern —
157
+ // buildMarkdown's reflow/hyphen-rejoin/list rules don't fit the 4-type Analysis block union, so this is an independent, intentionally coarser reconstruction (one word/block per visual LINE) — honest, not byte-identical to extractText (see AnalysisPage, src/types.ts).
158
+ // Converts a text-path Line's page-point geometry (src/lines.ts) into the
159
+ // same normalized box contract the OCR path uses (see AnalysisPage.words) — only the coordinate SYSTEM changes; `box.y` anchors on the line's baseline, `box.w` stays 0.
160
+ function textLineBox(l, dims) {
161
+ const { width, height } = dims;
162
+ return {
163
+ x: width > 0 ? l.x / width : 0,
164
+ y: height > 0 ? (height - l.y) / height : 0,
165
+ w: 0,
166
+ h: height > 0 ? l.h / height : 0
167
+ };
168
+ }
169
+ async function analyzeText(pdfPath, opts) {
170
+ var _opts_pages_last, _opts_pageMarkers;
171
+ const warn = (m)=>{
172
+ var _opts_onWarning;
173
+ return (_opts_onWarning = opts.onWarning) === null || _opts_onWarning === void 0 ? void 0 : _opts_onWarning.call(opts, m);
174
+ };
175
+ if (!dictAvailable()) warn(dictUnavailableWarning());
176
+ const range = opts.pages ? {
177
+ first: opts.pages.first,
178
+ last: (_opts_pages_last = opts.pages.last) !== null && _opts_pages_last !== void 0 ? _opts_pages_last : opts.pages.first
179
+ } : {};
180
+ const { numPages, pages, heightChars } = await collect(pdfPath, range);
181
+ const pageDims = new Map(pages.map((p)=>[
182
+ p.page,
183
+ {
184
+ width: p.width,
185
+ height: p.height
186
+ }
187
+ ]));
188
+ const pageLines = pages.map((p)=>({
189
+ page: p.page,
190
+ lines: toLines(p)
191
+ }));
192
+ const pageMarkers = (_opts_pageMarkers = opts.pageMarkers) !== null && _opts_pageMarkers !== void 0 ? _opts_pageMarkers : false;
193
+ const { md, stats, bodyH, headingSizes } = buildMarkdown(pageLines, heightChars, {
194
+ pageMarkers
195
+ });
196
+ const statsMd = pageMarkers ? md : buildMarkdown(pageLines, heightChars, {
197
+ pageMarkers: true
198
+ }).md;
199
+ const hybridPages = pages.filter((p)=>p.largeImage).map((p)=>p.page);
200
+ for (const n of hybridPages){
201
+ 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`);
202
+ }
203
+ // Report is measured from the REAL markdown above — accurate regardless of
204
+ // the coarser block/word reconstruction below.
205
+ const report = buildReport({
206
+ numPages,
207
+ md,
208
+ statsMd,
209
+ stats,
210
+ bodyH,
211
+ headingSizes,
212
+ hybridPages
213
+ });
214
+ // headingSizes is offset-independent — reused directly, then re-leveled
215
+ // here with offset 0 regardless of `opts.pageMarkers` (see AnalysisPage).
216
+ const levelOf = new Map(headingSizes.slice(0, 6).map((h, i)=>[
217
+ h,
218
+ Math.min(6, i + 1)
219
+ ]));
220
+ const analysisPages = pageLines.map(({ page, lines })=>{
221
+ const dims = pageDims.get(page);
222
+ const words = lines.map((l)=>({
223
+ text: l.text,
224
+ box: textLineBox(l, dims)
225
+ }));
226
+ const blocks = lines.map((l, i)=>{
227
+ const level = levelOf.get(l.h);
228
+ return level !== undefined ? {
229
+ type: 'heading',
230
+ text: l.text,
231
+ level,
232
+ wordIndexes: [
233
+ i
234
+ ]
235
+ } : {
236
+ type: 'paragraph',
237
+ text: l.text,
238
+ wordIndexes: [
239
+ i
240
+ ]
241
+ };
242
+ });
243
+ return {
244
+ page,
245
+ words,
246
+ blocks
247
+ };
248
+ });
249
+ return {
250
+ outputVersion: OUTPUT_VERSION,
251
+ report,
252
+ pages: analysisPages
253
+ };
254
+ }
255
+ /**
256
+ * Analyze a PDF (or replay an already-recognized OCR word dump) into
257
+ * `Analysis`: plain, JSON-serializable per-page `words`/`blocks` plus the same QA `report` the other entry points return. Render with toMarkdown()/toText(), or hand-edit/splice it (no merge API — see toMarkdown's doc comment).
258
+ *
259
+ * @param input a bare pdf path, or `{ words }` to replay a dump (always analyzed as OCR) — unlike extractOcr, never `{ pdfPath }`
260
+ * @param opts see {@link AnalyzeOptions}; for a pdf path, `opts.path` picks 'text' (default) or 'ocr' — analyze() makes no auto-fallback decision (that's pdfToMarkdown's job)
261
+ */ export async function analyze(input, opts = {}) {
262
+ validatePages(opts.pages);
263
+ if (typeof input !== 'string') return analyzeOcr({
264
+ words: validateWordsInput(input)
265
+ }, opts);
266
+ if (opts.path === 'ocr') return analyzeOcr({
267
+ pdfPath: input
268
+ }, opts);
269
+ return analyzeText(input, opts);
270
+ }