@precisa-saude/fhir-ocr-utils 0.21.0 → 0.21.1

package/dist/cli.js

@@ -18,7 +18,7 @@ var CONFIDENCE_NAME_ONLY = 0.7;
 var CONFIDENCE_AMBIGUOUS = 0.4;
 var MAX_OCCURRENCES_PER_NAME = 5;
 function normalize(text) {
-  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/[^\S\n]+/g, " ");
+  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/(?<=\p{L})-(?=[\p{L}\p{N}])/gu, " ").replace(/[^\S\n]+/g, " ");
 }
 var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "hdl",
@@ -449,7 +449,7 @@ async function main() {
     strict: false
   });
   if (values.version) {
-    process.stdout.write(`${"0.21.0"}
+    process.stdout.write(`${"0.21.1"}
 `);
     return;
   }

package/dist/index.cjs

@@ -9,7 +9,7 @@ var CONFIDENCE_NAME_ONLY = 0.7;
 var CONFIDENCE_AMBIGUOUS = 0.4;
 var MAX_OCCURRENCES_PER_NAME = 5;
 function normalize(text) {
-  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/[^\S\n]+/g, " ");
+  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/(?<=\p{L})-(?=[\p{L}\p{N}])/gu, " ").replace(/[^\S\n]+/g, " ");
 }
 var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "hdl",

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["/home/runner/work/fhir-brasil/fhir-brasil/packages/ocr-utils/dist/index.cjs","../src/anchor.ts"],"names":[],"mappings":"AAAA;ACaA;AAEE;AACA;AACA;AAAA,2CACK;AAwBA,IAAM,0BAAA,EAA4B,CAAA;AAMlC,IAAM,qBAAA,EAAuB,GAAA;AAM7B,IAAM,qBAAA,EAAuB,GAAA;AAGpC,IAAM,yBAAA,EAA2B,CAAA;AASjC,SAAS,SAAA,CAAU,IAAA,EAAsB;AACvC,EAAA,OAAO,IAAA,CACJ,SAAA,CAAU,KAAK,CAAA,CACf,OAAA,CAAQ,kBAAA,EAAoB,EAAE,CAAA,CAC9B,WAAA,CAAY,CAAA,CACZ,OAAA,CAAQ,WAAA,EAAa,GAAG,CAAA;AAC7B;AAEA,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACt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uBAAA,kBAAyB,IAAI,GAAA,CAAI;AAAA,EACrC,UAAA;AAAA;AAAA,EACA,WAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,YAAA;AAAA;AAAA,EACA,KAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,UAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,YAAA;AAAA,EACA,OAAA;AAAA,EACA,SAAA;AAAA,EACA;AACF,CAAC,CAAA;AAOD,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACtC,QAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,OAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,cAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,OAAA;AAAA,EACA,UAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,QAAA;AAAA,EACA,OAAA;AAAA,EACA,cAAA;AAAA,EACA;AACF,CAAC,CAAA;AASD,IAAM,yBAAA,EAAqC;AAAA,EACzC,qBAAA;AAAA;AAAA,EACA,kBAAA;AAAA;AAAA,EACA,iBAAA;AAAA;AAAA,EACA,uBAAA;AAAA;AAAA,EACA,cAAA;AAAA;AAAA,EACA,YAAA;AAAA,EACA,iBAAA;AAAA,EACA,YAAA;AAAA,EACA,cAAA;AAAA,EACA,gBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,gBAAA;AAAA,EACA,mBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,aAAA;AAAA,EACA,WAAA;AAAA,EACA,eAAA;AAAA,EACA,aAAA;AAAA,EACA;AACF,CAAA;AASA,IAAM,oBAAA,kBAAsB,IAAI,GAAA,CAAI;AAAA,EAClC,mBAAA;AAAA,EACA,eAAA;AAAA,EACA,qBAAA;AAAA,EACA,qBAAA;AAAA,EACA,oBAAA;AAAA,EACA,eAAA;AAAA,EACA;AACF,CAAC,CAAA;AAGD,IAAM,uBAAA,EAAmC;AAAA,EACvC,mBAAA;AAAA,EACA,sBAAA;AAAA,EACA,iBAAA;AAAA,EACA;AACF,CAAA;AAOA,IAAM,0BAAA,EAAsC,CAAC,aAAA,EAAe,kBAAA,EAAoB,aAAa,CAAA;AAa7F,IAAM,iBAAA,EAAmB,uBAAA;AAEzB,SAAS,eAAA,CAAgB,IAAA,EAAuB;AAC9C,EAAA,GAAA,CAAI,yBAAA,CAA0B,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,CAAA,EAAG;AACzD,IAAA,OAAO,KAAA;AAAA,EACT;AACA,EAAA,OAAO,sBAAA,CAAuB,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,EAAA,GAAK,gBAAA,CAAiB,IAAA,CAAK,IAAI,CAAA;AACzF;AAEA,IAAM,cAAA,EAAgB,IAAA;AAGtB,IAAI,iBAAA,EAAuC,IAAA;AAE3C,SAAS,aAAA,CAAA,EAA6B;AACpC,EAAA,GAAA,CAAI,CAAC,gBAAA,EAAkB;AACrB,IAAA,iBAAA,EAAmB,IAAI,GAAA;AAAA,MACrB,MAAA,CAAO,IAAA,CAAK,kBAAY,CAAA,CACrB,GAAA,CAAI,CAAC,IAAA,EAAA,GAAS,SAAA,CAAU,IAAI,CAAA,CAAE,IAAA,CAAK,CAAC,CAAA,CACpC,MAAA,CAAO,OAAO;AAAA,IACnB,CAAA;AAAA,EACF;AACA,EAAA,OAAO,gBAAA;AACT;AAqBA,IAAI,eAAA,EAAkD,IAAA;AACtD,IAAI,mBAAA,EAAsD,IAAA;AAE1D,SAAS,WAAA,CAAA,EAAwC;AAC/C,EAAA,GAAA,CAAI,CAAC,cAAA,EAAgB;AACnB,IAAA,eAAA,EAAiB,wCAAA,CAAqB;AAAA,EACxC;AACA,EAAA,OAAO,cAAA;AACT;AAEA,SAAS,YAAA,CAAa,IAAA,EAAsB;AAC1C,EAAA,OAAO,IAAA,CAAK,OAAA,CAAQ,qBAAA,EAAuB,MAAM,CAAA;AACnD;AAkBA,SAAS,gBAAA,CAAiB,cAAA,EAAgC;AACxD,EAAA,MAAM,KAAA,EAAO,cAAA,CAAe,KAAA,CAAM,GAAG,CAAA,CAAE,GAAA,CAAI,YAAY,CAAA,CAAE,IAAA,CAAK,YAAY,CAAA;AAC1E,EAAA,MAAM,OAAA,EAAS,SAAA,CAAU,IAAA,CAAK,cAAc,EAAA,EAAI,KAAA,EAAO,EAAA;AACvD,EAAA,OAAO,IAAI,MAAA,CAAO,CAAA,mBAAA,EAAsB,IAAI,CAAA,EAAA;AAC9C;AAEsE;AAC3B,EAAA;AACD,EAAA;AAC1C;AAOiD;AACb,EAAA;AACzB,IAAA;AACT,EAAA;AACkC,EAAA;AACpC;AAEqD;AAC1B,EAAA;AACkB,IAAA;AACJ,IAAA;AACD,MAAA;AACG,QAAA;AAClB,QAAA;AACf,UAAA;AACF,QAAA;AAC8B,QAAA;AAC5B,UAAA;AACF,QAAA;AAC6B,QAAA;AAClB,QAAA;AACyB,UAAA;AACV,UAAA;AAC1B,QAAA;AACkB,QAAA;AACW,UAAA;AACb,UAAA;AACwB,UAAA;AAC5B,UAAA;AACX,QAAA;AACH,MAAA;AACF,IAAA;AACqB,IAAA;AACvB,EAAA;AACO,EAAA;AACT;AAEuF;AACxC,EAAA;AACA,EAAA;AACP,EAAA;AACxC;AAEkD;AACV,EAAA;AACxC;AAMiD;AACjB,EAAA;AACrB,IAAA;AACT,EAAA;AACiC,EAAA;AACF,EAAA;AACU,IAAA;AAC9B,MAAA;AACT,IAAA;AACF,EAAA;AACO,EAAA;AACT;AASgE;AAC7B,EAAA;AACU,EAAA;AACL,IAAA;AAClC,MAAA;AACF,IAAA;AACoB,IAAA;AACU,IAAA;AACZ,IAAA;AACA,IAAA;AACmB,IAAA;AACE,IAAA;AACA,MAAA;AACtB,MAAA;AACkB,MAAA;AACnC,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAa+D;AAC9B,EAAA;AACU,IAAA;AACzC,EAAA;AAC+B,EAAA;AACC,EAAA;AACW,IAAA;AACd,IAAA;AAEE,MAAA;AAG7B,IAAA;AACgB,IAAA;AACS,MAAA;AACzB,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAUoE;AACvC,EAAA;AACa,EAAA;AACQ,EAAA;AACF,EAAA;AACF,EAAA;AACA,EAAA;AAEJ,EAAA;AACK,IAAA;AAEH,IAAA;AACb,IAAA;AACG,MAAA;AACO,MAAA;AACrC,IAAA;AACa,IAAA;AACX,MAAA;AACF,IAAA;AAEuC,IAAA;AACX,IAAA;AACE,MAAA;AACM,MAAA;AACpC,IAAA;AAEoC,IAAA;AACX,IAAA;AACgB,MAAA;AACR,MAAA;AACjC,IAAA;AAEuC,IAAA;AACH,MAAA;AAChC,QAAA;AACF,MAAA;AAEqC,MAAA;AACnC,QAAA;AACF,MAAA;AAEiB,MAAA;AACI,MAAA;AACN,QAAA;AACM,MAAA;AACN,QAAA;AACf,MAAA;AAEsC,MAAA;AAGvB,MAAA;AAEH,MAAA;AACiB,QAAA;AACb,UAAA;AACZ,UAAA;AACa,UAAA;AACM,UAAA;AACC,UAAA;AACrB,QAAA;AACH,MAAA;AACF,IAAA;AACF,EAAA;AAE6C,EAAA;AACb,EAAA;AAEzB,EAAA;AACc,IAAA;AACnB,IAAA;AACO,IAAA;AACiB,MAAA;AACtB,MAAA;AAC6B,MAAA;AAC/B,IAAA;AACF,EAAA;AACF;AAKgE;AACvB,EAAA;AACzC;AD9M+C;AACA;AACA;AACA;AACA;AACA;AACA","file":"/home/runner/work/fhir-brasil/fhir-brasil/packages/ocr-utils/dist/index.cjs","sourcesContent":[null,"/**\n * OCR Anchor — Biomarker text anchoring\n *\n * Scans OCR text for biomarker names BEFORE sending to LLM.\n * This prevents hallucination by constraining what biomarkers\n * the LLM is allowed to extract.\n *\n * Matching is deliberately conservative: a name only anchors when it appears\n * as a whole token, is not swallowed by a longer biomarker name, is not inside\n * a genetic report line, and — for generic single-word names — sits on a line\n * that actually carries a value.\n */\n\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\n  UNIT_TO_UCUM,\n} from '@precisa-saude/fhir';\n\nexport interface AnchorMatch {\n  code: string;\n  confidence: number;\n  loinc?: string;\n  matchedName: string;\n  position: number;\n}\n\nexport interface AnchorResult {\n  filteredReference: string;\n  matches: AnchorMatch[];\n  stats: {\n    totalPatterns: number;\n    matchedCount: number;\n    scanTimeMs: number;\n  };\n}\n\n/**\n * Confidence assigned to a specific biomarker name found on a line that also\n * carries a value (a number, a unit, or an expected qualitative term).\n */\nexport const CONFIDENCE_VALUE_ADJACENT = 1.0;\n\n/**\n * Confidence assigned to a specific biomarker name with no value evidence\n * nearby — a section heading, or a mention in prose.\n */\nexport const CONFIDENCE_NAME_ONLY = 0.7;\n\n/**\n * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,\n * `Blood`, …) that only anchored because a value was found next to it.\n */\nexport const CONFIDENCE_AMBIGUOUS = 0.4;\n\n/** Cap on how many occurrences of the same name are inspected per document. */\nconst MAX_OCCURRENCES_PER_NAME = 5;\n\n/**\n * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Collapses horizontal whitespace, but KEEPS line breaks — the line is the\n *   context window used to decide whether a match is a real biomarker mention\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/[^\\S\\n]+/g, ' ');\n}\n\nconst UNAMBIGUOUS_SHORT_NAMES = new Set([\n  'hdl',\n  'ldl',\n  'lh',\n  'tsh',\n  'crp',\n  'pcr',\n  'ggt',\n  'alt',\n  'ast',\n  'bun',\n  'wbc',\n  'rbc',\n  'mcv',\n  'mch',\n  'rdw',\n  'mpv',\n  'psa',\n  'fsh',\n  'hba1c',\n  'egfr',\n  'acr',\n  'esr',\n  'vhs',\n  'bmc',\n  'bmd',\n  'vat',\n  'dxa',\n  'dmo',\n  'cmo',\n  'ffm',\n  'lbm',\n  'mlg',\n  'tav',\n]);\n\n/**\n * Single-word catalog names that are ordinary words in EN/PT, so seeing them\n * proves nothing on its own. They only anchor when the line also carries a\n * value. Qualitative urine markers (`Color`, `Protein`, `Blood`, …) are\n * detected automatically — see `isQualitativeUrine` — and don't belong here.\n */\nconst CONTEXT_REQUIRED_NAMES = new Set([\n  'bacteria', // Bacteria_Urine — tem unidade, escapa da regra automática\n  'bacterias', // Bacteria_Urine\n  'lead', // Lead — verbo/substantivo comuníssimo em inglês\n  'peso', // TotalMass\n  'saturation', // TransferrinSaturation — \"oxygen saturation\", \"saturation index\"\n  'tap', // ProthrombinTime — \"tap\" em inglês\n  'volume', // VATVolume\n  'weight', // TotalMass\n  // Sítios de dobra pelo nome nu. São partes do corpo antes de serem medidas,\n  // e aparecem em prosa: num laudo de DEXA real, \"hips and thighs\" e\n  // \"abdominal region\" ancoravam dobra cutânea que o documento não tem.\n  // Exigir valor na linha separa a tabela do parágrafo.\n  'abdominal',\n  'chest',\n  'coxa',\n  'peitoral',\n  'subescapular',\n  'subscapular',\n  'suprailiac',\n  'thigh',\n  'triceps',\n  'tricipital',\n]);\n\n/**\n * Qualitative results expected next to a non-numeric biomarker\n * (urine dipstick, sediment, appearance). Normalized, single tokens —\n * \"não reagente\" is covered by `reagente`, \"não detectado\" by `detectado`.\n */\nconst QUALITATIVE_VALUE_TERMS = new Set([\n  'absent',\n  'alguns',\n  'amarela',\n  'amarelo',\n  'anormal',\n  'ausencia',\n  'ausente',\n  'ausentes',\n  'citrino',\n  'claro',\n  'clear',\n  'cloudy',\n  'colorless',\n  'detectado',\n  'detected',\n  'escuro',\n  'incolor',\n  'indetectavel',\n  'limpido',\n  'moderada',\n  'moderado',\n  'negativa',\n  'negative',\n  'negativo',\n  'normais',\n  'normal',\n  'numerosos',\n  'ocasional',\n  'positiva',\n  'positive',\n  'positivo',\n  'present',\n  'presente',\n  'presentes',\n  'raras',\n  'raro',\n  'raros',\n  'reagente',\n  'trace',\n  'traces',\n  'tracos',\n  'turvo',\n  'undetectable',\n  'yellow',\n]);\n\n/**\n * Signals that a line comes from a genetic/molecular report rather than from a\n * panel of measured values. Gene symbols collide with biomarker names (`APOB`\n * the gene vs. `ApoB` the lipoprotein), so the context — not a static HGNC\n * blocklist — is what tells them apart. Blocking the token itself would break\n * real lipid panels.\n */\nconst GENETIC_CONTEXT_PATTERNS: RegExp[] = [\n  /\\b[nx][mrpc]_\\d{6,}/, // RefSeq: NM_000384.2, NP_, NR_, XM_\n  /\\bens[gtp]\\d{6,}/, // Ensembl: ENSG00000084674\n  /\\bp\\.[a-z]{3}\\d/, // HGVS proteína: p.Trp448*\n  /\\bc\\.\\d+[acgt]?[>_+-]/, // HGVS codificante: c.1234A>G, c.76_78del\n  /\\brs\\d{4,}\\b/, // dbSNP\n  /\\bgenes?\\b/,\n  /\\bvariante?s?\\b/,\n  /\\bexons?\\b/,\n  /\\bzygosity\\b/,\n  /\\bzigosidade\\b/,\n  /\\balleles?\\b/,\n  /\\balelos?\\b/,\n  /\\bmutations?\\b/,\n  /\\bmutac(ao|oes)\\b/,\n  /\\bpathogenic/,\n  /\\bpatogenic/,\n  /\\bheterozyg/,\n  /\\bhomozyg/,\n  /\\bheterozigot/,\n  /\\bhomozigot/,\n  /\\bsequence change\\b/,\n];\n\n/**\n * Sítios de dobra cutânea cujo nome nu também nomeia uma circunferência:\n * \"Coxa\" aparece tanto em \"Dobra Cutânea Coxa\" quanto em \"Circunferência da\n * Coxa\". O termo nu precisa existir como alias, porque há laudo que imprime\n * só o sítio na coluna, então a desambiguação tem que vir do contexto da\n * linha, como já se faz com laudo genético.\n */\nconst SKINFOLD_SITE_CODES = new Set([\n  'SkinfoldAbdominal',\n  'SkinfoldChest',\n  'SkinfoldMidaxillary',\n  'SkinfoldSubscapular',\n  'SkinfoldSuprailiac',\n  'SkinfoldThigh',\n  'SkinfoldTriceps',\n]);\n\n/** Uma linha de circunferência ou perímetro não mede dobra. */\nconst GIRTH_CONTEXT_PATTERNS: RegExp[] = [\n  /\\bcircumference\\b/,\n  /\\bcircunferencias?\\b/,\n  /\\bperimetros?\\b/,\n  /\\bgirth\\b/,\n];\n\n/**\n * Só bloqueia quando a linha fala de circunferência e não fala de dobra:\n * \"Dobra Cutânea Coxa\" e \"Thigh Skinfold\" continuam ancorando normalmente,\n * e uma linha que traga as duas palavras é ambígua demais para descartar.\n */\nconst SKINFOLD_CONTEXT_PATTERNS: RegExp[] = [/\\bdobras?\\b/, /\\bskin ?folds?\\b/, /\\bpregas?\\b/];\n\n/**\n * Medida em centímetros numa linha de sítio corporal.\n *\n * Dobra cutânea é em milímetros, sempre: um valor em cm no mesmo sítio é\n * circunferência. É o desambiguador mais forte que existe aqui, porque não\n * depende de a folha escrever a palavra \"circunferência\", e num laudo de\n * antropometria a coluna costuma trazer só o sítio e o número.\n *\n * Rejeita cm em vez de exigir mm: há folha que imprime a unidade no cabeçalho\n * da coluna e não em cada linha, e exigir mm perderia essas.\n */\nconst CENTIMETRE_VALUE = /\\d\\s*(?:,\\d+\\s*)?cm\\b/;\n\nfunction hasGirthContext(line: string): boolean {\n  if (SKINFOLD_CONTEXT_PATTERNS.some((re) => re.test(line))) {\n    return false;\n  }\n  return GIRTH_CONTEXT_PATTERNS.some((re) => re.test(line)) || CENTIMETRE_VALUE.test(line);\n}\n\nconst DIGIT_PATTERN = /\\d/;\n\n/** Unit tokens reused from the core catalog instead of a parallel list. */\nlet cachedUnitTokens: Set<string> | null = null;\n\nfunction getUnitTokens(): Set<string> {\n  if (!cachedUnitTokens) {\n    cachedUnitTokens = new Set(\n      Object.keys(UNIT_TO_UCUM)\n        .map((unit) => normalize(unit).trim())\n        .filter(Boolean),\n    );\n  }\n  return cachedUnitTokens;\n}\n\ninterface PatternEntry {\n  ambiguous: boolean;\n  code: string;\n  loinc?: string;\n  original: string;\n}\n\ninterface NamePattern {\n  entries: PatternEntry[];\n  /** Built on first use — most names never match a given document. */\n  regex: RegExp | null;\n}\n\ninterface Candidate {\n  end: number;\n  entries: PatternEntry[];\n  start: number;\n}\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNamePatterns: Map<string, NamePattern> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&');\n}\n\n/**\n * Build a whole-token matcher for a normalized name.\n *\n * Lookarounds instead of `\\b` because names may start or end with a non-word\n * character (`Lp(a)`), where `\\b` asserts the wrong thing.\n *\n * A multi-word name must sit on a single line: in the column layouts labs\n * print, consecutive lines are separate biomarkers, and allowing a line break\n * inside a name turns \"Colesterol\\nHDL\" into the name \"Colesterol HDL\".\n * A wrapped name still anchors through its head token when that token is a\n * name of its own (\"Colesterol\\nTotal\" → `Cholesterol`).\n *\n * The trailing optional `s` keeps the plurals labs actually print\n * (\"Proteínas\", \"Cetonas\") anchored to the singular catalog name — without\n * letting `proteína` match inside `proteinúria`.\n */\nfunction buildNamePattern(normalizedName: string): RegExp {\n  const body = normalizedName.split(' ').map(escapeRegExp).join('[^\\\\S\\\\n]+');\n  const plural = /\\p{L}$/u.test(normalizedName) ? 's?' : '';\n  return new RegExp(`(?<![\\\\p{L}\\\\p{N}])${body}${plural}(?![\\\\p{L}\\\\p{N}])`, 'gu');\n}\n\nfunction isQualitativeUrine(pattern: BiomarkerSearchPattern): boolean {\n  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];\n  return categories.includes('urina') && !pattern.unit;\n}\n\n/**\n * A name is ambiguous when it is a single token that also reads as ordinary\n * text. Multi-word names (`Occult Blood`, `Urine Protein`) are specific enough\n * on their own.\n */\nfunction isAmbiguousName(normalizedName: string, pattern: BiomarkerSearchPattern): boolean {\n  if (normalizedName.includes(' ')) {\n    return false;\n  }\n  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);\n}\n\nfunction getNamePatterns(): Map<string, NamePattern> {\n  if (!cachedNamePatterns) {\n    const map = new Map<string, NamePattern>();\n    for (const pattern of getPatterns()) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name).trim();\n        if (!normalized) {\n          continue;\n        }\n        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {\n          continue;\n        }\n        let slot = map.get(normalized);\n        if (!slot) {\n          slot = { entries: [], regex: null };\n          map.set(normalized, slot);\n        }\n        slot.entries.push({\n          ambiguous: isAmbiguousName(normalized, pattern),\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n      }\n    }\n    cachedNamePatterns = map;\n  }\n  return cachedNamePatterns;\n}\n\nfunction getLineBounds(text: string, position: number): { end: number; start: number } {\n  const start = text.lastIndexOf('\\n', position) + 1;\n  const nextBreak = text.indexOf('\\n', position);\n  return { end: nextBreak === -1 ? text.length : nextBreak, start };\n}\n\nfunction hasGeneticContext(line: string): boolean {\n  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));\n}\n\n/**\n * Does this line carry something that looks like a measured result?\n * A digit, a known unit, or an expected qualitative term.\n */\nfunction hasValueEvidence(line: string): boolean {\n  if (DIGIT_PATTERN.test(line)) {\n    return true;\n  }\n  const unitTokens = getUnitTokens();\n  for (const token of line.split(/[^\\p{L}\\p{N}%/]+/u)) {\n    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {\n      return true;\n    }\n  }\n  return false;\n}\n\n/**\n * Cheap pre-filter before the (much costlier) boundary regex.\n *\n * Sound because `normalize` collapses horizontal whitespace to a single space\n * and a name never spans a line break: whenever the pattern can match, the\n * literal name is a substring of the text.\n */\nfunction collectCandidates(normalizedText: string): Candidate[] {\n  const candidates: Candidate[] = [];\n  for (const [name, slot] of getNamePatterns()) {\n    if (!normalizedText.includes(name)) {\n      continue;\n    }\n    const { entries } = slot;\n    const regex = (slot.regex ??= buildNamePattern(name));\n    regex.lastIndex = 0;\n    let occurrences = 0;\n    let match = regex.exec(normalizedText);\n    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {\n      candidates.push({ end: match.index + match[0].length, entries, start: match.index });\n      occurrences += 1;\n      match = regex.exec(normalizedText);\n    }\n  }\n  return candidates;\n}\n\n/**\n * Longest match wins: drop a match fully contained in a longer one, so\n * `Cholesterol` doesn't anchor inside `HDL Cholesterol` and `Blood` doesn't\n * anchor inside `Blood Glucose`.\n *\n * Strictly longer, not longer-or-equal: containment plus equal length means an\n * identical span, which only happens when two distinct catalog names match the\n * same text (a singular and its plural form, say). Dropping one of those by\n * catalog order would silently lose a code, and losing an anchor is worse than\n * keeping both — `findBiomarkersInText` dedups per code anyway.\n */\nfunction resolveOverlaps(candidates: Candidate[]): Candidate[] {\n  const sorted = [...candidates].sort(\n    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start,\n  );\n  const accepted: Candidate[] = [];\n  for (const candidate of sorted) {\n    const length = candidate.end - candidate.start;\n    const swallowed = accepted.some(\n      (other) =>\n        other.start <= candidate.start &&\n        candidate.end <= other.end &&\n        other.end - other.start > length,\n    );\n    if (!swallowed) {\n      accepted.push(candidate);\n    }\n  }\n  return accepted;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n *\n * Matching is whole-token, longest-match-wins, and context-aware: matches\n * inside genetic report lines are discarded, and generic names only anchor\n * when a value sits on the same line. Returns one match per biomarker code —\n * the highest-confidence occurrence.\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const bestByCode = new Map<string, AnchorMatch>();\n  const geneticLines = new Map<number, boolean>();\n  const valueLines = new Map<number, boolean>();\n  const girthLines = new Map<number, boolean>();\n\n  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {\n    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);\n\n    let genetic = geneticLines.get(lineStart);\n    if (genetic === undefined) {\n      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));\n      geneticLines.set(lineStart, genetic);\n    }\n    if (genetic) {\n      continue;\n    }\n\n    let hasValue = valueLines.get(lineStart);\n    if (hasValue === undefined) {\n      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));\n      valueLines.set(lineStart, hasValue);\n    }\n\n    let girth = girthLines.get(lineStart);\n    if (girth === undefined) {\n      girth = hasGirthContext(normalizedText.slice(lineStart, lineEnd));\n      girthLines.set(lineStart, girth);\n    }\n\n    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\n        continue;\n      }\n\n      if (girth && SKINFOLD_SITE_CODES.has(entry.code)) {\n        continue;\n      }\n\n      let confidence = CONFIDENCE_NAME_ONLY;\n      if (entry.ambiguous) {\n        confidence = CONFIDENCE_AMBIGUOUS;\n      } else if (hasValue) {\n        confidence = CONFIDENCE_VALUE_ADJACENT;\n      }\n\n      const existing = bestByCode.get(entry.code);\n      const better =\n        !existing ||\n        confidence > existing.confidence ||\n        (confidence === existing.confidence && candidate.start < existing.position);\n      if (better) {\n        bestByCode.set(entry.code, {\n          code: entry.code,\n          confidence,\n          loinc: entry.loinc,\n          matchedName: entry.original,\n          position: candidate.start,\n        });\n      }\n    }\n  }\n\n  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);\n  const scanTimeMs = Date.now() - startTime;\n\n  return {\n    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),\n    matches,\n    stats: {\n      matchedCount: matches.length,\n      scanTimeMs,\n      totalPatterns: getPatterns().length,\n    },\n  };\n}\n\n/**\n * Get the list of matched biomarker codes from an anchor result.\n */\nexport function getMatchedCodes(result: AnchorResult): string[] {\n  return result.matches.map((m) => m.code);\n}\n"]}
+{"version":3,"sources":["/home/runner/work/fhir-brasil/fhir-brasil/packages/ocr-utils/dist/index.cjs","../src/anchor.ts"],"names":[],"mappings":"AAAA;ACaA;AAEE;AACA;AACA;AAAA,2CACK;AAwBA,IAAM,0BAAA,EAA4B,CAAA;AAMlC,IAAM,qBAAA,EAAuB,GAAA;AAM7B,IAAM,qBAAA,EAAuB,GAAA;AAGpC,IAAM,yBAAA,EAA2B,CAAA;AA8BjC,SAAS,SAAA,CAAU,IAAA,EAAsB;AACvC,EAAA,OAAO,IAAA,CACJ,SAAA,CAAU,KAAK,CAAA,CACf,OAAA,CAAQ,kBAAA,EAAoB,EAAE,CAAA,CAC9B,WAAA,CAAY,CAAA,CACZ,OAAA,CAAQ,+BAAA,EAAiC,GAAG,CAAA,CAC5C,OAAA,CAAQ,WAAA,EAAa,GAAG,CAAA;AAC7B;AAEA,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACt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uBAAA,kBAAyB,IAAI,GAAA,CAAI;AAAA,EACrC,UAAA;AAAA;AAAA,EACA,WAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,YAAA;AAAA;AAAA,EACA,KAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,UAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,YAAA;AAAA,EACA,OAAA;AAAA,EACA,SAAA;AAAA,EACA;AACF,CAAC,CAAA;AAOD,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACtC,QAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,OAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,cAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,OAAA;AAAA,EACA,UAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,QAAA;AAAA,EACA,OAAA;AAAA,EACA,cAAA;AAAA,EACA;AACF,CAAC,CAAA;AASD,IAAM,yBAAA,EAAqC;AAAA,EACzC,qBAAA;AAAA;AAAA,EACA,kBAAA;AAAA;AAAA,EACA,iBAAA;AAAA;AAAA,EACA,uBAAA;AAAA;AAAA,EACA,cAAA;AAAA;AAAA,EACA,YAAA;AAAA,EACA,iBAAA;AAAA,EACA,YAAA;AAAA,EACA,cAAA;AAAA,EACA,gBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,gBAAA;AAAA,EACA,mBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,aAAA;AAAA,EACA,WAAA;AAAA,EACA,eAAA;AAAA,EACA,aAAA;AAAA,EACA;AACF,CAAA;AASA,IAAM,oBAAA,kBAAsB,IAAI,GAAA,CAAI;AAAA,EAClC,mBAAA;AAAA,EACA,eAAA;AAAA,EACA,qBAAA;AAAA,EACA,qBAAA;AAAA,EACA,oBAAA;AAAA,EACA,eAAA;AAAA,EACA;AACF,CAAC,CAAA;AAGD,IAAM,uBAAA,EAAmC;AAAA,EACvC,mBAAA;AAAA,EACA,sBAAA;AAAA,EACA,iBAAA;AAAA,EACA;AACF,CAAA;AAOA,IAAM,0BAAA,EAAsC,CAAC,aAAA,EAAe,kBAAA,EAAoB,aAAa,CAAA;AAa7F,IAAM,iBAAA,EAAmB,uBAAA;AAEzB,SAAS,eAAA,CAAgB,IAAA,EAAuB;AAC9C,EAAA,GAAA,CAAI,yBAAA,CAA0B,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,CAAA,EAAG;AACzD,IAAA,OAAO,KAAA;AAAA,EACT;AACA,EAAA,OAAO,sBAAA,CAAuB,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,EAAA,GAAK,gBAAA,CAAiB,IAAA,CAAK,IAAI,CAAA;AACzF;AAEA,IAAM,cAAA,EAAgB,IAAA;AAGtB,IAAI,iBAAA,EAAuC,IAAA;AAE3C,SAAS,aAAA,CAAA,EAA6B;AACpC,EAAA,GAAA,CAAI,CAAC,gBAAA,EAAkB;AACrB,IAAA,iBAAA,EAAmB,IAAI,GAAA;AAAA,MACrB,MAAA,CAAO,IAAA,CAAK,kBAAY,CAAA,CACrB,GAAA,CAAI,CAAC,IAAA,EAAA,GAAS,SAAA,CAAU,IAAI,CAAA,CAAE,IAAA,CAAK,CAAC,CAAA,CACpC,MAAA,CAAO,OAAO;AAAA,IACnB,CAAA;AAAA,EACF;AACA,EAAA,OAAO,gBAAA;AACT;AAqBA,IAAI,eAAA,EAAkD,IAAA;AACtD,IAAI,mBAAA,EAAsD,IAAA;AAE1D,SAAS,WAAA,CAAA,EAAwC;AAC/C,EAAA,GAAA,CAAI,CAAC,cAAA,EAAgB;AACnB,IAAA,eAAA,EAAiB,wCAAA,CAAqB;AAAA,EACxC;AACA,EAAA,OAAO,cAAA;AACT;AAEA,SAAS,YAAA,CAAa,IAAA,EAAsB;AAC1C,EAAA,OAAO,IAAA,CAAK,OAAA,CAAQ,qBAAA,EAAuB,MAAM,CAAA;AACnD;AAkBA,SAAS,gBAAA,CAAiB,cAAA,EAAgC;AACxD,EAAA,MAAM,KAAA,EAAO,cAAA,CAAe,KAAA,CAAM,GAAG,CAAA,CAAE,GAAA,CAAI,YAAY,CAAA,CAAE,IAAA,CAAK,YAAY,CAAA;AAC1E,EAAA,MAAM,OAAA,EAAS,SAAA,CAAU,IAAA,CAAK,cAAc,EAAA,EAAI,KAAA,EAAO,EAAA;AACvD,EAAA,OAAO,IAAI,MAAA,CAAO,CAAA,mBAAA,EAAsB,IAAI,CAAA,EAAA;AAC9C;AAEsE;AAC3B,EAAA;AACD,EAAA;AAC1C;AAOiD;AACb,EAAA;AACzB,IAAA;AACT,EAAA;AACkC,EAAA;AACpC;AAEqD;AAC1B,EAAA;AACkB,IAAA;AACJ,IAAA;AACD,MAAA;AACG,QAAA;AAClB,QAAA;AACf,UAAA;AACF,QAAA;AAC8B,QAAA;AAC5B,UAAA;AACF,QAAA;AAC6B,QAAA;AAClB,QAAA;AACyB,UAAA;AACV,UAAA;AAC1B,QAAA;AACkB,QAAA;AACW,UAAA;AACb,UAAA;AACwB,UAAA;AAC5B,UAAA;AACX,QAAA;AACH,MAAA;AACF,IAAA;AACqB,IAAA;AACvB,EAAA;AACO,EAAA;AACT;AAEuF;AACxC,EAAA;AACA,EAAA;AACP,EAAA;AACxC;AAEkD;AACV,EAAA;AACxC;AAMiD;AACjB,EAAA;AACrB,IAAA;AACT,EAAA;AACiC,EAAA;AACF,EAAA;AACU,IAAA;AAC9B,MAAA;AACT,IAAA;AACF,EAAA;AACO,EAAA;AACT;AASgE;AAC7B,EAAA;AACU,EAAA;AACL,IAAA;AAClC,MAAA;AACF,IAAA;AACoB,IAAA;AACU,IAAA;AACZ,IAAA;AACA,IAAA;AACmB,IAAA;AACE,IAAA;AACA,MAAA;AACtB,MAAA;AACkB,MAAA;AACnC,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAa+D;AAC9B,EAAA;AACU,IAAA;AACzC,EAAA;AAC+B,EAAA;AACC,EAAA;AACW,IAAA;AACd,IAAA;AAEE,MAAA;AAG7B,IAAA;AACgB,IAAA;AACS,MAAA;AACzB,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAUoE;AACvC,EAAA;AACa,EAAA;AACQ,EAAA;AACF,EAAA;AACF,EAAA;AACA,EAAA;AAEJ,EAAA;AACK,IAAA;AAEH,IAAA;AACb,IAAA;AACG,MAAA;AACO,MAAA;AACrC,IAAA;AACa,IAAA;AACX,MAAA;AACF,IAAA;AAEuC,IAAA;AACX,IAAA;AACE,MAAA;AACM,MAAA;AACpC,IAAA;AAEoC,IAAA;AACX,IAAA;AACgB,MAAA;AACR,MAAA;AACjC,IAAA;AAEuC,IAAA;AACH,MAAA;AAChC,QAAA;AACF,MAAA;AAEqC,MAAA;AACnC,QAAA;AACF,MAAA;AAEiB,MAAA;AACI,MAAA;AACN,QAAA;AACM,MAAA;AACN,QAAA;AACf,MAAA;AAEsC,MAAA;AAGvB,MAAA;AAEH,MAAA;AACiB,QAAA;AACb,UAAA;AACZ,UAAA;AACa,UAAA;AACM,UAAA;AACC,UAAA;AACrB,QAAA;AACH,MAAA;AACF,IAAA;AACF,EAAA;AAE6C,EAAA;AACb,EAAA;AAEzB,EAAA;AACc,IAAA;AACnB,IAAA;AACO,IAAA;AACiB,MAAA;AACtB,MAAA;AAC6B,MAAA;AAC/B,IAAA;AACF,EAAA;AACF;AAKgE;AACvB,EAAA;AACzC;ADpO+C;AACA;AACA;AACA;AACA;AACA;AACA","file":"/home/runner/work/fhir-brasil/fhir-brasil/packages/ocr-utils/dist/index.cjs","sourcesContent":[null,"/**\n * OCR Anchor — Biomarker text anchoring\n *\n * Scans OCR text for biomarker names BEFORE sending to LLM.\n * This prevents hallucination by constraining what biomarkers\n * the LLM is allowed to extract.\n *\n * Matching is deliberately conservative: a name only anchors when it appears\n * as a whole token, is not swallowed by a longer biomarker name, is not inside\n * a genetic report line, and — for generic single-word names — sits on a line\n * that actually carries a value.\n */\n\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\n  UNIT_TO_UCUM,\n} from '@precisa-saude/fhir';\n\nexport interface AnchorMatch {\n  code: string;\n  confidence: number;\n  loinc?: string;\n  matchedName: string;\n  position: number;\n}\n\nexport interface AnchorResult {\n  filteredReference: string;\n  matches: AnchorMatch[];\n  stats: {\n    totalPatterns: number;\n    matchedCount: number;\n    scanTimeMs: number;\n  };\n}\n\n/**\n * Confidence assigned to a specific biomarker name found on a line that also\n * carries a value (a number, a unit, or an expected qualitative term).\n */\nexport const CONFIDENCE_VALUE_ADJACENT = 1.0;\n\n/**\n * Confidence assigned to a specific biomarker name with no value evidence\n * nearby — a section heading, or a mention in prose.\n */\nexport const CONFIDENCE_NAME_ONLY = 0.7;\n\n/**\n * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,\n * `Blood`, …) that only anchored because a value was found next to it.\n */\nexport const CONFIDENCE_AMBIGUOUS = 0.4;\n\n/** Cap on how many occurrences of the same name are inspected per document. */\nconst MAX_OCCURRENCES_PER_NAME = 5;\n\n/**\n * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Treats a hyphen that joins words as a space\n * - Collapses horizontal whitespace, but KEEPS line breaks — the line is the\n *   context window used to decide whether a match is a real biomarker mention\n *\n * O hífen entre palavras vira espaço porque o catálogo e o laboratório\n * discordam sobre ele o tempo todo: o catálogo escreve \"Proteína C-Reativa\" e\n * \"High-Density Lipoprotein\", e os laudos imprimem \"Proteína C Reativa\" e\n * \"High Density Lipoprotein\". Sem essa equivalência, 82 dos 159 nomes com\n * hífen deixam de ancorar na grafia que o documento usa.\n *\n * Não era teórico: um GGT de verdade foi descartado como alucinação em 27\n * laudos porque o documento escrevia \"Gama glutamil transferase\" e o catálogo\n * \"Gama-Glutamil Transferase\". Um caractere derrubava o valor antes de\n * qualquer validação.\n *\n * A troca exige **letra antes** do hífen, e por isso não toca em número:\n * o `-2.5` de um T-score e o `0-5` de uma faixa de urina seguem intactos.\n * Trocar sem essa guarda apagaria o sinal de um valor negativo, que é bem\n * pior que o problema original.\n *\n * O pré-filtro de substring do `collectCandidates` compara a chave com o texto\n * já normalizado, então a equivalência precisa nascer aqui: aplicada só na\n * regex, o `includes` descartaria o nome antes de ela rodar.\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/(?<=\\p{L})-(?=[\\p{L}\\p{N}])/gu, ' ')\n    .replace(/[^\\S\\n]+/g, ' ');\n}\n\nconst UNAMBIGUOUS_SHORT_NAMES = new Set([\n  'hdl',\n  'ldl',\n  'lh',\n  'tsh',\n  'crp',\n  'pcr',\n  'ggt',\n  'alt',\n  'ast',\n  'bun',\n  'wbc',\n  'rbc',\n  'mcv',\n  'mch',\n  'rdw',\n  'mpv',\n  'psa',\n  'fsh',\n  'hba1c',\n  'egfr',\n  'acr',\n  'esr',\n  'vhs',\n  'bmc',\n  'bmd',\n  'vat',\n  'dxa',\n  'dmo',\n  'cmo',\n  'ffm',\n  'lbm',\n  'mlg',\n  'tav',\n]);\n\n/**\n * Single-word catalog names that are ordinary words in EN/PT, so seeing them\n * proves nothing on its own. They only anchor when the line also carries a\n * value. Qualitative urine markers (`Color`, `Protein`, `Blood`, …) are\n * detected automatically — see `isQualitativeUrine` — and don't belong here.\n */\nconst CONTEXT_REQUIRED_NAMES = new Set([\n  'bacteria', // Bacteria_Urine — tem unidade, escapa da regra automática\n  'bacterias', // Bacteria_Urine\n  'lead', // Lead — verbo/substantivo comuníssimo em inglês\n  'peso', // TotalMass\n  'saturation', // TransferrinSaturation — \"oxygen saturation\", \"saturation index\"\n  'tap', // ProthrombinTime — \"tap\" em inglês\n  'volume', // VATVolume\n  'weight', // TotalMass\n  // Sítios de dobra pelo nome nu. São partes do corpo antes de serem medidas,\n  // e aparecem em prosa: num laudo de DEXA real, \"hips and thighs\" e\n  // \"abdominal region\" ancoravam dobra cutânea que o documento não tem.\n  // Exigir valor na linha separa a tabela do parágrafo.\n  'abdominal',\n  'chest',\n  'coxa',\n  'peitoral',\n  'subescapular',\n  'subscapular',\n  'suprailiac',\n  'thigh',\n  'triceps',\n  'tricipital',\n]);\n\n/**\n * Qualitative results expected next to a non-numeric biomarker\n * (urine dipstick, sediment, appearance). Normalized, single tokens —\n * \"não reagente\" is covered by `reagente`, \"não detectado\" by `detectado`.\n */\nconst QUALITATIVE_VALUE_TERMS = new Set([\n  'absent',\n  'alguns',\n  'amarela',\n  'amarelo',\n  'anormal',\n  'ausencia',\n  'ausente',\n  'ausentes',\n  'citrino',\n  'claro',\n  'clear',\n  'cloudy',\n  'colorless',\n  'detectado',\n  'detected',\n  'escuro',\n  'incolor',\n  'indetectavel',\n  'limpido',\n  'moderada',\n  'moderado',\n  'negativa',\n  'negative',\n  'negativo',\n  'normais',\n  'normal',\n  'numerosos',\n  'ocasional',\n  'positiva',\n  'positive',\n  'positivo',\n  'present',\n  'presente',\n  'presentes',\n  'raras',\n  'raro',\n  'raros',\n  'reagente',\n  'trace',\n  'traces',\n  'tracos',\n  'turvo',\n  'undetectable',\n  'yellow',\n]);\n\n/**\n * Signals that a line comes from a genetic/molecular report rather than from a\n * panel of measured values. Gene symbols collide with biomarker names (`APOB`\n * the gene vs. `ApoB` the lipoprotein), so the context — not a static HGNC\n * blocklist — is what tells them apart. Blocking the token itself would break\n * real lipid panels.\n */\nconst GENETIC_CONTEXT_PATTERNS: RegExp[] = [\n  /\\b[nx][mrpc]_\\d{6,}/, // RefSeq: NM_000384.2, NP_, NR_, XM_\n  /\\bens[gtp]\\d{6,}/, // Ensembl: ENSG00000084674\n  /\\bp\\.[a-z]{3}\\d/, // HGVS proteína: p.Trp448*\n  /\\bc\\.\\d+[acgt]?[>_+-]/, // HGVS codificante: c.1234A>G, c.76_78del\n  /\\brs\\d{4,}\\b/, // dbSNP\n  /\\bgenes?\\b/,\n  /\\bvariante?s?\\b/,\n  /\\bexons?\\b/,\n  /\\bzygosity\\b/,\n  /\\bzigosidade\\b/,\n  /\\balleles?\\b/,\n  /\\balelos?\\b/,\n  /\\bmutations?\\b/,\n  /\\bmutac(ao|oes)\\b/,\n  /\\bpathogenic/,\n  /\\bpatogenic/,\n  /\\bheterozyg/,\n  /\\bhomozyg/,\n  /\\bheterozigot/,\n  /\\bhomozigot/,\n  /\\bsequence change\\b/,\n];\n\n/**\n * Sítios de dobra cutânea cujo nome nu também nomeia uma circunferência:\n * \"Coxa\" aparece tanto em \"Dobra Cutânea Coxa\" quanto em \"Circunferência da\n * Coxa\". O termo nu precisa existir como alias, porque há laudo que imprime\n * só o sítio na coluna, então a desambiguação tem que vir do contexto da\n * linha, como já se faz com laudo genético.\n */\nconst SKINFOLD_SITE_CODES = new Set([\n  'SkinfoldAbdominal',\n  'SkinfoldChest',\n  'SkinfoldMidaxillary',\n  'SkinfoldSubscapular',\n  'SkinfoldSuprailiac',\n  'SkinfoldThigh',\n  'SkinfoldTriceps',\n]);\n\n/** Uma linha de circunferência ou perímetro não mede dobra. */\nconst GIRTH_CONTEXT_PATTERNS: RegExp[] = [\n  /\\bcircumference\\b/,\n  /\\bcircunferencias?\\b/,\n  /\\bperimetros?\\b/,\n  /\\bgirth\\b/,\n];\n\n/**\n * Só bloqueia quando a linha fala de circunferência e não fala de dobra:\n * \"Dobra Cutânea Coxa\" e \"Thigh Skinfold\" continuam ancorando normalmente,\n * e uma linha que traga as duas palavras é ambígua demais para descartar.\n */\nconst SKINFOLD_CONTEXT_PATTERNS: RegExp[] = [/\\bdobras?\\b/, /\\bskin ?folds?\\b/, /\\bpregas?\\b/];\n\n/**\n * Medida em centímetros numa linha de sítio corporal.\n *\n * Dobra cutânea é em milímetros, sempre: um valor em cm no mesmo sítio é\n * circunferência. É o desambiguador mais forte que existe aqui, porque não\n * depende de a folha escrever a palavra \"circunferência\", e num laudo de\n * antropometria a coluna costuma trazer só o sítio e o número.\n *\n * Rejeita cm em vez de exigir mm: há folha que imprime a unidade no cabeçalho\n * da coluna e não em cada linha, e exigir mm perderia essas.\n */\nconst CENTIMETRE_VALUE = /\\d\\s*(?:,\\d+\\s*)?cm\\b/;\n\nfunction hasGirthContext(line: string): boolean {\n  if (SKINFOLD_CONTEXT_PATTERNS.some((re) => re.test(line))) {\n    return false;\n  }\n  return GIRTH_CONTEXT_PATTERNS.some((re) => re.test(line)) || CENTIMETRE_VALUE.test(line);\n}\n\nconst DIGIT_PATTERN = /\\d/;\n\n/** Unit tokens reused from the core catalog instead of a parallel list. */\nlet cachedUnitTokens: Set<string> | null = null;\n\nfunction getUnitTokens(): Set<string> {\n  if (!cachedUnitTokens) {\n    cachedUnitTokens = new Set(\n      Object.keys(UNIT_TO_UCUM)\n        .map((unit) => normalize(unit).trim())\n        .filter(Boolean),\n    );\n  }\n  return cachedUnitTokens;\n}\n\ninterface PatternEntry {\n  ambiguous: boolean;\n  code: string;\n  loinc?: string;\n  original: string;\n}\n\ninterface NamePattern {\n  entries: PatternEntry[];\n  /** Built on first use — most names never match a given document. */\n  regex: RegExp | null;\n}\n\ninterface Candidate {\n  end: number;\n  entries: PatternEntry[];\n  start: number;\n}\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNamePatterns: Map<string, NamePattern> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&');\n}\n\n/**\n * Build a whole-token matcher for a normalized name.\n *\n * Lookarounds instead of `\\b` because names may start or end with a non-word\n * character (`Lp(a)`), where `\\b` asserts the wrong thing.\n *\n * A multi-word name must sit on a single line: in the column layouts labs\n * print, consecutive lines are separate biomarkers, and allowing a line break\n * inside a name turns \"Colesterol\\nHDL\" into the name \"Colesterol HDL\".\n * A wrapped name still anchors through its head token when that token is a\n * name of its own (\"Colesterol\\nTotal\" → `Cholesterol`).\n *\n * The trailing optional `s` keeps the plurals labs actually print\n * (\"Proteínas\", \"Cetonas\") anchored to the singular catalog name — without\n * letting `proteína` match inside `proteinúria`.\n */\nfunction buildNamePattern(normalizedName: string): RegExp {\n  const body = normalizedName.split(' ').map(escapeRegExp).join('[^\\\\S\\\\n]+');\n  const plural = /\\p{L}$/u.test(normalizedName) ? 's?' : '';\n  return new RegExp(`(?<![\\\\p{L}\\\\p{N}])${body}${plural}(?![\\\\p{L}\\\\p{N}])`, 'gu');\n}\n\nfunction isQualitativeUrine(pattern: BiomarkerSearchPattern): boolean {\n  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];\n  return categories.includes('urina') && !pattern.unit;\n}\n\n/**\n * A name is ambiguous when it is a single token that also reads as ordinary\n * text. Multi-word names (`Occult Blood`, `Urine Protein`) are specific enough\n * on their own.\n */\nfunction isAmbiguousName(normalizedName: string, pattern: BiomarkerSearchPattern): boolean {\n  if (normalizedName.includes(' ')) {\n    return false;\n  }\n  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);\n}\n\nfunction getNamePatterns(): Map<string, NamePattern> {\n  if (!cachedNamePatterns) {\n    const map = new Map<string, NamePattern>();\n    for (const pattern of getPatterns()) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name).trim();\n        if (!normalized) {\n          continue;\n        }\n        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {\n          continue;\n        }\n        let slot = map.get(normalized);\n        if (!slot) {\n          slot = { entries: [], regex: null };\n          map.set(normalized, slot);\n        }\n        slot.entries.push({\n          ambiguous: isAmbiguousName(normalized, pattern),\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n      }\n    }\n    cachedNamePatterns = map;\n  }\n  return cachedNamePatterns;\n}\n\nfunction getLineBounds(text: string, position: number): { end: number; start: number } {\n  const start = text.lastIndexOf('\\n', position) + 1;\n  const nextBreak = text.indexOf('\\n', position);\n  return { end: nextBreak === -1 ? text.length : nextBreak, start };\n}\n\nfunction hasGeneticContext(line: string): boolean {\n  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));\n}\n\n/**\n * Does this line carry something that looks like a measured result?\n * A digit, a known unit, or an expected qualitative term.\n */\nfunction hasValueEvidence(line: string): boolean {\n  if (DIGIT_PATTERN.test(line)) {\n    return true;\n  }\n  const unitTokens = getUnitTokens();\n  for (const token of line.split(/[^\\p{L}\\p{N}%/]+/u)) {\n    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {\n      return true;\n    }\n  }\n  return false;\n}\n\n/**\n * Cheap pre-filter before the (much costlier) boundary regex.\n *\n * Sound because `normalize` collapses horizontal whitespace to a single space\n * and a name never spans a line break: whenever the pattern can match, the\n * literal name is a substring of the text.\n */\nfunction collectCandidates(normalizedText: string): Candidate[] {\n  const candidates: Candidate[] = [];\n  for (const [name, slot] of getNamePatterns()) {\n    if (!normalizedText.includes(name)) {\n      continue;\n    }\n    const { entries } = slot;\n    const regex = (slot.regex ??= buildNamePattern(name));\n    regex.lastIndex = 0;\n    let occurrences = 0;\n    let match = regex.exec(normalizedText);\n    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {\n      candidates.push({ end: match.index + match[0].length, entries, start: match.index });\n      occurrences += 1;\n      match = regex.exec(normalizedText);\n    }\n  }\n  return candidates;\n}\n\n/**\n * Longest match wins: drop a match fully contained in a longer one, so\n * `Cholesterol` doesn't anchor inside `HDL Cholesterol` and `Blood` doesn't\n * anchor inside `Blood Glucose`.\n *\n * Strictly longer, not longer-or-equal: containment plus equal length means an\n * identical span, which only happens when two distinct catalog names match the\n * same text (a singular and its plural form, say). Dropping one of those by\n * catalog order would silently lose a code, and losing an anchor is worse than\n * keeping both — `findBiomarkersInText` dedups per code anyway.\n */\nfunction resolveOverlaps(candidates: Candidate[]): Candidate[] {\n  const sorted = [...candidates].sort(\n    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start,\n  );\n  const accepted: Candidate[] = [];\n  for (const candidate of sorted) {\n    const length = candidate.end - candidate.start;\n    const swallowed = accepted.some(\n      (other) =>\n        other.start <= candidate.start &&\n        candidate.end <= other.end &&\n        other.end - other.start > length,\n    );\n    if (!swallowed) {\n      accepted.push(candidate);\n    }\n  }\n  return accepted;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n *\n * Matching is whole-token, longest-match-wins, and context-aware: matches\n * inside genetic report lines are discarded, and generic names only anchor\n * when a value sits on the same line. Returns one match per biomarker code —\n * the highest-confidence occurrence.\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const bestByCode = new Map<string, AnchorMatch>();\n  const geneticLines = new Map<number, boolean>();\n  const valueLines = new Map<number, boolean>();\n  const girthLines = new Map<number, boolean>();\n\n  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {\n    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);\n\n    let genetic = geneticLines.get(lineStart);\n    if (genetic === undefined) {\n      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));\n      geneticLines.set(lineStart, genetic);\n    }\n    if (genetic) {\n      continue;\n    }\n\n    let hasValue = valueLines.get(lineStart);\n    if (hasValue === undefined) {\n      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));\n      valueLines.set(lineStart, hasValue);\n    }\n\n    let girth = girthLines.get(lineStart);\n    if (girth === undefined) {\n      girth = hasGirthContext(normalizedText.slice(lineStart, lineEnd));\n      girthLines.set(lineStart, girth);\n    }\n\n    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\n        continue;\n      }\n\n      if (girth && SKINFOLD_SITE_CODES.has(entry.code)) {\n        continue;\n      }\n\n      let confidence = CONFIDENCE_NAME_ONLY;\n      if (entry.ambiguous) {\n        confidence = CONFIDENCE_AMBIGUOUS;\n      } else if (hasValue) {\n        confidence = CONFIDENCE_VALUE_ADJACENT;\n      }\n\n      const existing = bestByCode.get(entry.code);\n      const better =\n        !existing ||\n        confidence > existing.confidence ||\n        (confidence === existing.confidence && candidate.start < existing.position);\n      if (better) {\n        bestByCode.set(entry.code, {\n          code: entry.code,\n          confidence,\n          loinc: entry.loinc,\n          matchedName: entry.original,\n          position: candidate.start,\n        });\n      }\n    }\n  }\n\n  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);\n  const scanTimeMs = Date.now() - startTime;\n\n  return {\n    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),\n    matches,\n    stats: {\n      matchedCount: matches.length,\n      scanTimeMs,\n      totalPatterns: getPatterns().length,\n    },\n  };\n}\n\n/**\n * Get the list of matched biomarker codes from an anchor result.\n */\nexport function getMatchedCodes(result: AnchorResult): string[] {\n  return result.matches.map((m) => m.code);\n}\n"]}

package/dist/index.js

@@ -9,7 +9,7 @@ var CONFIDENCE_NAME_ONLY = 0.7;
 var CONFIDENCE_AMBIGUOUS = 0.4;
 var MAX_OCCURRENCES_PER_NAME = 5;
 function normalize(text) {
-  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/[^\S\n]+/g, " ");
+  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/(?<=\p{L})-(?=[\p{L}\p{N}])/gu, " ").replace(/[^\S\n]+/g, " ");
 }
 var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "hdl",

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/anchor.ts"],"sourcesContent":["/**\n * OCR Anchor — Biomarker text anchoring\n *\n * Scans OCR text for biomarker names BEFORE sending to LLM.\n * This prevents hallucination by constraining what biomarkers\n * the LLM is allowed to extract.\n *\n * Matching is deliberately conservative: a name only anchors when it appears\n * as a whole token, is not swallowed by a longer biomarker name, is not inside\n * a genetic report line, and — for generic single-word names — sits on a line\n * that actually carries a value.\n */\n\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\n  UNIT_TO_UCUM,\n} from '@precisa-saude/fhir';\n\nexport interface AnchorMatch {\n  code: string;\n  confidence: number;\n  loinc?: string;\n  matchedName: string;\n  position: number;\n}\n\nexport interface AnchorResult {\n  filteredReference: string;\n  matches: AnchorMatch[];\n  stats: {\n    totalPatterns: number;\n    matchedCount: number;\n    scanTimeMs: number;\n  };\n}\n\n/**\n * Confidence assigned to a specific biomarker name found on a line that also\n * carries a value (a number, a unit, or an expected qualitative term).\n */\nexport const CONFIDENCE_VALUE_ADJACENT = 1.0;\n\n/**\n * Confidence assigned to a specific biomarker name with no value evidence\n * nearby — a section heading, or a mention in prose.\n */\nexport const CONFIDENCE_NAME_ONLY = 0.7;\n\n/**\n * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,\n * `Blood`, …) that only anchored because a value was found next to it.\n */\nexport const CONFIDENCE_AMBIGUOUS = 0.4;\n\n/** Cap on how many occurrences of the same name are inspected per document. */\nconst MAX_OCCURRENCES_PER_NAME = 5;\n\n/**\n * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Collapses horizontal whitespace, but KEEPS line breaks — the line is the\n *   context window used to decide whether a match is a real biomarker mention\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/[^\\S\\n]+/g, ' ');\n}\n\nconst UNAMBIGUOUS_SHORT_NAMES = new Set([\n  'hdl',\n  'ldl',\n  'lh',\n  'tsh',\n  'crp',\n  'pcr',\n  'ggt',\n  'alt',\n  'ast',\n  'bun',\n  'wbc',\n  'rbc',\n  'mcv',\n  'mch',\n  'rdw',\n  'mpv',\n  'psa',\n  'fsh',\n  'hba1c',\n  'egfr',\n  'acr',\n  'esr',\n  'vhs',\n  'bmc',\n  'bmd',\n  'vat',\n  'dxa',\n  'dmo',\n  'cmo',\n  'ffm',\n  'lbm',\n  'mlg',\n  'tav',\n]);\n\n/**\n * Single-word catalog names that are ordinary words in EN/PT, so seeing them\n * proves nothing on its own. They only anchor when the line also carries a\n * value. Qualitative urine markers (`Color`, `Protein`, `Blood`, …) are\n * detected automatically — see `isQualitativeUrine` — and don't belong here.\n */\nconst CONTEXT_REQUIRED_NAMES = new Set([\n  'bacteria', // Bacteria_Urine — tem unidade, escapa da regra automática\n  'bacterias', // Bacteria_Urine\n  'lead', // Lead — verbo/substantivo comuníssimo em inglês\n  'peso', // TotalMass\n  'saturation', // TransferrinSaturation — \"oxygen saturation\", \"saturation index\"\n  'tap', // ProthrombinTime — \"tap\" em inglês\n  'volume', // VATVolume\n  'weight', // TotalMass\n  // Sítios de dobra pelo nome nu. São partes do corpo antes de serem medidas,\n  // e aparecem em prosa: num laudo de DEXA real, \"hips and thighs\" e\n  // \"abdominal region\" ancoravam dobra cutânea que o documento não tem.\n  // Exigir valor na linha separa a tabela do parágrafo.\n  'abdominal',\n  'chest',\n  'coxa',\n  'peitoral',\n  'subescapular',\n  'subscapular',\n  'suprailiac',\n  'thigh',\n  'triceps',\n  'tricipital',\n]);\n\n/**\n * Qualitative results expected next to a non-numeric biomarker\n * (urine dipstick, sediment, appearance). Normalized, single tokens —\n * \"não reagente\" is covered by `reagente`, \"não detectado\" by `detectado`.\n */\nconst QUALITATIVE_VALUE_TERMS = new Set([\n  'absent',\n  'alguns',\n  'amarela',\n  'amarelo',\n  'anormal',\n  'ausencia',\n  'ausente',\n  'ausentes',\n  'citrino',\n  'claro',\n  'clear',\n  'cloudy',\n  'colorless',\n  'detectado',\n  'detected',\n  'escuro',\n  'incolor',\n  'indetectavel',\n  'limpido',\n  'moderada',\n  'moderado',\n  'negativa',\n  'negative',\n  'negativo',\n  'normais',\n  'normal',\n  'numerosos',\n  'ocasional',\n  'positiva',\n  'positive',\n  'positivo',\n  'present',\n  'presente',\n  'presentes',\n  'raras',\n  'raro',\n  'raros',\n  'reagente',\n  'trace',\n  'traces',\n  'tracos',\n  'turvo',\n  'undetectable',\n  'yellow',\n]);\n\n/**\n * Signals that a line comes from a genetic/molecular report rather than from a\n * panel of measured values. Gene symbols collide with biomarker names (`APOB`\n * the gene vs. `ApoB` the lipoprotein), so the context — not a static HGNC\n * blocklist — is what tells them apart. Blocking the token itself would break\n * real lipid panels.\n */\nconst GENETIC_CONTEXT_PATTERNS: RegExp[] = [\n  /\\b[nx][mrpc]_\\d{6,}/, // RefSeq: NM_000384.2, NP_, NR_, XM_\n  /\\bens[gtp]\\d{6,}/, // Ensembl: ENSG00000084674\n  /\\bp\\.[a-z]{3}\\d/, // HGVS proteína: p.Trp448*\n  /\\bc\\.\\d+[acgt]?[>_+-]/, // HGVS codificante: c.1234A>G, c.76_78del\n  /\\brs\\d{4,}\\b/, // dbSNP\n  /\\bgenes?\\b/,\n  /\\bvariante?s?\\b/,\n  /\\bexons?\\b/,\n  /\\bzygosity\\b/,\n  /\\bzigosidade\\b/,\n  /\\balleles?\\b/,\n  /\\balelos?\\b/,\n  /\\bmutations?\\b/,\n  /\\bmutac(ao|oes)\\b/,\n  /\\bpathogenic/,\n  /\\bpatogenic/,\n  /\\bheterozyg/,\n  /\\bhomozyg/,\n  /\\bheterozigot/,\n  /\\bhomozigot/,\n  /\\bsequence change\\b/,\n];\n\n/**\n * Sítios de dobra cutânea cujo nome nu também nomeia uma circunferência:\n * \"Coxa\" aparece tanto em \"Dobra Cutânea Coxa\" quanto em \"Circunferência da\n * Coxa\". O termo nu precisa existir como alias, porque há laudo que imprime\n * só o sítio na coluna, então a desambiguação tem que vir do contexto da\n * linha, como já se faz com laudo genético.\n */\nconst SKINFOLD_SITE_CODES = new Set([\n  'SkinfoldAbdominal',\n  'SkinfoldChest',\n  'SkinfoldMidaxillary',\n  'SkinfoldSubscapular',\n  'SkinfoldSuprailiac',\n  'SkinfoldThigh',\n  'SkinfoldTriceps',\n]);\n\n/** Uma linha de circunferência ou perímetro não mede dobra. */\nconst GIRTH_CONTEXT_PATTERNS: RegExp[] = [\n  /\\bcircumference\\b/,\n  /\\bcircunferencias?\\b/,\n  /\\bperimetros?\\b/,\n  /\\bgirth\\b/,\n];\n\n/**\n * Só bloqueia quando a linha fala de circunferência e não fala de dobra:\n * \"Dobra Cutânea Coxa\" e \"Thigh Skinfold\" continuam ancorando normalmente,\n * e uma linha que traga as duas palavras é ambígua demais para descartar.\n */\nconst SKINFOLD_CONTEXT_PATTERNS: RegExp[] = [/\\bdobras?\\b/, /\\bskin ?folds?\\b/, /\\bpregas?\\b/];\n\n/**\n * Medida em centímetros numa linha de sítio corporal.\n *\n * Dobra cutânea é em milímetros, sempre: um valor em cm no mesmo sítio é\n * circunferência. É o desambiguador mais forte que existe aqui, porque não\n * depende de a folha escrever a palavra \"circunferência\", e num laudo de\n * antropometria a coluna costuma trazer só o sítio e o número.\n *\n * Rejeita cm em vez de exigir mm: há folha que imprime a unidade no cabeçalho\n * da coluna e não em cada linha, e exigir mm perderia essas.\n */\nconst CENTIMETRE_VALUE = /\\d\\s*(?:,\\d+\\s*)?cm\\b/;\n\nfunction hasGirthContext(line: string): boolean {\n  if (SKINFOLD_CONTEXT_PATTERNS.some((re) => re.test(line))) {\n    return false;\n  }\n  return GIRTH_CONTEXT_PATTERNS.some((re) => re.test(line)) || CENTIMETRE_VALUE.test(line);\n}\n\nconst DIGIT_PATTERN = /\\d/;\n\n/** Unit tokens reused from the core catalog instead of a parallel list. */\nlet cachedUnitTokens: Set<string> | null = null;\n\nfunction getUnitTokens(): Set<string> {\n  if (!cachedUnitTokens) {\n    cachedUnitTokens = new Set(\n      Object.keys(UNIT_TO_UCUM)\n        .map((unit) => normalize(unit).trim())\n        .filter(Boolean),\n    );\n  }\n  return cachedUnitTokens;\n}\n\ninterface PatternEntry {\n  ambiguous: boolean;\n  code: string;\n  loinc?: string;\n  original: string;\n}\n\ninterface NamePattern {\n  entries: PatternEntry[];\n  /** Built on first use — most names never match a given document. */\n  regex: RegExp | null;\n}\n\ninterface Candidate {\n  end: number;\n  entries: PatternEntry[];\n  start: number;\n}\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNamePatterns: Map<string, NamePattern> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&');\n}\n\n/**\n * Build a whole-token matcher for a normalized name.\n *\n * Lookarounds instead of `\\b` because names may start or end with a non-word\n * character (`Lp(a)`), where `\\b` asserts the wrong thing.\n *\n * A multi-word name must sit on a single line: in the column layouts labs\n * print, consecutive lines are separate biomarkers, and allowing a line break\n * inside a name turns \"Colesterol\\nHDL\" into the name \"Colesterol HDL\".\n * A wrapped name still anchors through its head token when that token is a\n * name of its own (\"Colesterol\\nTotal\" → `Cholesterol`).\n *\n * The trailing optional `s` keeps the plurals labs actually print\n * (\"Proteínas\", \"Cetonas\") anchored to the singular catalog name — without\n * letting `proteína` match inside `proteinúria`.\n */\nfunction buildNamePattern(normalizedName: string): RegExp {\n  const body = normalizedName.split(' ').map(escapeRegExp).join('[^\\\\S\\\\n]+');\n  const plural = /\\p{L}$/u.test(normalizedName) ? 's?' : '';\n  return new RegExp(`(?<![\\\\p{L}\\\\p{N}])${body}${plural}(?![\\\\p{L}\\\\p{N}])`, 'gu');\n}\n\nfunction isQualitativeUrine(pattern: BiomarkerSearchPattern): boolean {\n  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];\n  return categories.includes('urina') && !pattern.unit;\n}\n\n/**\n * A name is ambiguous when it is a single token that also reads as ordinary\n * text. Multi-word names (`Occult Blood`, `Urine Protein`) are specific enough\n * on their own.\n */\nfunction isAmbiguousName(normalizedName: string, pattern: BiomarkerSearchPattern): boolean {\n  if (normalizedName.includes(' ')) {\n    return false;\n  }\n  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);\n}\n\nfunction getNamePatterns(): Map<string, NamePattern> {\n  if (!cachedNamePatterns) {\n    const map = new Map<string, NamePattern>();\n    for (const pattern of getPatterns()) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name).trim();\n        if (!normalized) {\n          continue;\n        }\n        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {\n          continue;\n        }\n        let slot = map.get(normalized);\n        if (!slot) {\n          slot = { entries: [], regex: null };\n          map.set(normalized, slot);\n        }\n        slot.entries.push({\n          ambiguous: isAmbiguousName(normalized, pattern),\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n      }\n    }\n    cachedNamePatterns = map;\n  }\n  return cachedNamePatterns;\n}\n\nfunction getLineBounds(text: string, position: number): { end: number; start: number } {\n  const start = text.lastIndexOf('\\n', position) + 1;\n  const nextBreak = text.indexOf('\\n', position);\n  return { end: nextBreak === -1 ? text.length : nextBreak, start };\n}\n\nfunction hasGeneticContext(line: string): boolean {\n  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));\n}\n\n/**\n * Does this line carry something that looks like a measured result?\n * A digit, a known unit, or an expected qualitative term.\n */\nfunction hasValueEvidence(line: string): boolean {\n  if (DIGIT_PATTERN.test(line)) {\n    return true;\n  }\n  const unitTokens = getUnitTokens();\n  for (const token of line.split(/[^\\p{L}\\p{N}%/]+/u)) {\n    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {\n      return true;\n    }\n  }\n  return false;\n}\n\n/**\n * Cheap pre-filter before the (much costlier) boundary regex.\n *\n * Sound because `normalize` collapses horizontal whitespace to a single space\n * and a name never spans a line break: whenever the pattern can match, the\n * literal name is a substring of the text.\n */\nfunction collectCandidates(normalizedText: string): Candidate[] {\n  const candidates: Candidate[] = [];\n  for (const [name, slot] of getNamePatterns()) {\n    if (!normalizedText.includes(name)) {\n      continue;\n    }\n    const { entries } = slot;\n    const regex = (slot.regex ??= buildNamePattern(name));\n    regex.lastIndex = 0;\n    let occurrences = 0;\n    let match = regex.exec(normalizedText);\n    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {\n      candidates.push({ end: match.index + match[0].length, entries, start: match.index });\n      occurrences += 1;\n      match = regex.exec(normalizedText);\n    }\n  }\n  return candidates;\n}\n\n/**\n * Longest match wins: drop a match fully contained in a longer one, so\n * `Cholesterol` doesn't anchor inside `HDL Cholesterol` and `Blood` doesn't\n * anchor inside `Blood Glucose`.\n *\n * Strictly longer, not longer-or-equal: containment plus equal length means an\n * identical span, which only happens when two distinct catalog names match the\n * same text (a singular and its plural form, say). Dropping one of those by\n * catalog order would silently lose a code, and losing an anchor is worse than\n * keeping both — `findBiomarkersInText` dedups per code anyway.\n */\nfunction resolveOverlaps(candidates: Candidate[]): Candidate[] {\n  const sorted = [...candidates].sort(\n    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start,\n  );\n  const accepted: Candidate[] = [];\n  for (const candidate of sorted) {\n    const length = candidate.end - candidate.start;\n    const swallowed = accepted.some(\n      (other) =>\n        other.start <= candidate.start &&\n        candidate.end <= other.end &&\n        other.end - other.start > length,\n    );\n    if (!swallowed) {\n      accepted.push(candidate);\n    }\n  }\n  return accepted;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n *\n * Matching is whole-token, longest-match-wins, and context-aware: matches\n * inside genetic report lines are discarded, and generic names only anchor\n * when a value sits on the same line. Returns one match per biomarker code —\n * the highest-confidence occurrence.\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const bestByCode = new Map<string, AnchorMatch>();\n  const geneticLines = new Map<number, boolean>();\n  const valueLines = new Map<number, boolean>();\n  const girthLines = new Map<number, boolean>();\n\n  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {\n    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);\n\n    let genetic = geneticLines.get(lineStart);\n    if (genetic === undefined) {\n      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));\n      geneticLines.set(lineStart, genetic);\n    }\n    if (genetic) {\n      continue;\n    }\n\n    let hasValue = valueLines.get(lineStart);\n    if (hasValue === undefined) {\n      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));\n      valueLines.set(lineStart, hasValue);\n    }\n\n    let girth = girthLines.get(lineStart);\n    if (girth === undefined) {\n      girth = hasGirthContext(normalizedText.slice(lineStart, lineEnd));\n      girthLines.set(lineStart, girth);\n    }\n\n    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\n        continue;\n      }\n\n      if (girth && SKINFOLD_SITE_CODES.has(entry.code)) {\n        continue;\n      }\n\n      let confidence = CONFIDENCE_NAME_ONLY;\n      if (entry.ambiguous) {\n        confidence = CONFIDENCE_AMBIGUOUS;\n      } else if (hasValue) {\n        confidence = CONFIDENCE_VALUE_ADJACENT;\n      }\n\n      const existing = bestByCode.get(entry.code);\n      const better =\n        !existing ||\n        confidence > existing.confidence ||\n        (confidence === existing.confidence && candidate.start < existing.position);\n      if (better) {\n        bestByCode.set(entry.code, {\n          code: entry.code,\n          confidence,\n          loinc: entry.loinc,\n          matchedName: entry.original,\n          position: candidate.start,\n        });\n      }\n    }\n  }\n\n  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);\n  const scanTimeMs = Date.now() - startTime;\n\n  return {\n    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),\n    matches,\n    stats: {\n      matchedCount: matches.length,\n      scanTimeMs,\n      totalPatterns: getPatterns().length,\n    },\n  };\n}\n\n/**\n * Get the list of matched biomarker codes from an anchor result.\n */\nexport function getMatchedCodes(result: AnchorResult): string[] {\n  return result.matches.map((m) => m.code);\n}\n"],"mappings":";AAaA;AAAA,EAEE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAwBA,IAAM,4BAA4B;AAMlC,IAAM,uBAAuB;AAM7B,IAAM,uBAAuB;AAGpC,IAAM,2BAA2B;AASjC,SAAS,UAAU,MAAsB;AACvC,SAAO,KACJ,UAAU,KAAK,EACf,QAAQ,oBAAoB,EAAE,EAC9B,YAAY,EACZ,QAAQ,aAAa,GAAG;AAC7B;AAEA,IAAM,0BAA0B,oBAAI,IAAI;AAAA,EACtC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAQD,IAAM,yBAAyB,oBAAI,IAAI;AAAA,EACrC;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAOD,IAAM,0BAA0B,oBAAI,IAAI;AAAA,EACtC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AASD,IAAM,2BAAqC;AAAA,EACzC;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AASA,IAAM,sBAAsB,oBAAI,IAAI;AAAA,EAClC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAGD,IAAM,yBAAmC;AAAA,EACvC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAOA,IAAM,4BAAsC,CAAC,eAAe,oBAAoB,aAAa;AAa7F,IAAM,mBAAmB;AAEzB,SAAS,gBAAgB,MAAuB;AAC9C,MAAI,0BAA0B,KAAK,CAAC,OAAO,GAAG,KAAK,IAAI,CAAC,GAAG;AACzD,WAAO;AAAA,EACT;AACA,SAAO,uBAAuB,KAAK,CAAC,OAAO,GAAG,KAAK,IAAI,CAAC,KAAK,iBAAiB,KAAK,IAAI;AACzF;AAEA,IAAM,gBAAgB;AAGtB,IAAI,mBAAuC;AAE3C,SAAS,gBAA6B;AACpC,MAAI,CAAC,kBAAkB;AACrB,uBAAmB,IAAI;AAAA,MACrB,OAAO,KAAK,YAAY,EACrB,IAAI,CAAC,SAAS,UAAU,IAAI,EAAE,KAAK,CAAC,EACpC,OAAO,OAAO;AAAA,IACnB;AAAA,EACF;AACA,SAAO;AACT;AAqBA,IAAI,iBAAkD;AACtD,IAAI,qBAAsD;AAE1D,SAAS,cAAwC;AAC/C,MAAI,CAAC,gBAAgB;AACnB,qBAAiB,qBAAqB;AAAA,EACxC;AACA,SAAO;AACT;AAEA,SAAS,aAAa,MAAsB;AAC1C,SAAO,KAAK,QAAQ,uBAAuB,MAAM;AACnD;AAkBA,SAAS,iBAAiB,gBAAgC;AACxD,QAAM,OAAO,eAAe,MAAM,GAAG,EAAE,IAAI,YAAY,EAAE,KAAK,YAAY;AAC1E,QAAM,SAAS,UAAU,KAAK,cAAc,IAAI,OAAO;AACvD,SAAO,IAAI,OAAO,sBAAsB,IAAI,GAAG,MAAM,sBAAsB,IAAI;AACjF;AAEA,SAAS,mBAAmB,SAA0C;AACpE,QAAM,aAAa,MAAM,QAAQ,QAAQ,QAAQ,IAAI,QAAQ,WAAW,CAAC,QAAQ,QAAQ;AACzF,SAAO,WAAW,SAAS,OAAO,KAAK,CAAC,QAAQ;AAClD;AAOA,SAAS,gBAAgB,gBAAwB,SAA0C;AACzF,MAAI,eAAe,SAAS,GAAG,GAAG;AAChC,WAAO;AAAA,EACT;AACA,SAAO,uBAAuB,IAAI,cAAc,KAAK,mBAAmB,OAAO;AACjF;AAEA,SAAS,kBAA4C;AACnD,MAAI,CAAC,oBAAoB;AACvB,UAAM,MAAM,oBAAI,IAAyB;AACzC,eAAW,WAAW,YAAY,GAAG;AACnC,iBAAW,QAAQ,QAAQ,OAAO;AAChC,cAAM,aAAa,UAAU,IAAI,EAAE,KAAK;AACxC,YAAI,CAAC,YAAY;AACf;AAAA,QACF;AACA,YAAI,WAAW,SAAS,KAAK,CAAC,wBAAwB,IAAI,UAAU,GAAG;AACrE;AAAA,QACF;AACA,YAAI,OAAO,IAAI,IAAI,UAAU;AAC7B,YAAI,CAAC,MAAM;AACT,iBAAO,EAAE,SAAS,CAAC,GAAG,OAAO,KAAK;AAClC,cAAI,IAAI,YAAY,IAAI;AAAA,QAC1B;AACA,aAAK,QAAQ,KAAK;AAAA,UAChB,WAAW,gBAAgB,YAAY,OAAO;AAAA,UAC9C,MAAM,QAAQ;AAAA,UACd,GAAI,QAAQ,SAAS,EAAE,OAAO,QAAQ,MAAM;AAAA,UAC5C,UAAU;AAAA,QACZ,CAAC;AAAA,MACH;AAAA,IACF;AACA,yBAAqB;AAAA,EACvB;AACA,SAAO;AACT;AAEA,SAAS,cAAc,MAAc,UAAkD;AACrF,QAAM,QAAQ,KAAK,YAAY,MAAM,QAAQ,IAAI;AACjD,QAAM,YAAY,KAAK,QAAQ,MAAM,QAAQ;AAC7C,SAAO,EAAE,KAAK,cAAc,KAAK,KAAK,SAAS,WAAW,MAAM;AAClE;AAEA,SAAS,kBAAkB,MAAuB;AAChD,SAAO,yBAAyB,KAAK,CAAC,YAAY,QAAQ,KAAK,IAAI,CAAC;AACtE;AAMA,SAAS,iBAAiB,MAAuB;AAC/C,MAAI,cAAc,KAAK,IAAI,GAAG;AAC5B,WAAO;AAAA,EACT;AACA,QAAM,aAAa,cAAc;AACjC,aAAW,SAAS,KAAK,MAAM,mBAAmB,GAAG;AACnD,QAAI,UAAU,WAAW,IAAI,KAAK,KAAK,wBAAwB,IAAI,KAAK,IAAI;AAC1E,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;AASA,SAAS,kBAAkB,gBAAqC;AAC9D,QAAM,aAA0B,CAAC;AACjC,aAAW,CAAC,MAAM,IAAI,KAAK,gBAAgB,GAAG;AAC5C,QAAI,CAAC,eAAe,SAAS,IAAI,GAAG;AAClC;AAAA,IACF;AACA,UAAM,EAAE,QAAQ,IAAI;AACpB,UAAM,QAAS,KAAK,UAAU,iBAAiB,IAAI;AACnD,UAAM,YAAY;AAClB,QAAI,cAAc;AAClB,QAAI,QAAQ,MAAM,KAAK,cAAc;AACrC,WAAO,UAAU,QAAQ,cAAc,0BAA0B;AAC/D,iBAAW,KAAK,EAAE,KAAK,MAAM,QAAQ,MAAM,CAAC,EAAE,QAAQ,SAAS,OAAO,MAAM,MAAM,CAAC;AACnF,qBAAe;AACf,cAAQ,MAAM,KAAK,cAAc;AAAA,IACnC;AAAA,EACF;AACA,SAAO;AACT;AAaA,SAAS,gBAAgB,YAAsC;AAC7D,QAAM,SAAS,CAAC,GAAG,UAAU,EAAE;AAAA,IAC7B,CAAC,GAAG,MAAM,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,UAAU,EAAE,QAAQ,EAAE;AAAA,EAC/D;AACA,QAAM,WAAwB,CAAC;AAC/B,aAAW,aAAa,QAAQ;AAC9B,UAAM,SAAS,UAAU,MAAM,UAAU;AACzC,UAAM,YAAY,SAAS;AAAA,MACzB,CAAC,UACC,MAAM,SAAS,UAAU,SACzB,UAAU,OAAO,MAAM,OACvB,MAAM,MAAM,MAAM,QAAQ;AAAA,IAC9B;AACA,QAAI,CAAC,WAAW;AACd,eAAS,KAAK,SAAS;AAAA,IACzB;AAAA,EACF;AACA,SAAO;AACT;AAUO,SAAS,qBAAqB,SAA+B;AAClE,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,iBAAiB,UAAU,OAAO;AACxC,QAAM,aAAa,oBAAI,IAAyB;AAChD,QAAM,eAAe,oBAAI,IAAqB;AAC9C,QAAM,aAAa,oBAAI,IAAqB;AAC5C,QAAM,aAAa,oBAAI,IAAqB;AAE5C,aAAW,aAAa,gBAAgB,kBAAkB,cAAc,CAAC,GAAG;AAC1E,UAAM,EAAE,KAAK,SAAS,OAAO,UAAU,IAAI,cAAc,gBAAgB,UAAU,KAAK;AAExF,QAAI,UAAU,aAAa,IAAI,SAAS;AACxC,QAAI,YAAY,QAAW;AACzB,gBAAU,kBAAkB,eAAe,MAAM,WAAW,OAAO,CAAC;AACpE,mBAAa,IAAI,WAAW,OAAO;AAAA,IACrC;AACA,QAAI,SAAS;AACX;AAAA,IACF;AAEA,QAAI,WAAW,WAAW,IAAI,SAAS;AACvC,QAAI,aAAa,QAAW;AAC1B,iBAAW,iBAAiB,eAAe,MAAM,WAAW,OAAO,CAAC;AACpE,iBAAW,IAAI,WAAW,QAAQ;AAAA,IACpC;AAEA,QAAI,QAAQ,WAAW,IAAI,SAAS;AACpC,QAAI,UAAU,QAAW;AACvB,cAAQ,gBAAgB,eAAe,MAAM,WAAW,OAAO,CAAC;AAChE,iBAAW,IAAI,WAAW,KAAK;AAAA,IACjC;AAEA,eAAW,SAAS,UAAU,SAAS;AACrC,UAAI,MAAM,aAAa,CAAC,UAAU;AAChC;AAAA,MACF;AAEA,UAAI,SAAS,oBAAoB,IAAI,MAAM,IAAI,GAAG;AAChD;AAAA,MACF;AAEA,UAAI,aAAa;AACjB,UAAI,MAAM,WAAW;AACnB,qBAAa;AAAA,MACf,WAAW,UAAU;AACnB,qBAAa;AAAA,MACf;AAEA,YAAM,WAAW,WAAW,IAAI,MAAM,IAAI;AAC1C,YAAM,SACJ,CAAC,YACD,aAAa,SAAS,cACrB,eAAe,SAAS,cAAc,UAAU,QAAQ,SAAS;AACpE,UAAI,QAAQ;AACV,mBAAW,IAAI,MAAM,MAAM;AAAA,UACzB,MAAM,MAAM;AAAA,UACZ;AAAA,UACA,OAAO,MAAM;AAAA,UACb,aAAa,MAAM;AAAA,UACnB,UAAU,UAAU;AAAA,QACtB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAEA,QAAM,UAAU,MAAM,KAAK,WAAW,OAAO,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AACtF,QAAM,aAAa,KAAK,IAAI,IAAI;AAEhC,SAAO;AAAA,IACL,mBAAmB,6BAA6B,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC;AAAA,IAC1E;AAAA,IACA,OAAO;AAAA,MACL,cAAc,QAAQ;AAAA,MACtB;AAAA,MACA,eAAe,YAAY,EAAE;AAAA,IAC/B;AAAA,EACF;AACF;AAKO,SAAS,gBAAgB,QAAgC;AAC9D,SAAO,OAAO,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI;AACzC;","names":[]}
+{"version":3,"sources":["../src/anchor.ts"],"sourcesContent":["/**\n * OCR Anchor — Biomarker text anchoring\n *\n * Scans OCR text for biomarker names BEFORE sending to LLM.\n * This prevents hallucination by constraining what biomarkers\n * the LLM is allowed to extract.\n *\n * Matching is deliberately conservative: a name only anchors when it appears\n * as a whole token, is not swallowed by a longer biomarker name, is not inside\n * a genetic report line, and — for generic single-word names — sits on a line\n * that actually carries a value.\n */\n\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\n  UNIT_TO_UCUM,\n} from '@precisa-saude/fhir';\n\nexport interface AnchorMatch {\n  code: string;\n  confidence: number;\n  loinc?: string;\n  matchedName: string;\n  position: number;\n}\n\nexport interface AnchorResult {\n  filteredReference: string;\n  matches: AnchorMatch[];\n  stats: {\n    totalPatterns: number;\n    matchedCount: number;\n    scanTimeMs: number;\n  };\n}\n\n/**\n * Confidence assigned to a specific biomarker name found on a line that also\n * carries a value (a number, a unit, or an expected qualitative term).\n */\nexport const CONFIDENCE_VALUE_ADJACENT = 1.0;\n\n/**\n * Confidence assigned to a specific biomarker name with no value evidence\n * nearby — a section heading, or a mention in prose.\n */\nexport const CONFIDENCE_NAME_ONLY = 0.7;\n\n/**\n * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,\n * `Blood`, …) that only anchored because a value was found next to it.\n */\nexport const CONFIDENCE_AMBIGUOUS = 0.4;\n\n/** Cap on how many occurrences of the same name are inspected per document. */\nconst MAX_OCCURRENCES_PER_NAME = 5;\n\n/**\n * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Treats a hyphen that joins words as a space\n * - Collapses horizontal whitespace, but KEEPS line breaks — the line is the\n *   context window used to decide whether a match is a real biomarker mention\n *\n * O hífen entre palavras vira espaço porque o catálogo e o laboratório\n * discordam sobre ele o tempo todo: o catálogo escreve \"Proteína C-Reativa\" e\n * \"High-Density Lipoprotein\", e os laudos imprimem \"Proteína C Reativa\" e\n * \"High Density Lipoprotein\". Sem essa equivalência, 82 dos 159 nomes com\n * hífen deixam de ancorar na grafia que o documento usa.\n *\n * Não era teórico: um GGT de verdade foi descartado como alucinação em 27\n * laudos porque o documento escrevia \"Gama glutamil transferase\" e o catálogo\n * \"Gama-Glutamil Transferase\". Um caractere derrubava o valor antes de\n * qualquer validação.\n *\n * A troca exige **letra antes** do hífen, e por isso não toca em número:\n * o `-2.5` de um T-score e o `0-5` de uma faixa de urina seguem intactos.\n * Trocar sem essa guarda apagaria o sinal de um valor negativo, que é bem\n * pior que o problema original.\n *\n * O pré-filtro de substring do `collectCandidates` compara a chave com o texto\n * já normalizado, então a equivalência precisa nascer aqui: aplicada só na\n * regex, o `includes` descartaria o nome antes de ela rodar.\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/(?<=\\p{L})-(?=[\\p{L}\\p{N}])/gu, ' ')\n    .replace(/[^\\S\\n]+/g, ' ');\n}\n\nconst UNAMBIGUOUS_SHORT_NAMES = new Set([\n  'hdl',\n  'ldl',\n  'lh',\n  'tsh',\n  'crp',\n  'pcr',\n  'ggt',\n  'alt',\n  'ast',\n  'bun',\n  'wbc',\n  'rbc',\n  'mcv',\n  'mch',\n  'rdw',\n  'mpv',\n  'psa',\n  'fsh',\n  'hba1c',\n  'egfr',\n  'acr',\n  'esr',\n  'vhs',\n  'bmc',\n  'bmd',\n  'vat',\n  'dxa',\n  'dmo',\n  'cmo',\n  'ffm',\n  'lbm',\n  'mlg',\n  'tav',\n]);\n\n/**\n * Single-word catalog names that are ordinary words in EN/PT, so seeing them\n * proves nothing on its own. They only anchor when the line also carries a\n * value. Qualitative urine markers (`Color`, `Protein`, `Blood`, …) are\n * detected automatically — see `isQualitativeUrine` — and don't belong here.\n */\nconst CONTEXT_REQUIRED_NAMES = new Set([\n  'bacteria', // Bacteria_Urine — tem unidade, escapa da regra automática\n  'bacterias', // Bacteria_Urine\n  'lead', // Lead — verbo/substantivo comuníssimo em inglês\n  'peso', // TotalMass\n  'saturation', // TransferrinSaturation — \"oxygen saturation\", \"saturation index\"\n  'tap', // ProthrombinTime — \"tap\" em inglês\n  'volume', // VATVolume\n  'weight', // TotalMass\n  // Sítios de dobra pelo nome nu. São partes do corpo antes de serem medidas,\n  // e aparecem em prosa: num laudo de DEXA real, \"hips and thighs\" e\n  // \"abdominal region\" ancoravam dobra cutânea que o documento não tem.\n  // Exigir valor na linha separa a tabela do parágrafo.\n  'abdominal',\n  'chest',\n  'coxa',\n  'peitoral',\n  'subescapular',\n  'subscapular',\n  'suprailiac',\n  'thigh',\n  'triceps',\n  'tricipital',\n]);\n\n/**\n * Qualitative results expected next to a non-numeric biomarker\n * (urine dipstick, sediment, appearance). Normalized, single tokens —\n * \"não reagente\" is covered by `reagente`, \"não detectado\" by `detectado`.\n */\nconst QUALITATIVE_VALUE_TERMS = new Set([\n  'absent',\n  'alguns',\n  'amarela',\n  'amarelo',\n  'anormal',\n  'ausencia',\n  'ausente',\n  'ausentes',\n  'citrino',\n  'claro',\n  'clear',\n  'cloudy',\n  'colorless',\n  'detectado',\n  'detected',\n  'escuro',\n  'incolor',\n  'indetectavel',\n  'limpido',\n  'moderada',\n  'moderado',\n  'negativa',\n  'negative',\n  'negativo',\n  'normais',\n  'normal',\n  'numerosos',\n  'ocasional',\n  'positiva',\n  'positive',\n  'positivo',\n  'present',\n  'presente',\n  'presentes',\n  'raras',\n  'raro',\n  'raros',\n  'reagente',\n  'trace',\n  'traces',\n  'tracos',\n  'turvo',\n  'undetectable',\n  'yellow',\n]);\n\n/**\n * Signals that a line comes from a genetic/molecular report rather than from a\n * panel of measured values. Gene symbols collide with biomarker names (`APOB`\n * the gene vs. `ApoB` the lipoprotein), so the context — not a static HGNC\n * blocklist — is what tells them apart. Blocking the token itself would break\n * real lipid panels.\n */\nconst GENETIC_CONTEXT_PATTERNS: RegExp[] = [\n  /\\b[nx][mrpc]_\\d{6,}/, // RefSeq: NM_000384.2, NP_, NR_, XM_\n  /\\bens[gtp]\\d{6,}/, // Ensembl: ENSG00000084674\n  /\\bp\\.[a-z]{3}\\d/, // HGVS proteína: p.Trp448*\n  /\\bc\\.\\d+[acgt]?[>_+-]/, // HGVS codificante: c.1234A>G, c.76_78del\n  /\\brs\\d{4,}\\b/, // dbSNP\n  /\\bgenes?\\b/,\n  /\\bvariante?s?\\b/,\n  /\\bexons?\\b/,\n  /\\bzygosity\\b/,\n  /\\bzigosidade\\b/,\n  /\\balleles?\\b/,\n  /\\balelos?\\b/,\n  /\\bmutations?\\b/,\n  /\\bmutac(ao|oes)\\b/,\n  /\\bpathogenic/,\n  /\\bpatogenic/,\n  /\\bheterozyg/,\n  /\\bhomozyg/,\n  /\\bheterozigot/,\n  /\\bhomozigot/,\n  /\\bsequence change\\b/,\n];\n\n/**\n * Sítios de dobra cutânea cujo nome nu também nomeia uma circunferência:\n * \"Coxa\" aparece tanto em \"Dobra Cutânea Coxa\" quanto em \"Circunferência da\n * Coxa\". O termo nu precisa existir como alias, porque há laudo que imprime\n * só o sítio na coluna, então a desambiguação tem que vir do contexto da\n * linha, como já se faz com laudo genético.\n */\nconst SKINFOLD_SITE_CODES = new Set([\n  'SkinfoldAbdominal',\n  'SkinfoldChest',\n  'SkinfoldMidaxillary',\n  'SkinfoldSubscapular',\n  'SkinfoldSuprailiac',\n  'SkinfoldThigh',\n  'SkinfoldTriceps',\n]);\n\n/** Uma linha de circunferência ou perímetro não mede dobra. */\nconst GIRTH_CONTEXT_PATTERNS: RegExp[] = [\n  /\\bcircumference\\b/,\n  /\\bcircunferencias?\\b/,\n  /\\bperimetros?\\b/,\n  /\\bgirth\\b/,\n];\n\n/**\n * Só bloqueia quando a linha fala de circunferência e não fala de dobra:\n * \"Dobra Cutânea Coxa\" e \"Thigh Skinfold\" continuam ancorando normalmente,\n * e uma linha que traga as duas palavras é ambígua demais para descartar.\n */\nconst SKINFOLD_CONTEXT_PATTERNS: RegExp[] = [/\\bdobras?\\b/, /\\bskin ?folds?\\b/, /\\bpregas?\\b/];\n\n/**\n * Medida em centímetros numa linha de sítio corporal.\n *\n * Dobra cutânea é em milímetros, sempre: um valor em cm no mesmo sítio é\n * circunferência. É o desambiguador mais forte que existe aqui, porque não\n * depende de a folha escrever a palavra \"circunferência\", e num laudo de\n * antropometria a coluna costuma trazer só o sítio e o número.\n *\n * Rejeita cm em vez de exigir mm: há folha que imprime a unidade no cabeçalho\n * da coluna e não em cada linha, e exigir mm perderia essas.\n */\nconst CENTIMETRE_VALUE = /\\d\\s*(?:,\\d+\\s*)?cm\\b/;\n\nfunction hasGirthContext(line: string): boolean {\n  if (SKINFOLD_CONTEXT_PATTERNS.some((re) => re.test(line))) {\n    return false;\n  }\n  return GIRTH_CONTEXT_PATTERNS.some((re) => re.test(line)) || CENTIMETRE_VALUE.test(line);\n}\n\nconst DIGIT_PATTERN = /\\d/;\n\n/** Unit tokens reused from the core catalog instead of a parallel list. */\nlet cachedUnitTokens: Set<string> | null = null;\n\nfunction getUnitTokens(): Set<string> {\n  if (!cachedUnitTokens) {\n    cachedUnitTokens = new Set(\n      Object.keys(UNIT_TO_UCUM)\n        .map((unit) => normalize(unit).trim())\n        .filter(Boolean),\n    );\n  }\n  return cachedUnitTokens;\n}\n\ninterface PatternEntry {\n  ambiguous: boolean;\n  code: string;\n  loinc?: string;\n  original: string;\n}\n\ninterface NamePattern {\n  entries: PatternEntry[];\n  /** Built on first use — most names never match a given document. */\n  regex: RegExp | null;\n}\n\ninterface Candidate {\n  end: number;\n  entries: PatternEntry[];\n  start: number;\n}\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNamePatterns: Map<string, NamePattern> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&');\n}\n\n/**\n * Build a whole-token matcher for a normalized name.\n *\n * Lookarounds instead of `\\b` because names may start or end with a non-word\n * character (`Lp(a)`), where `\\b` asserts the wrong thing.\n *\n * A multi-word name must sit on a single line: in the column layouts labs\n * print, consecutive lines are separate biomarkers, and allowing a line break\n * inside a name turns \"Colesterol\\nHDL\" into the name \"Colesterol HDL\".\n * A wrapped name still anchors through its head token when that token is a\n * name of its own (\"Colesterol\\nTotal\" → `Cholesterol`).\n *\n * The trailing optional `s` keeps the plurals labs actually print\n * (\"Proteínas\", \"Cetonas\") anchored to the singular catalog name — without\n * letting `proteína` match inside `proteinúria`.\n */\nfunction buildNamePattern(normalizedName: string): RegExp {\n  const body = normalizedName.split(' ').map(escapeRegExp).join('[^\\\\S\\\\n]+');\n  const plural = /\\p{L}$/u.test(normalizedName) ? 's?' : '';\n  return new RegExp(`(?<![\\\\p{L}\\\\p{N}])${body}${plural}(?![\\\\p{L}\\\\p{N}])`, 'gu');\n}\n\nfunction isQualitativeUrine(pattern: BiomarkerSearchPattern): boolean {\n  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];\n  return categories.includes('urina') && !pattern.unit;\n}\n\n/**\n * A name is ambiguous when it is a single token that also reads as ordinary\n * text. Multi-word names (`Occult Blood`, `Urine Protein`) are specific enough\n * on their own.\n */\nfunction isAmbiguousName(normalizedName: string, pattern: BiomarkerSearchPattern): boolean {\n  if (normalizedName.includes(' ')) {\n    return false;\n  }\n  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);\n}\n\nfunction getNamePatterns(): Map<string, NamePattern> {\n  if (!cachedNamePatterns) {\n    const map = new Map<string, NamePattern>();\n    for (const pattern of getPatterns()) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name).trim();\n        if (!normalized) {\n          continue;\n        }\n        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {\n          continue;\n        }\n        let slot = map.get(normalized);\n        if (!slot) {\n          slot = { entries: [], regex: null };\n          map.set(normalized, slot);\n        }\n        slot.entries.push({\n          ambiguous: isAmbiguousName(normalized, pattern),\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n      }\n    }\n    cachedNamePatterns = map;\n  }\n  return cachedNamePatterns;\n}\n\nfunction getLineBounds(text: string, position: number): { end: number; start: number } {\n  const start = text.lastIndexOf('\\n', position) + 1;\n  const nextBreak = text.indexOf('\\n', position);\n  return { end: nextBreak === -1 ? text.length : nextBreak, start };\n}\n\nfunction hasGeneticContext(line: string): boolean {\n  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));\n}\n\n/**\n * Does this line carry something that looks like a measured result?\n * A digit, a known unit, or an expected qualitative term.\n */\nfunction hasValueEvidence(line: string): boolean {\n  if (DIGIT_PATTERN.test(line)) {\n    return true;\n  }\n  const unitTokens = getUnitTokens();\n  for (const token of line.split(/[^\\p{L}\\p{N}%/]+/u)) {\n    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {\n      return true;\n    }\n  }\n  return false;\n}\n\n/**\n * Cheap pre-filter before the (much costlier) boundary regex.\n *\n * Sound because `normalize` collapses horizontal whitespace to a single space\n * and a name never spans a line break: whenever the pattern can match, the\n * literal name is a substring of the text.\n */\nfunction collectCandidates(normalizedText: string): Candidate[] {\n  const candidates: Candidate[] = [];\n  for (const [name, slot] of getNamePatterns()) {\n    if (!normalizedText.includes(name)) {\n      continue;\n    }\n    const { entries } = slot;\n    const regex = (slot.regex ??= buildNamePattern(name));\n    regex.lastIndex = 0;\n    let occurrences = 0;\n    let match = regex.exec(normalizedText);\n    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {\n      candidates.push({ end: match.index + match[0].length, entries, start: match.index });\n      occurrences += 1;\n      match = regex.exec(normalizedText);\n    }\n  }\n  return candidates;\n}\n\n/**\n * Longest match wins: drop a match fully contained in a longer one, so\n * `Cholesterol` doesn't anchor inside `HDL Cholesterol` and `Blood` doesn't\n * anchor inside `Blood Glucose`.\n *\n * Strictly longer, not longer-or-equal: containment plus equal length means an\n * identical span, which only happens when two distinct catalog names match the\n * same text (a singular and its plural form, say). Dropping one of those by\n * catalog order would silently lose a code, and losing an anchor is worse than\n * keeping both — `findBiomarkersInText` dedups per code anyway.\n */\nfunction resolveOverlaps(candidates: Candidate[]): Candidate[] {\n  const sorted = [...candidates].sort(\n    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start,\n  );\n  const accepted: Candidate[] = [];\n  for (const candidate of sorted) {\n    const length = candidate.end - candidate.start;\n    const swallowed = accepted.some(\n      (other) =>\n        other.start <= candidate.start &&\n        candidate.end <= other.end &&\n        other.end - other.start > length,\n    );\n    if (!swallowed) {\n      accepted.push(candidate);\n    }\n  }\n  return accepted;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n *\n * Matching is whole-token, longest-match-wins, and context-aware: matches\n * inside genetic report lines are discarded, and generic names only anchor\n * when a value sits on the same line. Returns one match per biomarker code —\n * the highest-confidence occurrence.\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const bestByCode = new Map<string, AnchorMatch>();\n  const geneticLines = new Map<number, boolean>();\n  const valueLines = new Map<number, boolean>();\n  const girthLines = new Map<number, boolean>();\n\n  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {\n    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);\n\n    let genetic = geneticLines.get(lineStart);\n    if (genetic === undefined) {\n      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));\n      geneticLines.set(lineStart, genetic);\n    }\n    if (genetic) {\n      continue;\n    }\n\n    let hasValue = valueLines.get(lineStart);\n    if (hasValue === undefined) {\n      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));\n      valueLines.set(lineStart, hasValue);\n    }\n\n    let girth = girthLines.get(lineStart);\n    if (girth === undefined) {\n      girth = hasGirthContext(normalizedText.slice(lineStart, lineEnd));\n      girthLines.set(lineStart, girth);\n    }\n\n    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\n        continue;\n      }\n\n      if (girth && SKINFOLD_SITE_CODES.has(entry.code)) {\n        continue;\n      }\n\n      let confidence = CONFIDENCE_NAME_ONLY;\n      if (entry.ambiguous) {\n        confidence = CONFIDENCE_AMBIGUOUS;\n      } else if (hasValue) {\n        confidence = CONFIDENCE_VALUE_ADJACENT;\n      }\n\n      const existing = bestByCode.get(entry.code);\n      const better =\n        !existing ||\n        confidence > existing.confidence ||\n        (confidence === existing.confidence && candidate.start < existing.position);\n      if (better) {\n        bestByCode.set(entry.code, {\n          code: entry.code,\n          confidence,\n          loinc: entry.loinc,\n          matchedName: entry.original,\n          position: candidate.start,\n        });\n      }\n    }\n  }\n\n  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);\n  const scanTimeMs = Date.now() - startTime;\n\n  return {\n    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),\n    matches,\n    stats: {\n      matchedCount: matches.length,\n      scanTimeMs,\n      totalPatterns: getPatterns().length,\n    },\n  };\n}\n\n/**\n * Get the list of matched biomarker codes from an anchor result.\n */\nexport function getMatchedCodes(result: AnchorResult): string[] {\n  return result.matches.map((m) => m.code);\n}\n"],"mappings":";AAaA;AAAA,EAEE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAwBA,IAAM,4BAA4B;AAMlC,IAAM,uBAAuB;AAM7B,IAAM,uBAAuB;AAGpC,IAAM,2BAA2B;AA8BjC,SAAS,UAAU,MAAsB;AACvC,SAAO,KACJ,UAAU,KAAK,EACf,QAAQ,oBAAoB,EAAE,EAC9B,YAAY,EACZ,QAAQ,iCAAiC,GAAG,EAC5C,QAAQ,aAAa,GAAG;AAC7B;AAEA,IAAM,0BAA0B,oBAAI,IAAI;AAAA,EACtC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAQD,IAAM,yBAAyB,oBAAI,IAAI;AAAA,EACrC;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAOD,IAAM,0BAA0B,oBAAI,IAAI;AAAA,EACtC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AASD,IAAM,2BAAqC;AAAA,EACzC;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AASA,IAAM,sBAAsB,oBAAI,IAAI;AAAA,EAClC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAGD,IAAM,yBAAmC;AAAA,EACvC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAOA,IAAM,4BAAsC,CAAC,eAAe,oBAAoB,aAAa;AAa7F,IAAM,mBAAmB;AAEzB,SAAS,gBAAgB,MAAuB;AAC9C,MAAI,0BAA0B,KAAK,CAAC,OAAO,GAAG,KAAK,IAAI,CAAC,GAAG;AACzD,WAAO;AAAA,EACT;AACA,SAAO,uBAAuB,KAAK,CAAC,OAAO,GAAG,KAAK,IAAI,CAAC,KAAK,iBAAiB,KAAK,IAAI;AACzF;AAEA,IAAM,gBAAgB;AAGtB,IAAI,mBAAuC;AAE3C,SAAS,gBAA6B;AACpC,MAAI,CAAC,kBAAkB;AACrB,uBAAmB,IAAI;AAAA,MACrB,OAAO,KAAK,YAAY,EACrB,IAAI,CAAC,SAAS,UAAU,IAAI,EAAE,KAAK,CAAC,EACpC,OAAO,OAAO;AAAA,IACnB;AAAA,EACF;AACA,SAAO;AACT;AAqBA,IAAI,iBAAkD;AACtD,IAAI,qBAAsD;AAE1D,SAAS,cAAwC;AAC/C,MAAI,CAAC,gBAAgB;AACnB,qBAAiB,qBAAqB;AAAA,EACxC;AACA,SAAO;AACT;AAEA,SAAS,aAAa,MAAsB;AAC1C,SAAO,KAAK,QAAQ,uBAAuB,MAAM;AACnD;AAkBA,SAAS,iBAAiB,gBAAgC;AACxD,QAAM,OAAO,eAAe,MAAM,GAAG,EAAE,IAAI,YAAY,EAAE,KAAK,YAAY;AAC1E,QAAM,SAAS,UAAU,KAAK,cAAc,IAAI,OAAO;AACvD,SAAO,IAAI,OAAO,sBAAsB,IAAI,GAAG,MAAM,sBAAsB,IAAI;AACjF;AAEA,SAAS,mBAAmB,SAA0C;AACpE,QAAM,aAAa,MAAM,QAAQ,QAAQ,QAAQ,IAAI,QAAQ,WAAW,CAAC,QAAQ,QAAQ;AACzF,SAAO,WAAW,SAAS,OAAO,KAAK,CAAC,QAAQ;AAClD;AAOA,SAAS,gBAAgB,gBAAwB,SAA0C;AACzF,MAAI,eAAe,SAAS,GAAG,GAAG;AAChC,WAAO;AAAA,EACT;AACA,SAAO,uBAAuB,IAAI,cAAc,KAAK,mBAAmB,OAAO;AACjF;AAEA,SAAS,kBAA4C;AACnD,MAAI,CAAC,oBAAoB;AACvB,UAAM,MAAM,oBAAI,IAAyB;AACzC,eAAW,WAAW,YAAY,GAAG;AACnC,iBAAW,QAAQ,QAAQ,OAAO;AAChC,cAAM,aAAa,UAAU,IAAI,EAAE,KAAK;AACxC,YAAI,CAAC,YAAY;AACf;AAAA,QACF;AACA,YAAI,WAAW,SAAS,KAAK,CAAC,wBAAwB,IAAI,UAAU,GAAG;AACrE;AAAA,QACF;AACA,YAAI,OAAO,IAAI,IAAI,UAAU;AAC7B,YAAI,CAAC,MAAM;AACT,iBAAO,EAAE,SAAS,CAAC,GAAG,OAAO,KAAK;AAClC,cAAI,IAAI,YAAY,IAAI;AAAA,QAC1B;AACA,aAAK,QAAQ,KAAK;AAAA,UAChB,WAAW,gBAAgB,YAAY,OAAO;AAAA,UAC9C,MAAM,QAAQ;AAAA,UACd,GAAI,QAAQ,SAAS,EAAE,OAAO,QAAQ,MAAM;AAAA,UAC5C,UAAU;AAAA,QACZ,CAAC;AAAA,MACH;AAAA,IACF;AACA,yBAAqB;AAAA,EACvB;AACA,SAAO;AACT;AAEA,SAAS,cAAc,MAAc,UAAkD;AACrF,QAAM,QAAQ,KAAK,YAAY,MAAM,QAAQ,IAAI;AACjD,QAAM,YAAY,KAAK,QAAQ,MAAM,QAAQ;AAC7C,SAAO,EAAE,KAAK,cAAc,KAAK,KAAK,SAAS,WAAW,MAAM;AAClE;AAEA,SAAS,kBAAkB,MAAuB;AAChD,SAAO,yBAAyB,KAAK,CAAC,YAAY,QAAQ,KAAK,IAAI,CAAC;AACtE;AAMA,SAAS,iBAAiB,MAAuB;AAC/C,MAAI,cAAc,KAAK,IAAI,GAAG;AAC5B,WAAO;AAAA,EACT;AACA,QAAM,aAAa,cAAc;AACjC,aAAW,SAAS,KAAK,MAAM,mBAAmB,GAAG;AACnD,QAAI,UAAU,WAAW,IAAI,KAAK,KAAK,wBAAwB,IAAI,KAAK,IAAI;AAC1E,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;AASA,SAAS,kBAAkB,gBAAqC;AAC9D,QAAM,aAA0B,CAAC;AACjC,aAAW,CAAC,MAAM,IAAI,KAAK,gBAAgB,GAAG;AAC5C,QAAI,CAAC,eAAe,SAAS,IAAI,GAAG;AAClC;AAAA,IACF;AACA,UAAM,EAAE,QAAQ,IAAI;AACpB,UAAM,QAAS,KAAK,UAAU,iBAAiB,IAAI;AACnD,UAAM,YAAY;AAClB,QAAI,cAAc;AAClB,QAAI,QAAQ,MAAM,KAAK,cAAc;AACrC,WAAO,UAAU,QAAQ,cAAc,0BAA0B;AAC/D,iBAAW,KAAK,EAAE,KAAK,MAAM,QAAQ,MAAM,CAAC,EAAE,QAAQ,SAAS,OAAO,MAAM,MAAM,CAAC;AACnF,qBAAe;AACf,cAAQ,MAAM,KAAK,cAAc;AAAA,IACnC;AAAA,EACF;AACA,SAAO;AACT;AAaA,SAAS,gBAAgB,YAAsC;AAC7D,QAAM,SAAS,CAAC,GAAG,UAAU,EAAE;AAAA,IAC7B,CAAC,GAAG,MAAM,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,UAAU,EAAE,QAAQ,EAAE;AAAA,EAC/D;AACA,QAAM,WAAwB,CAAC;AAC/B,aAAW,aAAa,QAAQ;AAC9B,UAAM,SAAS,UAAU,MAAM,UAAU;AACzC,UAAM,YAAY,SAAS;AAAA,MACzB,CAAC,UACC,MAAM,SAAS,UAAU,SACzB,UAAU,OAAO,MAAM,OACvB,MAAM,MAAM,MAAM,QAAQ;AAAA,IAC9B;AACA,QAAI,CAAC,WAAW;AACd,eAAS,KAAK,SAAS;AAAA,IACzB;AAAA,EACF;AACA,SAAO;AACT;AAUO,SAAS,qBAAqB,SAA+B;AAClE,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,iBAAiB,UAAU,OAAO;AACxC,QAAM,aAAa,oBAAI,IAAyB;AAChD,QAAM,eAAe,oBAAI,IAAqB;AAC9C,QAAM,aAAa,oBAAI,IAAqB;AAC5C,QAAM,aAAa,oBAAI,IAAqB;AAE5C,aAAW,aAAa,gBAAgB,kBAAkB,cAAc,CAAC,GAAG;AAC1E,UAAM,EAAE,KAAK,SAAS,OAAO,UAAU,IAAI,cAAc,gBAAgB,UAAU,KAAK;AAExF,QAAI,UAAU,aAAa,IAAI,SAAS;AACxC,QAAI,YAAY,QAAW;AACzB,gBAAU,kBAAkB,eAAe,MAAM,WAAW,OAAO,CAAC;AACpE,mBAAa,IAAI,WAAW,OAAO;AAAA,IACrC;AACA,QAAI,SAAS;AACX;AAAA,IACF;AAEA,QAAI,WAAW,WAAW,IAAI,SAAS;AACvC,QAAI,aAAa,QAAW;AAC1B,iBAAW,iBAAiB,eAAe,MAAM,WAAW,OAAO,CAAC;AACpE,iBAAW,IAAI,WAAW,QAAQ;AAAA,IACpC;AAEA,QAAI,QAAQ,WAAW,IAAI,SAAS;AACpC,QAAI,UAAU,QAAW;AACvB,cAAQ,gBAAgB,eAAe,MAAM,WAAW,OAAO,CAAC;AAChE,iBAAW,IAAI,WAAW,KAAK;AAAA,IACjC;AAEA,eAAW,SAAS,UAAU,SAAS;AACrC,UAAI,MAAM,aAAa,CAAC,UAAU;AAChC;AAAA,MACF;AAEA,UAAI,SAAS,oBAAoB,IAAI,MAAM,IAAI,GAAG;AAChD;AAAA,MACF;AAEA,UAAI,aAAa;AACjB,UAAI,MAAM,WAAW;AACnB,qBAAa;AAAA,MACf,WAAW,UAAU;AACnB,qBAAa;AAAA,MACf;AAEA,YAAM,WAAW,WAAW,IAAI,MAAM,IAAI;AAC1C,YAAM,SACJ,CAAC,YACD,aAAa,SAAS,cACrB,eAAe,SAAS,cAAc,UAAU,QAAQ,SAAS;AACpE,UAAI,QAAQ;AACV,mBAAW,IAAI,MAAM,MAAM;AAAA,UACzB,MAAM,MAAM;AAAA,UACZ;AAAA,UACA,OAAO,MAAM;AAAA,UACb,aAAa,MAAM;AAAA,UACnB,UAAU,UAAU;AAAA,QACtB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAEA,QAAM,UAAU,MAAM,KAAK,WAAW,OAAO,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AACtF,QAAM,aAAa,KAAK,IAAI,IAAI;AAEhC,SAAO;AAAA,IACL,mBAAmB,6BAA6B,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC;AAAA,IAC1E;AAAA,IACA,OAAO;AAAA,MACL,cAAc,QAAQ;AAAA,MACtB;AAAA,MACA,eAAe,YAAY,EAAE;AAAA,IAC/B;AAAA,EACF;AACF;AAKO,SAAS,gBAAgB,QAAgC;AAC9D,SAAO,OAAO,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI;AACzC;","names":[]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@precisa-saude/fhir-ocr-utils",
-  "version": "0.21.0",
+  "version": "0.21.1",
   "description": "Utilitários de ancoragem OCR para extração de biomarcadores de PDFs de resultados laboratoriais",
   "keywords": [
     "fhir",
@@ -41,7 +41,7 @@
     "dist"
   ],
   "dependencies": {
-    "@precisa-saude/fhir": "^0.21.0"
+    "@precisa-saude/fhir": "^0.21.1"
   },
   "devDependencies": {
     "tsup": "^8.3.5",