@precisa-saude/fhir-ocr-utils 0.16.2 → 0.16.4

package/README.md

@@ -47,7 +47,7 @@ console.log(result.filteredReference);
 
 Este pacote implementa um padrão de **ancoragem antes do LLM** para prevenir alucinações na extração de dados laboratoriais:
 
-1. **Ancoragem (este pacote):** Escaneia o texto OCR bruto procurando nomes de biomarcadores conhecidos usando correspondência exata de strings contra as 180+ definições de `@precisa-saude/fhir`.
+1. **Ancoragem (este pacote):** Escaneia o texto OCR bruto procurando nomes de biomarcadores conhecidos contra as 180+ definições de `@precisa-saude/fhir`.
 
 2. **Filtragem:** Gera uma referência LLM filtrada (`filteredReference`) contendo apenas os biomarcadores que foram realmente encontrados no texto. O LLM só pode extrair valores para biomarcadores presentes nesta lista.
 
@@ -62,16 +62,58 @@ PDF → OCR → findBiomarkersInText() → filteredReference → LLM → valores
                         o LLM pode extrair
 ```
 
+### Regras de correspondência
+
+Uma âncora falsa é cara: ela envia ao LLM uma referência de biomarcadores que
+não estão no documento, convidando o modelo a preencher valores inexistentes.
+Por isso a correspondência é conservadora:
+
+- **Fronteira de token** — o nome precisa aparecer inteiro. `Color` não casa
+  dentro de "Colorectal", nem `Cholesterol` dentro de "Hypercholesterolemia".
+  Plurais impressos pelos laboratórios ("Proteínas", "Cetonas") continuam
+  ancorando no nome singular do catálogo.
+- **Nome mais longo vence** — em sobreposição, o casamento mais específico
+  prevalece: "HDL Cholesterol" ancora `HDL`, não `Cholesterol`; "Blood Glucose"
+  ancora `Glucose`, não `Blood_Urine`.
+- **Nomes de uma linha só** — nos layouts em coluna dos laudos, linhas
+  consecutivas são biomarcadores distintos, então um nome composto não
+  atravessa quebra de linha ("Colesterol\nHDL" não vira "Colesterol HDL").
+- **Co-ocorrência de valor para nomes genéricos** — nomes que também são
+  palavras comuns (`Color`, `Protein`, `Blood`, `Volume`, `Peso`) só ancoram se
+  a linha carregar um valor: um número, uma unidade conhecida, ou um termo
+  qualitativo esperado ("Negativo", "Ausente", "Amarelo Citrino"). "Specimen
+  type: Blood" não ancora; "Sangue Oculto: Negativo" ancora.
+- **Contexto genético é descartado** — símbolos de gene colidem com nomes de
+  biomarcador (o gene `APOB` vs. a lipoproteína `ApoB`). Linhas com acesso
+  RefSeq (`NM_000384.2`), notação HGVS (`p.Trp448*`, `c.1234A>G`), `rs` do dbSNP
+  ou vocabulário de laudo genético não ancoram. É o contexto que decide, não uma
+  lista de símbolos proibidos — bani-los quebraria laudos lipídicos reais.
+
 ## API
 
 ### `findBiomarkersInText(ocrText: string): AnchorResult`
 
-Escaneia texto OCR e retorna todos os biomarcadores encontrados.
+Escaneia texto OCR e retorna todos os biomarcadores encontrados — um casamento
+por código, o de maior confiança.
 
-- `result.matches` — Lista de biomarcadores encontrados com código, LOINC, nome e posição
+- `result.matches` — Lista de biomarcadores encontrados com código, LOINC, nome, confiança e posição
 - `result.filteredReference` — Referência formatada para enviar ao LLM
 - `result.stats` — Estatísticas de execução (total de padrões, encontrados, tempo)
 
+`position` é o índice no texto normalizado (sem acentos, minúsculas, espaços
+horizontais colapsados), não no texto OCR original.
+
+#### Confiança
+
+`match.confidence` reflete a qualidade do casamento — nome completo ao lado de
+um valor não é a mesma coisa que uma menção solta em prosa:
+
+| Valor | Constante                   | Significado                                                              |
+| ----- | --------------------------- | ------------------------------------------------------------------------ |
+| `1.0` | `CONFIDENCE_VALUE_ADJACENT` | Nome específico com valor na mesma linha (`Glicose: 95 mg/dL`)           |
+| `0.7` | `CONFIDENCE_NAME_ONLY`      | Nome específico sem valor por perto (cabeçalho de seção, menção solta)   |
+| `0.4` | `CONFIDENCE_AMBIGUOUS`      | Nome genérico que só ancorou por causa do valor ao lado (`Cor: Amarelo`) |
+
 ### `getMatchedCodes(result: AnchorResult): string[]`
 
 Extrai a lista de códigos de biomarcadores de um resultado de ancoragem.

package/dist/cli.js

@@ -10,10 +10,15 @@ import { getInput, outputJson, outputText } from "@precisa-saude/fhir/cli-utils"
 // src/anchor.ts
 import {
   generateFilteredLLMReference,
-  getAllSearchPatterns
+  getAllSearchPatterns,
+  UNIT_TO_UCUM
 } from "@precisa-saude/fhir";
+var CONFIDENCE_VALUE_ADJACENT = 1;
+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+/g, " ");
+  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/[^\S\n]+/g, " ");
 }
 var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "hdl",
@@ -50,73 +55,266 @@ var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "mlg",
   "tav"
 ]);
+var CONTEXT_REQUIRED_NAMES = /* @__PURE__ */ new Set([
+  "bacteria",
+  // Bacteria_Urine — tem unidade, escapa da regra automática
+  "bacterias",
+  // Bacteria_Urine
+  "lead",
+  // Lead — verbo/substantivo comuníssimo em inglês
+  "peso",
+  // TotalMass
+  "saturation",
+  // TransferrinSaturation — "oxygen saturation", "saturation index"
+  "tap",
+  // ProthrombinTime — "tap" em inglês
+  "volume",
+  // VATVolume
+  "weight"
+  // TotalMass
+]);
+var QUALITATIVE_VALUE_TERMS = /* @__PURE__ */ new Set([
+  "absent",
+  "alguns",
+  "amarela",
+  "amarelo",
+  "anormal",
+  "ausencia",
+  "ausente",
+  "ausentes",
+  "citrino",
+  "claro",
+  "clear",
+  "cloudy",
+  "colorless",
+  "detectado",
+  "detected",
+  "escuro",
+  "incolor",
+  "indetectavel",
+  "limpido",
+  "moderada",
+  "moderado",
+  "negativa",
+  "negative",
+  "negativo",
+  "normais",
+  "normal",
+  "numerosos",
+  "ocasional",
+  "positiva",
+  "positive",
+  "positivo",
+  "present",
+  "presente",
+  "presentes",
+  "raras",
+  "raro",
+  "raros",
+  "reagente",
+  "trace",
+  "traces",
+  "tracos",
+  "turvo",
+  "undetectable",
+  "yellow"
+]);
+var GENETIC_CONTEXT_PATTERNS = [
+  /\b[nx][mrpc]_\d{6,}/,
+  // RefSeq: NM_000384.2, NP_, NR_, XM_
+  /\bens[gtp]\d{6,}/,
+  // Ensembl: ENSG00000084674
+  /\bp\.[a-z]{3}\d/,
+  // HGVS proteína: p.Trp448*
+  /\bc\.\d+[acgt]?[>_+-]/,
+  // HGVS codificante: c.1234A>G, c.76_78del
+  /\brs\d{4,}\b/,
+  // dbSNP
+  /\bgenes?\b/,
+  /\bvariante?s?\b/,
+  /\bexons?\b/,
+  /\bzygosity\b/,
+  /\bzigosidade\b/,
+  /\balleles?\b/,
+  /\balelos?\b/,
+  /\bmutations?\b/,
+  /\bmutac(ao|oes)\b/,
+  /\bpathogenic/,
+  /\bpatogenic/,
+  /\bheterozyg/,
+  /\bhomozyg/,
+  /\bheterozigot/,
+  /\bhomozigot/,
+  /\bsequence change\b/
+];
+var DIGIT_PATTERN = /\d/;
+var cachedUnitTokens = null;
+function getUnitTokens() {
+  if (!cachedUnitTokens) {
+    cachedUnitTokens = new Set(
+      Object.keys(UNIT_TO_UCUM).map((unit) => normalize(unit).trim()).filter(Boolean)
+    );
+  }
+  return cachedUnitTokens;
+}
 var cachedPatterns = null;
-var cachedNormalized = null;
+var cachedNamePatterns = null;
 function getPatterns() {
   if (!cachedPatterns) {
     cachedPatterns = getAllSearchPatterns();
   }
   return cachedPatterns;
 }
-function getNormalizedPatterns() {
-  if (!cachedNormalized) {
-    const patterns = getPatterns();
+function escapeRegExp(text) {
+  return text.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+function buildNamePattern(normalizedName) {
+  const body = normalizedName.split(" ").map(escapeRegExp).join("[^\\S\\n]+");
+  const plural = /\p{L}$/u.test(normalizedName) ? "s?" : "";
+  return new RegExp(`(?<![\\p{L}\\p{N}])${body}${plural}(?![\\p{L}\\p{N}])`, "gu");
+}
+function isQualitativeUrine(pattern) {
+  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];
+  return categories.includes("urina") && !pattern.unit;
+}
+function isAmbiguousName(normalizedName, pattern) {
+  if (normalizedName.includes(" ")) {
+    return false;
+  }
+  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);
+}
+function getNamePatterns() {
+  if (!cachedNamePatterns) {
     const map = /* @__PURE__ */ new Map();
-    for (const pattern of patterns) {
+    for (const pattern of getPatterns()) {
       for (const name of pattern.names) {
-        const normalized = normalize(name);
-        const existing = map.get(normalized) || [];
-        existing.push({
+        const normalized = normalize(name).trim();
+        if (!normalized) {
+          continue;
+        }
+        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {
+          continue;
+        }
+        let slot = map.get(normalized);
+        if (!slot) {
+          slot = { entries: [], regex: null };
+          map.set(normalized, slot);
+        }
+        slot.entries.push({
+          ambiguous: isAmbiguousName(normalized, pattern),
           code: pattern.code,
           ...pattern.loinc && { loinc: pattern.loinc },
           original: name
         });
-        map.set(normalized, existing);
       }
     }
-    cachedNormalized = map;
+    cachedNamePatterns = map;
   }
-  return cachedNormalized;
+  return cachedNamePatterns;
+}
+function getLineBounds(text, position) {
+  const start = text.lastIndexOf("\n", position) + 1;
+  const nextBreak = text.indexOf("\n", position);
+  return { end: nextBreak === -1 ? text.length : nextBreak, start };
+}
+function hasGeneticContext(line) {
+  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));
+}
+function hasValueEvidence(line) {
+  if (DIGIT_PATTERN.test(line)) {
+    return true;
+  }
+  const unitTokens = getUnitTokens();
+  for (const token of line.split(/[^\p{L}\p{N}%/]+/u)) {
+    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {
+      return true;
+    }
+  }
+  return false;
+}
+function collectCandidates(normalizedText) {
+  const candidates = [];
+  for (const [name, slot] of getNamePatterns()) {
+    if (!normalizedText.includes(name)) {
+      continue;
+    }
+    const { entries } = slot;
+    const regex = slot.regex ??= buildNamePattern(name);
+    regex.lastIndex = 0;
+    let occurrences = 0;
+    let match = regex.exec(normalizedText);
+    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {
+      candidates.push({ end: match.index + match[0].length, entries, start: match.index });
+      occurrences += 1;
+      match = regex.exec(normalizedText);
+    }
+  }
+  return candidates;
+}
+function resolveOverlaps(candidates) {
+  const sorted = [...candidates].sort(
+    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start
+  );
+  const accepted = [];
+  for (const candidate of sorted) {
+    const length = candidate.end - candidate.start;
+    const swallowed = accepted.some(
+      (other) => other.start <= candidate.start && candidate.end <= other.end && other.end - other.start > length
+    );
+    if (!swallowed) {
+      accepted.push(candidate);
+    }
+  }
+  return accepted;
 }
 function findBiomarkersInText(ocrText) {
   const startTime = Date.now();
   const normalizedText = normalize(ocrText);
-  const matchedCodes = /* @__PURE__ */ new Set();
-  const matches = [];
-  const normalizedPatterns = getNormalizedPatterns();
-  for (const [normalizedName, entries] of normalizedPatterns) {
-    if (normalizedName.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalizedName)) {
+  const bestByCode = /* @__PURE__ */ new Map();
+  const geneticLines = /* @__PURE__ */ new Map();
+  const valueLines = /* @__PURE__ */ new Map();
+  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {
+    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);
+    let genetic = geneticLines.get(lineStart);
+    if (genetic === void 0) {
+      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));
+      geneticLines.set(lineStart, genetic);
+    }
+    if (genetic) {
       continue;
     }
-    let position = -1;
-    if (normalizedName.length <= 4) {
-      const regex = new RegExp(`\\b${normalizedName}\\b`);
-      const match = regex.exec(normalizedText);
-      if (match) {
-        position = match.index;
-      }
-    } else {
-      position = normalizedText.indexOf(normalizedName);
+    let hasValue = valueLines.get(lineStart);
+    if (hasValue === void 0) {
+      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));
+      valueLines.set(lineStart, hasValue);
     }
-    if (position !== -1) {
-      for (const entry of entries) {
-        if (!matchedCodes.has(entry.code)) {
-          matchedCodes.add(entry.code);
-          matches.push({
-            code: entry.code,
-            confidence: 1,
-            loinc: entry.loinc,
-            matchedName: entry.original,
-            position
-          });
-        }
+    for (const entry of candidate.entries) {
+      if (entry.ambiguous && !hasValue) {
+        continue;
+      }
+      let confidence = CONFIDENCE_NAME_ONLY;
+      if (entry.ambiguous) {
+        confidence = CONFIDENCE_AMBIGUOUS;
+      } else if (hasValue) {
+        confidence = CONFIDENCE_VALUE_ADJACENT;
+      }
+      const existing = bestByCode.get(entry.code);
+      const better = !existing || confidence > existing.confidence || confidence === existing.confidence && candidate.start < existing.position;
+      if (better) {
+        bestByCode.set(entry.code, {
+          code: entry.code,
+          confidence,
+          loinc: entry.loinc,
+          matchedName: entry.original,
+          position: candidate.start
+        });
       }
     }
   }
+  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);
   const scanTimeMs = Date.now() - startTime;
-  const matchedCodesArray = Array.from(matchedCodes);
   return {
-    filteredReference: generateFilteredLLMReference(matchedCodesArray),
+    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),
     matches,
     stats: {
       matchedCount: matches.length,
@@ -205,7 +403,7 @@ async function main() {
     strict: false
   });
   if (values.version) {
-    process.stdout.write(`${"0.16.2"}
+    process.stdout.write(`${"0.16.4"}
 `);
     return;
   }

package/dist/index.cjs

@@ -2,9 +2,14 @@
 
 
 
+
 var _fhir = require('@precisa-saude/fhir');
+var CONFIDENCE_VALUE_ADJACENT = 1;
+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+/g, " ");
+  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/[^\S\n]+/g, " ");
 }
 var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "hdl",
@@ -41,73 +46,266 @@ var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "mlg",
   "tav"
 ]);
+var CONTEXT_REQUIRED_NAMES = /* @__PURE__ */ new Set([
+  "bacteria",
+  // Bacteria_Urine — tem unidade, escapa da regra automática
+  "bacterias",
+  // Bacteria_Urine
+  "lead",
+  // Lead — verbo/substantivo comuníssimo em inglês
+  "peso",
+  // TotalMass
+  "saturation",
+  // TransferrinSaturation — "oxygen saturation", "saturation index"
+  "tap",
+  // ProthrombinTime — "tap" em inglês
+  "volume",
+  // VATVolume
+  "weight"
+  // TotalMass
+]);
+var QUALITATIVE_VALUE_TERMS = /* @__PURE__ */ new Set([
+  "absent",
+  "alguns",
+  "amarela",
+  "amarelo",
+  "anormal",
+  "ausencia",
+  "ausente",
+  "ausentes",
+  "citrino",
+  "claro",
+  "clear",
+  "cloudy",
+  "colorless",
+  "detectado",
+  "detected",
+  "escuro",
+  "incolor",
+  "indetectavel",
+  "limpido",
+  "moderada",
+  "moderado",
+  "negativa",
+  "negative",
+  "negativo",
+  "normais",
+  "normal",
+  "numerosos",
+  "ocasional",
+  "positiva",
+  "positive",
+  "positivo",
+  "present",
+  "presente",
+  "presentes",
+  "raras",
+  "raro",
+  "raros",
+  "reagente",
+  "trace",
+  "traces",
+  "tracos",
+  "turvo",
+  "undetectable",
+  "yellow"
+]);
+var GENETIC_CONTEXT_PATTERNS = [
+  /\b[nx][mrpc]_\d{6,}/,
+  // RefSeq: NM_000384.2, NP_, NR_, XM_
+  /\bens[gtp]\d{6,}/,
+  // Ensembl: ENSG00000084674
+  /\bp\.[a-z]{3}\d/,
+  // HGVS proteína: p.Trp448*
+  /\bc\.\d+[acgt]?[>_+-]/,
+  // HGVS codificante: c.1234A>G, c.76_78del
+  /\brs\d{4,}\b/,
+  // dbSNP
+  /\bgenes?\b/,
+  /\bvariante?s?\b/,
+  /\bexons?\b/,
+  /\bzygosity\b/,
+  /\bzigosidade\b/,
+  /\balleles?\b/,
+  /\balelos?\b/,
+  /\bmutations?\b/,
+  /\bmutac(ao|oes)\b/,
+  /\bpathogenic/,
+  /\bpatogenic/,
+  /\bheterozyg/,
+  /\bhomozyg/,
+  /\bheterozigot/,
+  /\bhomozigot/,
+  /\bsequence change\b/
+];
+var DIGIT_PATTERN = /\d/;
+var cachedUnitTokens = null;
+function getUnitTokens() {
+  if (!cachedUnitTokens) {
+    cachedUnitTokens = new Set(
+      Object.keys(_fhir.UNIT_TO_UCUM).map((unit) => normalize(unit).trim()).filter(Boolean)
+    );
+  }
+  return cachedUnitTokens;
+}
 var cachedPatterns = null;
-var cachedNormalized = null;
+var cachedNamePatterns = null;
 function getPatterns() {
   if (!cachedPatterns) {
     cachedPatterns = _fhir.getAllSearchPatterns.call(void 0, );
   }
   return cachedPatterns;
 }
-function getNormalizedPatterns() {
-  if (!cachedNormalized) {
-    const patterns = getPatterns();
+function escapeRegExp(text) {
+  return text.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+function buildNamePattern(normalizedName) {
+  const body = normalizedName.split(" ").map(escapeRegExp).join("[^\\S\\n]+");
+  const plural = /\p{L}$/u.test(normalizedName) ? "s?" : "";
+  return new RegExp(`(?<![\\p{L}\\p{N}])${body}${plural}(?![\\p{L}\\p{N}])`, "gu");
+}
+function isQualitativeUrine(pattern) {
+  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];
+  return categories.includes("urina") && !pattern.unit;
+}
+function isAmbiguousName(normalizedName, pattern) {
+  if (normalizedName.includes(" ")) {
+    return false;
+  }
+  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);
+}
+function getNamePatterns() {
+  if (!cachedNamePatterns) {
     const map = /* @__PURE__ */ new Map();
-    for (const pattern of patterns) {
+    for (const pattern of getPatterns()) {
       for (const name of pattern.names) {
-        const normalized = normalize(name);
-        const existing = map.get(normalized) || [];
-        existing.push({
+        const normalized = normalize(name).trim();
+        if (!normalized) {
+          continue;
+        }
+        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {
+          continue;
+        }
+        let slot = map.get(normalized);
+        if (!slot) {
+          slot = { entries: [], regex: null };
+          map.set(normalized, slot);
+        }
+        slot.entries.push({
+          ambiguous: isAmbiguousName(normalized, pattern),
           code: pattern.code,
           ...pattern.loinc && { loinc: pattern.loinc },
           original: name
         });
-        map.set(normalized, existing);
       }
     }
-    cachedNormalized = map;
+    cachedNamePatterns = map;
+  }
+  return cachedNamePatterns;
+}
+function getLineBounds(text, position) {
+  const start = text.lastIndexOf("\n", position) + 1;
+  const nextBreak = text.indexOf("\n", position);
+  return { end: nextBreak === -1 ? text.length : nextBreak, start };
+}
+function hasGeneticContext(line) {
+  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));
+}
+function hasValueEvidence(line) {
+  if (DIGIT_PATTERN.test(line)) {
+    return true;
+  }
+  const unitTokens = getUnitTokens();
+  for (const token of line.split(/[^\p{L}\p{N}%/]+/u)) {
+    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {
+      return true;
+    }
+  }
+  return false;
+}
+function collectCandidates(normalizedText) {
+  const candidates = [];
+  for (const [name, slot] of getNamePatterns()) {
+    if (!normalizedText.includes(name)) {
+      continue;
+    }
+    const { entries } = slot;
+    const regex = slot.regex ??= buildNamePattern(name);
+    regex.lastIndex = 0;
+    let occurrences = 0;
+    let match = regex.exec(normalizedText);
+    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {
+      candidates.push({ end: match.index + match[0].length, entries, start: match.index });
+      occurrences += 1;
+      match = regex.exec(normalizedText);
+    }
+  }
+  return candidates;
+}
+function resolveOverlaps(candidates) {
+  const sorted = [...candidates].sort(
+    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start
+  );
+  const accepted = [];
+  for (const candidate of sorted) {
+    const length = candidate.end - candidate.start;
+    const swallowed = accepted.some(
+      (other) => other.start <= candidate.start && candidate.end <= other.end && other.end - other.start > length
+    );
+    if (!swallowed) {
+      accepted.push(candidate);
+    }
   }
-  return cachedNormalized;
+  return accepted;
 }
 function findBiomarkersInText(ocrText) {
   const startTime = Date.now();
   const normalizedText = normalize(ocrText);
-  const matchedCodes = /* @__PURE__ */ new Set();
-  const matches = [];
-  const normalizedPatterns = getNormalizedPatterns();
-  for (const [normalizedName, entries] of normalizedPatterns) {
-    if (normalizedName.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalizedName)) {
+  const bestByCode = /* @__PURE__ */ new Map();
+  const geneticLines = /* @__PURE__ */ new Map();
+  const valueLines = /* @__PURE__ */ new Map();
+  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {
+    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);
+    let genetic = geneticLines.get(lineStart);
+    if (genetic === void 0) {
+      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));
+      geneticLines.set(lineStart, genetic);
+    }
+    if (genetic) {
       continue;
     }
-    let position = -1;
-    if (normalizedName.length <= 4) {
-      const regex = new RegExp(`\\b${normalizedName}\\b`);
-      const match = regex.exec(normalizedText);
-      if (match) {
-        position = match.index;
-      }
-    } else {
-      position = normalizedText.indexOf(normalizedName);
+    let hasValue = valueLines.get(lineStart);
+    if (hasValue === void 0) {
+      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));
+      valueLines.set(lineStart, hasValue);
     }
-    if (position !== -1) {
-      for (const entry of entries) {
-        if (!matchedCodes.has(entry.code)) {
-          matchedCodes.add(entry.code);
-          matches.push({
-            code: entry.code,
-            confidence: 1,
-            loinc: entry.loinc,
-            matchedName: entry.original,
-            position
-          });
-        }
+    for (const entry of candidate.entries) {
+      if (entry.ambiguous && !hasValue) {
+        continue;
+      }
+      let confidence = CONFIDENCE_NAME_ONLY;
+      if (entry.ambiguous) {
+        confidence = CONFIDENCE_AMBIGUOUS;
+      } else if (hasValue) {
+        confidence = CONFIDENCE_VALUE_ADJACENT;
+      }
+      const existing = bestByCode.get(entry.code);
+      const better = !existing || confidence > existing.confidence || confidence === existing.confidence && candidate.start < existing.position;
+      if (better) {
+        bestByCode.set(entry.code, {
+          code: entry.code,
+          confidence,
+          loinc: entry.loinc,
+          matchedName: entry.original,
+          position: candidate.start
+        });
       }
     }
   }
+  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);
   const scanTimeMs = Date.now() - startTime;
-  const matchedCodesArray = Array.from(matchedCodes);
   return {
-    filteredReference: _fhir.generateFilteredLLMReference.call(void 0, matchedCodesArray),
+    filteredReference: _fhir.generateFilteredLLMReference.call(void 0, matches.map((m) => m.code)),
     matches,
     stats: {
       matchedCount: matches.length,
@@ -122,5 +320,8 @@ function getMatchedCodes(result) {
 
 
 
-exports.findBiomarkersInText = findBiomarkersInText; exports.getMatchedCodes = getMatchedCodes;
+
+
+
+exports.CONFIDENCE_AMBIGUOUS = CONFIDENCE_AMBIGUOUS; exports.CONFIDENCE_NAME_ONLY = CONFIDENCE_NAME_ONLY; exports.CONFIDENCE_VALUE_ADJACENT = CONFIDENCE_VALUE_ADJACENT; exports.findBiomarkersInText = findBiomarkersInText; exports.getMatchedCodes = getMatchedCodes;
 //# sourceMappingURL=index.cjs.map

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;ACQA;AAEE;AACA;AAAA,2CACK;AA0BP,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,MAAA,EAAQ,GAAG,CAAA;AACxB;AAEA,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACt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eAAA,EAAkD,IAAA;AACtD,IAAI,iBAAA,EAAuD,IAAA;AAE3D,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,qBAAA,CAAA,EAAqD;AAC5D,EAAA,GAAA,CAAI,CAAC,gBAAA,EAAkB;AACrB,IAAA,MAAM,SAAA,EAAW,WAAA,CAAY,CAAA;AAC7B,IAAA,MAAM,IAAA,kBAAM,IAAI,GAAA,CAA4B,CAAA;AAC5C,IAAA,IAAA,CAAA,MAAW,QAAA,GAAW,QAAA,EAAU;AAC9B,MAAA,IAAA,CAAA,MAAW,KAAA,GAAQ,OAAA,CAAQ,KAAA,EAAO;AAChC,QAAA,MAAM,WAAA,EAAa,SAAA,CAAU,IAAI,CAAA;AACjC,QAAA,MAAM,SAAA,EAAW,GAAA,CAAI,GAAA,CAAI,UAAU,EAAA,GAAK,CAAC,CAAA;AACzC,QAAA,QAAA,CAAS,IAAA,CAAK;AAAA,UACZ,IAAA,EAAM,OAAA,CAAQ,IAAA;AAAA,UACd,GAAI,OAAA,CAAQ,MAAA,GAAS,EAAE,KAAA,EAAO,OAAA,CAAQ,MAAM,CAAA;AAAA,UAC5C,QAAA,EAAU;AAAA,QACZ,CAAC,CAAA;AACD,QAAA,GAAA,CAAI,GAAA,CAAI,UAAA,EAAY,QAAQ,CAAA;AAAA,MAC9B;AAAA,IACF;AACA,IAAA,iBAAA,EAAmB,GAAA;AAAA,EACrB;AACA,EAAA,OAAO,gBAAA;AACT;AAOO,SAAS,oBAAA,CAAqB,OAAA,EAA+B;AAClE,EAAA,MAAM,UAAA,EAAY,IAAA,CAAK,GAAA,CAAI,CAAA;AAC3B,EAAA,MAAM,eAAA,EAAiB,SAAA,CAAU,OAAO,CAAA;AACxC,EAAA,MAAM,aAAA,kBAAe,IAAI,GAAA,CAAY,CAAA;AACrC,EAAA,MAAM,QAAA,EAAyB,CAAC,CAAA;AAChC,EAAA,MAAM,mBAAA,EAAqB,qBAAA,CAAsB,CAAA;AAEjD,EAAA,IAAA,CAAA,MAAW,CAAC,cAAA,EAAgB,OAAO,EAAA,GAAK,kBAAA,EAAoB;AAC1D,IAAA,GAAA,CAAI,cAAA,CAAe,OAAA,EAAS,EAAA,GAAK,CAAC,uBAAA,CAAwB,GAAA,CAAI,cAAc,CAAA,EAAG;AAC7E,MAAA,QAAA;AAAA,IACF;AAEA,IAAA,IAAI,SAAA,EAAW,CAAA,CAAA;AACf,IAAA,GAAA,CAAI,cAAA,CAAe,OAAA,GAAU,CAAA,EAAG;AAC9B,MAAA,MAAM,MAAA,EAAQ,IAAI,MAAA,CAAO,CAAA,GAAA,EAAM,cAAc,CAAA,GAAA,CAAK,CAAA;AAClD,MAAA,MAAM,MAAA,EAAQ,KAAA,CAAM,IAAA,CAAK,cAAc,CAAA;AACvC,MAAA,GAAA,CAAI,KAAA,EAAO;AACT,QAAA,SAAA,EAAW,KAAA,CAAM,KAAA;AAAA,MACnB;AAAA,IACF,EAAA,KAAO;AACL,MAAA,SAAA,EAAW,cAAA,CAAe,OAAA,CAAQ,cAAc,CAAA;AAAA,IAClD;AAEA,IAAA,GAAA,CAAI,SAAA,IAAa,CAAA,CAAA,EAAI;AACnB,MAAA,IAAA,CAAA,MAAW,MAAA,GAAS,OAAA,EAAS;AAC3B,QAAA,GAAA,CAAI,CAAC,YAAA,CAAa,GAAA,CAAI,KAAA,CAAM,IAAI,CAAA,EAAG;AACjC,UAAA,YAAA,CAAa,GAAA,CAAI,KAAA,CAAM,IAAI,CAAA;AAC3B,UAAA,OAAA,CAAQ,IAAA,CAAK;AAAA,YACX,IAAA,EAAM,KAAA,CAAM,IAAA;AAAA,YACZ,UAAA,EAAY,CAAA;AAAA,YACZ,KAAA,EAAO,KAAA,CAAM,KAAA;AAAA,YACb,WAAA,EAAa,KAAA,CAAM,QAAA;AAAA,YACnB;AAAA,UACF,CAAC,CAAA;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,EAAA,MAAM,WAAA,EAAa,IAAA,CAAK,GAAA,CAAI,EAAA,EAAI,SAAA;AAChC,EAAA,MAAM,kBAAA,EAAoB,KAAA,CAAM,IAAA,CAAK,YAAY,CAAA;AAEjD,EAAA,OAAO;AAAA,IACL,iBAAA,EAAmB,gDAAA,iBAA8C,CAAA;AAAA,IACjE,OAAA;AAAA,IACA,KAAA,EAAO;AAAA,MACL,YAAA,EAAc,OAAA,CAAQ,MAAA;AAAA,MACtB,UAAA;AAAA,MACA,aAAA,EAAe,WAAA,CAAY,CAAA,CAAE;AAAA,IAC/B;AAAA,EACF,CAAA;AACF;AAKO,SAAS,eAAA,CAAgB,MAAA,EAAgC;AAC9D,EAAA,OAAO,MAAA,CAAO,OAAA,CAAQ,GAAA,CAAI,CAAC,CAAA,EAAA,GAAM,CAAA,CAAE,IAAI,CAAA;AACzC;ADzDA;AACE;AACA;AACF,+FAAC","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\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\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 * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Normalizes whitespace\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/\\s+/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\ntype PatternEntry = { original: string; code: string; loinc?: string };\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNormalized: Map<string, PatternEntry[]> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction getNormalizedPatterns(): Map<string, PatternEntry[]> {\n  if (!cachedNormalized) {\n    const patterns = getPatterns();\n    const map = new Map<string, PatternEntry[]>();\n    for (const pattern of patterns) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name);\n        const existing = map.get(normalized) || [];\n        existing.push({\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n        map.set(normalized, existing);\n      }\n    }\n    cachedNormalized = map;\n  }\n  return cachedNormalized;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n * Uses exact string matching on normalized text.\n * Returns unique matches (same biomarker won't be matched twice).\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const matchedCodes = new Set<string>();\n  const matches: AnchorMatch[] = [];\n  const normalizedPatterns = getNormalizedPatterns();\n\n  for (const [normalizedName, entries] of normalizedPatterns) {\n    if (normalizedName.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalizedName)) {\n      continue;\n    }\n\n    let position = -1;\n    if (normalizedName.length <= 4) {\n      const regex = new RegExp(`\\\\b${normalizedName}\\\\b`);\n      const match = regex.exec(normalizedText);\n      if (match) {\n        position = match.index;\n      }\n    } else {\n      position = normalizedText.indexOf(normalizedName);\n    }\n\n    if (position !== -1) {\n      for (const entry of entries) {\n        if (!matchedCodes.has(entry.code)) {\n          matchedCodes.add(entry.code);\n          matches.push({\n            code: entry.code,\n            confidence: 1.0,\n            loinc: entry.loinc,\n            matchedName: entry.original,\n            position,\n          });\n        }\n      }\n    }\n  }\n\n  const scanTimeMs = Date.now() - startTime;\n  const matchedCodesArray = Array.from(matchedCodes);\n\n  return {\n    filteredReference: generateFilteredLLMReference(matchedCodesArray),\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;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;AAAA;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;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;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;AAEuC,IAAA;AACH,MAAA;AAChC,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;AD/K+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]);\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\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\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    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\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.d.cts

@@ -4,6 +4,11 @@
  * Scans OCR text for biomarker names BEFORE sending to LLM.
  * This prevents hallucination by constraining what biomarkers
  * the LLM is allowed to extract.
+ *
+ * Matching is deliberately conservative: a name only anchors when it appears
+ * as a whole token, is not swallowed by a longer biomarker name, is not inside
+ * a genetic report line, and — for generic single-word names — sits on a line
+ * that actually carries a value.
  */
 interface AnchorMatch {
     code: string;
@@ -21,10 +26,28 @@ interface AnchorResult {
         scanTimeMs: number;
     };
 }
+/**
+ * Confidence assigned to a specific biomarker name found on a line that also
+ * carries a value (a number, a unit, or an expected qualitative term).
+ */
+declare const CONFIDENCE_VALUE_ADJACENT = 1;
+/**
+ * Confidence assigned to a specific biomarker name with no value evidence
+ * nearby — a section heading, or a mention in prose.
+ */
+declare const CONFIDENCE_NAME_ONLY = 0.7;
+/**
+ * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,
+ * `Blood`, …) that only anchored because a value was found next to it.
+ */
+declare const CONFIDENCE_AMBIGUOUS = 0.4;
 /**
  * Find all biomarker names present in OCR text.
- * Uses exact string matching on normalized text.
- * Returns unique matches (same biomarker won't be matched twice).
+ *
+ * Matching is whole-token, longest-match-wins, and context-aware: matches
+ * inside genetic report lines are discarded, and generic names only anchor
+ * when a value sits on the same line. Returns one match per biomarker code —
+ * the highest-confidence occurrence.
  */
 declare function findBiomarkersInText(ocrText: string): AnchorResult;
 /**
@@ -32,4 +55,4 @@ declare function findBiomarkersInText(ocrText: string): AnchorResult;
  */
 declare function getMatchedCodes(result: AnchorResult): string[];
 
-export { type AnchorMatch, type AnchorResult, findBiomarkersInText, getMatchedCodes };
+export { type AnchorMatch, type AnchorResult, CONFIDENCE_AMBIGUOUS, CONFIDENCE_NAME_ONLY, CONFIDENCE_VALUE_ADJACENT, findBiomarkersInText, getMatchedCodes };

package/dist/index.d.ts

@@ -4,6 +4,11 @@
  * Scans OCR text for biomarker names BEFORE sending to LLM.
  * This prevents hallucination by constraining what biomarkers
  * the LLM is allowed to extract.
+ *
+ * Matching is deliberately conservative: a name only anchors when it appears
+ * as a whole token, is not swallowed by a longer biomarker name, is not inside
+ * a genetic report line, and — for generic single-word names — sits on a line
+ * that actually carries a value.
  */
 interface AnchorMatch {
     code: string;
@@ -21,10 +26,28 @@ interface AnchorResult {
         scanTimeMs: number;
     };
 }
+/**
+ * Confidence assigned to a specific biomarker name found on a line that also
+ * carries a value (a number, a unit, or an expected qualitative term).
+ */
+declare const CONFIDENCE_VALUE_ADJACENT = 1;
+/**
+ * Confidence assigned to a specific biomarker name with no value evidence
+ * nearby — a section heading, or a mention in prose.
+ */
+declare const CONFIDENCE_NAME_ONLY = 0.7;
+/**
+ * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,
+ * `Blood`, …) that only anchored because a value was found next to it.
+ */
+declare const CONFIDENCE_AMBIGUOUS = 0.4;
 /**
  * Find all biomarker names present in OCR text.
- * Uses exact string matching on normalized text.
- * Returns unique matches (same biomarker won't be matched twice).
+ *
+ * Matching is whole-token, longest-match-wins, and context-aware: matches
+ * inside genetic report lines are discarded, and generic names only anchor
+ * when a value sits on the same line. Returns one match per biomarker code —
+ * the highest-confidence occurrence.
  */
 declare function findBiomarkersInText(ocrText: string): AnchorResult;
 /**
@@ -32,4 +55,4 @@ declare function findBiomarkersInText(ocrText: string): AnchorResult;
  */
 declare function getMatchedCodes(result: AnchorResult): string[];
 
-export { type AnchorMatch, type AnchorResult, findBiomarkersInText, getMatchedCodes };
+export { type AnchorMatch, type AnchorResult, CONFIDENCE_AMBIGUOUS, CONFIDENCE_NAME_ONLY, CONFIDENCE_VALUE_ADJACENT, findBiomarkersInText, getMatchedCodes };

package/dist/index.js

@@ -1,10 +1,15 @@
 // src/anchor.ts
 import {
   generateFilteredLLMReference,
-  getAllSearchPatterns
+  getAllSearchPatterns,
+  UNIT_TO_UCUM
 } from "@precisa-saude/fhir";
+var CONFIDENCE_VALUE_ADJACENT = 1;
+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+/g, " ");
+  return text.normalize("NFD").replace(/[\u0300-\u036f]/g, "").toLowerCase().replace(/[^\S\n]+/g, " ");
 }
 var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "hdl",
@@ -41,73 +46,266 @@ var UNAMBIGUOUS_SHORT_NAMES = /* @__PURE__ */ new Set([
   "mlg",
   "tav"
 ]);
+var CONTEXT_REQUIRED_NAMES = /* @__PURE__ */ new Set([
+  "bacteria",
+  // Bacteria_Urine — tem unidade, escapa da regra automática
+  "bacterias",
+  // Bacteria_Urine
+  "lead",
+  // Lead — verbo/substantivo comuníssimo em inglês
+  "peso",
+  // TotalMass
+  "saturation",
+  // TransferrinSaturation — "oxygen saturation", "saturation index"
+  "tap",
+  // ProthrombinTime — "tap" em inglês
+  "volume",
+  // VATVolume
+  "weight"
+  // TotalMass
+]);
+var QUALITATIVE_VALUE_TERMS = /* @__PURE__ */ new Set([
+  "absent",
+  "alguns",
+  "amarela",
+  "amarelo",
+  "anormal",
+  "ausencia",
+  "ausente",
+  "ausentes",
+  "citrino",
+  "claro",
+  "clear",
+  "cloudy",
+  "colorless",
+  "detectado",
+  "detected",
+  "escuro",
+  "incolor",
+  "indetectavel",
+  "limpido",
+  "moderada",
+  "moderado",
+  "negativa",
+  "negative",
+  "negativo",
+  "normais",
+  "normal",
+  "numerosos",
+  "ocasional",
+  "positiva",
+  "positive",
+  "positivo",
+  "present",
+  "presente",
+  "presentes",
+  "raras",
+  "raro",
+  "raros",
+  "reagente",
+  "trace",
+  "traces",
+  "tracos",
+  "turvo",
+  "undetectable",
+  "yellow"
+]);
+var GENETIC_CONTEXT_PATTERNS = [
+  /\b[nx][mrpc]_\d{6,}/,
+  // RefSeq: NM_000384.2, NP_, NR_, XM_
+  /\bens[gtp]\d{6,}/,
+  // Ensembl: ENSG00000084674
+  /\bp\.[a-z]{3}\d/,
+  // HGVS proteína: p.Trp448*
+  /\bc\.\d+[acgt]?[>_+-]/,
+  // HGVS codificante: c.1234A>G, c.76_78del
+  /\brs\d{4,}\b/,
+  // dbSNP
+  /\bgenes?\b/,
+  /\bvariante?s?\b/,
+  /\bexons?\b/,
+  /\bzygosity\b/,
+  /\bzigosidade\b/,
+  /\balleles?\b/,
+  /\balelos?\b/,
+  /\bmutations?\b/,
+  /\bmutac(ao|oes)\b/,
+  /\bpathogenic/,
+  /\bpatogenic/,
+  /\bheterozyg/,
+  /\bhomozyg/,
+  /\bheterozigot/,
+  /\bhomozigot/,
+  /\bsequence change\b/
+];
+var DIGIT_PATTERN = /\d/;
+var cachedUnitTokens = null;
+function getUnitTokens() {
+  if (!cachedUnitTokens) {
+    cachedUnitTokens = new Set(
+      Object.keys(UNIT_TO_UCUM).map((unit) => normalize(unit).trim()).filter(Boolean)
+    );
+  }
+  return cachedUnitTokens;
+}
 var cachedPatterns = null;
-var cachedNormalized = null;
+var cachedNamePatterns = null;
 function getPatterns() {
   if (!cachedPatterns) {
     cachedPatterns = getAllSearchPatterns();
   }
   return cachedPatterns;
 }
-function getNormalizedPatterns() {
-  if (!cachedNormalized) {
-    const patterns = getPatterns();
+function escapeRegExp(text) {
+  return text.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+function buildNamePattern(normalizedName) {
+  const body = normalizedName.split(" ").map(escapeRegExp).join("[^\\S\\n]+");
+  const plural = /\p{L}$/u.test(normalizedName) ? "s?" : "";
+  return new RegExp(`(?<![\\p{L}\\p{N}])${body}${plural}(?![\\p{L}\\p{N}])`, "gu");
+}
+function isQualitativeUrine(pattern) {
+  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];
+  return categories.includes("urina") && !pattern.unit;
+}
+function isAmbiguousName(normalizedName, pattern) {
+  if (normalizedName.includes(" ")) {
+    return false;
+  }
+  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);
+}
+function getNamePatterns() {
+  if (!cachedNamePatterns) {
     const map = /* @__PURE__ */ new Map();
-    for (const pattern of patterns) {
+    for (const pattern of getPatterns()) {
       for (const name of pattern.names) {
-        const normalized = normalize(name);
-        const existing = map.get(normalized) || [];
-        existing.push({
+        const normalized = normalize(name).trim();
+        if (!normalized) {
+          continue;
+        }
+        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {
+          continue;
+        }
+        let slot = map.get(normalized);
+        if (!slot) {
+          slot = { entries: [], regex: null };
+          map.set(normalized, slot);
+        }
+        slot.entries.push({
+          ambiguous: isAmbiguousName(normalized, pattern),
           code: pattern.code,
           ...pattern.loinc && { loinc: pattern.loinc },
           original: name
         });
-        map.set(normalized, existing);
       }
     }
-    cachedNormalized = map;
+    cachedNamePatterns = map;
+  }
+  return cachedNamePatterns;
+}
+function getLineBounds(text, position) {
+  const start = text.lastIndexOf("\n", position) + 1;
+  const nextBreak = text.indexOf("\n", position);
+  return { end: nextBreak === -1 ? text.length : nextBreak, start };
+}
+function hasGeneticContext(line) {
+  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));
+}
+function hasValueEvidence(line) {
+  if (DIGIT_PATTERN.test(line)) {
+    return true;
+  }
+  const unitTokens = getUnitTokens();
+  for (const token of line.split(/[^\p{L}\p{N}%/]+/u)) {
+    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {
+      return true;
+    }
+  }
+  return false;
+}
+function collectCandidates(normalizedText) {
+  const candidates = [];
+  for (const [name, slot] of getNamePatterns()) {
+    if (!normalizedText.includes(name)) {
+      continue;
+    }
+    const { entries } = slot;
+    const regex = slot.regex ??= buildNamePattern(name);
+    regex.lastIndex = 0;
+    let occurrences = 0;
+    let match = regex.exec(normalizedText);
+    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {
+      candidates.push({ end: match.index + match[0].length, entries, start: match.index });
+      occurrences += 1;
+      match = regex.exec(normalizedText);
+    }
+  }
+  return candidates;
+}
+function resolveOverlaps(candidates) {
+  const sorted = [...candidates].sort(
+    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start
+  );
+  const accepted = [];
+  for (const candidate of sorted) {
+    const length = candidate.end - candidate.start;
+    const swallowed = accepted.some(
+      (other) => other.start <= candidate.start && candidate.end <= other.end && other.end - other.start > length
+    );
+    if (!swallowed) {
+      accepted.push(candidate);
+    }
   }
-  return cachedNormalized;
+  return accepted;
 }
 function findBiomarkersInText(ocrText) {
   const startTime = Date.now();
   const normalizedText = normalize(ocrText);
-  const matchedCodes = /* @__PURE__ */ new Set();
-  const matches = [];
-  const normalizedPatterns = getNormalizedPatterns();
-  for (const [normalizedName, entries] of normalizedPatterns) {
-    if (normalizedName.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalizedName)) {
+  const bestByCode = /* @__PURE__ */ new Map();
+  const geneticLines = /* @__PURE__ */ new Map();
+  const valueLines = /* @__PURE__ */ new Map();
+  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {
+    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);
+    let genetic = geneticLines.get(lineStart);
+    if (genetic === void 0) {
+      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));
+      geneticLines.set(lineStart, genetic);
+    }
+    if (genetic) {
       continue;
     }
-    let position = -1;
-    if (normalizedName.length <= 4) {
-      const regex = new RegExp(`\\b${normalizedName}\\b`);
-      const match = regex.exec(normalizedText);
-      if (match) {
-        position = match.index;
-      }
-    } else {
-      position = normalizedText.indexOf(normalizedName);
+    let hasValue = valueLines.get(lineStart);
+    if (hasValue === void 0) {
+      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));
+      valueLines.set(lineStart, hasValue);
     }
-    if (position !== -1) {
-      for (const entry of entries) {
-        if (!matchedCodes.has(entry.code)) {
-          matchedCodes.add(entry.code);
-          matches.push({
-            code: entry.code,
-            confidence: 1,
-            loinc: entry.loinc,
-            matchedName: entry.original,
-            position
-          });
-        }
+    for (const entry of candidate.entries) {
+      if (entry.ambiguous && !hasValue) {
+        continue;
+      }
+      let confidence = CONFIDENCE_NAME_ONLY;
+      if (entry.ambiguous) {
+        confidence = CONFIDENCE_AMBIGUOUS;
+      } else if (hasValue) {
+        confidence = CONFIDENCE_VALUE_ADJACENT;
+      }
+      const existing = bestByCode.get(entry.code);
+      const better = !existing || confidence > existing.confidence || confidence === existing.confidence && candidate.start < existing.position;
+      if (better) {
+        bestByCode.set(entry.code, {
+          code: entry.code,
+          confidence,
+          loinc: entry.loinc,
+          matchedName: entry.original,
+          position: candidate.start
+        });
       }
     }
   }
+  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);
   const scanTimeMs = Date.now() - startTime;
-  const matchedCodesArray = Array.from(matchedCodes);
   return {
-    filteredReference: generateFilteredLLMReference(matchedCodesArray),
+    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),
     matches,
     stats: {
       matchedCount: matches.length,
@@ -120,6 +318,9 @@ function getMatchedCodes(result) {
   return result.matches.map((m) => m.code);
 }
 export {
+  CONFIDENCE_AMBIGUOUS,
+  CONFIDENCE_NAME_ONLY,
+  CONFIDENCE_VALUE_ADJACENT,
   findBiomarkersInText,
   getMatchedCodes
 };

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\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\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 * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Normalizes whitespace\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/\\s+/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\ntype PatternEntry = { original: string; code: string; loinc?: string };\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNormalized: Map<string, PatternEntry[]> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction getNormalizedPatterns(): Map<string, PatternEntry[]> {\n  if (!cachedNormalized) {\n    const patterns = getPatterns();\n    const map = new Map<string, PatternEntry[]>();\n    for (const pattern of patterns) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name);\n        const existing = map.get(normalized) || [];\n        existing.push({\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n        map.set(normalized, existing);\n      }\n    }\n    cachedNormalized = map;\n  }\n  return cachedNormalized;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n * Uses exact string matching on normalized text.\n * Returns unique matches (same biomarker won't be matched twice).\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const matchedCodes = new Set<string>();\n  const matches: AnchorMatch[] = [];\n  const normalizedPatterns = getNormalizedPatterns();\n\n  for (const [normalizedName, entries] of normalizedPatterns) {\n    if (normalizedName.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalizedName)) {\n      continue;\n    }\n\n    let position = -1;\n    if (normalizedName.length <= 4) {\n      const regex = new RegExp(`\\\\b${normalizedName}\\\\b`);\n      const match = regex.exec(normalizedText);\n      if (match) {\n        position = match.index;\n      }\n    } else {\n      position = normalizedText.indexOf(normalizedName);\n    }\n\n    if (position !== -1) {\n      for (const entry of entries) {\n        if (!matchedCodes.has(entry.code)) {\n          matchedCodes.add(entry.code);\n          matches.push({\n            code: entry.code,\n            confidence: 1.0,\n            loinc: entry.loinc,\n            matchedName: entry.original,\n            position,\n          });\n        }\n      }\n    }\n  }\n\n  const scanTimeMs = Date.now() - startTime;\n  const matchedCodesArray = Array.from(matchedCodes);\n\n  return {\n    filteredReference: generateFilteredLLMReference(matchedCodesArray),\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":";AAQA;AAAA,EAEE;AAAA,EACA;AAAA,OACK;AA0BP,SAAS,UAAU,MAAsB;AACvC,SAAO,KACJ,UAAU,KAAK,EACf,QAAQ,oBAAoB,EAAE,EAC9B,YAAY,EACZ,QAAQ,QAAQ,GAAG;AACxB;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;AAID,IAAI,iBAAkD;AACtD,IAAI,mBAAuD;AAE3D,SAAS,cAAwC;AAC/C,MAAI,CAAC,gBAAgB;AACnB,qBAAiB,qBAAqB;AAAA,EACxC;AACA,SAAO;AACT;AAEA,SAAS,wBAAqD;AAC5D,MAAI,CAAC,kBAAkB;AACrB,UAAM,WAAW,YAAY;AAC7B,UAAM,MAAM,oBAAI,IAA4B;AAC5C,eAAW,WAAW,UAAU;AAC9B,iBAAW,QAAQ,QAAQ,OAAO;AAChC,cAAM,aAAa,UAAU,IAAI;AACjC,cAAM,WAAW,IAAI,IAAI,UAAU,KAAK,CAAC;AACzC,iBAAS,KAAK;AAAA,UACZ,MAAM,QAAQ;AAAA,UACd,GAAI,QAAQ,SAAS,EAAE,OAAO,QAAQ,MAAM;AAAA,UAC5C,UAAU;AAAA,QACZ,CAAC;AACD,YAAI,IAAI,YAAY,QAAQ;AAAA,MAC9B;AAAA,IACF;AACA,uBAAmB;AAAA,EACrB;AACA,SAAO;AACT;AAOO,SAAS,qBAAqB,SAA+B;AAClE,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,iBAAiB,UAAU,OAAO;AACxC,QAAM,eAAe,oBAAI,IAAY;AACrC,QAAM,UAAyB,CAAC;AAChC,QAAM,qBAAqB,sBAAsB;AAEjD,aAAW,CAAC,gBAAgB,OAAO,KAAK,oBAAoB;AAC1D,QAAI,eAAe,SAAS,KAAK,CAAC,wBAAwB,IAAI,cAAc,GAAG;AAC7E;AAAA,IACF;AAEA,QAAI,WAAW;AACf,QAAI,eAAe,UAAU,GAAG;AAC9B,YAAM,QAAQ,IAAI,OAAO,MAAM,cAAc,KAAK;AAClD,YAAM,QAAQ,MAAM,KAAK,cAAc;AACvC,UAAI,OAAO;AACT,mBAAW,MAAM;AAAA,MACnB;AAAA,IACF,OAAO;AACL,iBAAW,eAAe,QAAQ,cAAc;AAAA,IAClD;AAEA,QAAI,aAAa,IAAI;AACnB,iBAAW,SAAS,SAAS;AAC3B,YAAI,CAAC,aAAa,IAAI,MAAM,IAAI,GAAG;AACjC,uBAAa,IAAI,MAAM,IAAI;AAC3B,kBAAQ,KAAK;AAAA,YACX,MAAM,MAAM;AAAA,YACZ,YAAY;AAAA,YACZ,OAAO,MAAM;AAAA,YACb,aAAa,MAAM;AAAA,YACnB;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,aAAa,KAAK,IAAI,IAAI;AAChC,QAAM,oBAAoB,MAAM,KAAK,YAAY;AAEjD,SAAO;AAAA,IACL,mBAAmB,6BAA6B,iBAAiB;AAAA,IACjE;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 * - 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]);\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\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\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    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\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;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;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;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,eAAW,SAAS,UAAU,SAAS;AACrC,UAAI,MAAM,aAAa,CAAC,UAAU;AAChC;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": "0.16.2",
+  "version": "0.16.4",
   "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": "^0.16.2"
+    "@precisa-saude/fhir": "^0.16.4"
   },
   "devDependencies": {
     "tsup": "^8.3.5",