@precisa-saude/fhir-ocr-utils 0.25.0 → 0.26.0

package/dist/index.cjs

@@ -1,4 +1,4 @@
-"use strict";Object.defineProperty(exports, "__esModule", {value: true});// src/anchor.ts
+"use strict";Object.defineProperty(exports, "__esModule", {value: true}); function _nullishCoalesce(lhs, rhsFn) { if (lhs != null) { return lhs; } else { return rhsFn(); } } function _optionalChain(ops) { let lastAccessLHS = undefined; let value = ops[0]; let i = 1; while (i < ops.length) { const op = ops[i]; const fn = ops[i + 1]; i += 2; if ((op === 'optionalAccess' || op === 'optionalCall') && value == null) { return undefined; } if (op === 'access' || op === 'optionalAccess') { lastAccessLHS = value; value = fn(value); } else if (op === 'call' || op === 'optionalCall') { value = fn((...args) => value.call(lastAccessLHS, ...args)); lastAccessLHS = undefined; } } return value; }// src/anchor.ts
 
 
 
@@ -364,10 +364,262 @@ function getMatchedCodes(result) {
   return result.matches.map((m) => m.code);
 }
 
+// src/extraction-schema.ts
+var LAB_EXTRACTION_SCHEMA = {
+  $id: "https://fhir-brasil.dev.br/schemas/lab-extraction.json",
+  $schema: "https://json-schema.org/draft/2020-12/schema",
+  additionalProperties: false,
+  properties: {
+    biomarkers: {
+      description: "The measurements read from the report.",
+      items: {
+        additionalProperties: false,
+        properties: {
+          confidence: {
+            description: "Confidence in this reading, from 0 to 1.",
+            maximum: 1,
+            minimum: 0,
+            type: "number"
+          },
+          loinc: {
+            anyOf: [{ type: "string" }, { type: "null" }],
+            description: "A LOINC code from the allowed list, or null when none of them applies."
+          },
+          name: {
+            description: "The measurement name as the report prints it.",
+            type: "string"
+          },
+          referenceMax: {
+            anyOf: [{ type: "number" }, { type: "null" }],
+            description: "Upper bound of the range printed on the report, or null."
+          },
+          referenceMin: {
+            anyOf: [{ type: "number" }, { type: "null" }],
+            description: "Lower bound of the range printed on the report, or null."
+          },
+          sourceText: {
+            description: "The snippet of the report carrying this measurement and its value.",
+            type: "string"
+          },
+          unit: {
+            description: "Unit as the report prints it. Empty string when there is none.",
+            type: "string"
+          },
+          value: {
+            anyOf: [{ type: "number" }, { type: "string" }],
+            description: "Numeric value, or text for a qualitative result."
+          }
+        },
+        required: ["name", "value", "unit", "sourceText", "confidence"],
+        type: "object"
+      },
+      type: "array"
+    }
+  },
+  required: ["biomarkers"],
+  title: "Laboratory report extraction",
+  type: "object"
+};
+
+// src/extraction-to-lab-result.ts
+
+function flagFor(b) {
+  if (typeof b.value !== "number") return "";
+  if (typeof b.referenceMax === "number" && b.value > b.referenceMax) return "H";
+  if (typeof b.referenceMin === "number" && b.value < b.referenceMin) return "L";
+  return "";
+}
+function extractionToLabResult(biomarkers, options = {}) {
+  const reportId = _nullishCoalesce(options.reportId, () => ( "laudo-demo"));
+  const userId = _nullishCoalesce(options.userId, () => ( "paciente-demo"));
+  const collectionDate = _nullishCoalesce(options.collectionDate, () => ( (/* @__PURE__ */ new Date()).toISOString().slice(0, 10)));
+  const observations = biomarkers.flatMap((b) => {
+    const code = b.loinc ? _fhir.loincToCode.call(void 0, b.loinc) : void 0;
+    if (!code) return [];
+    return [
+      {
+        biomarkerCode: code,
+        biomarkerName: b.name,
+        flag: flagFor(b),
+        ...typeof b.referenceMax === "number" ? { referenceMax: b.referenceMax } : {},
+        ...typeof b.referenceMin === "number" ? { referenceMin: b.referenceMin } : {},
+        reportId,
+        unit: b.unit,
+        value: b.value
+      }
+    ];
+  });
+  return {
+    observations,
+    profile: { name: "Paciente de Demonstra\xE7\xE3o", userId },
+    report: {
+      collectionDate,
+      createdAt: `${collectionDate}T00:00:00Z`,
+      overallStatus: observations.some((o) => o.flag !== "") ? "ANORMAL" : "NORMAL",
+      processingStatus: "complete",
+      reportId,
+      userId
+    }
+  };
+}
+
+// src/extraction-validator.ts
+
+var isRecord = (v) => typeof v === "object" && v !== null && !Array.isArray(v);
+function schemaErrors(raw) {
+  if (!isRecord(raw)) return ["n\xE3o \xE9 um objeto"];
+  const errors = [];
+  const { confidence, loinc, name, referenceMax, referenceMin, sourceText, unit, value } = raw;
+  if (typeof name !== "string" || name.length === 0) errors.push("`name` ausente ou vazio");
+  if (typeof sourceText !== "string" || sourceText.length === 0)
+    errors.push("`sourceText` ausente ou vazio");
+  if (typeof unit !== "string") errors.push("`unit` ausente");
+  if (typeof value !== "number" && typeof value !== "string") errors.push("`value` ausente");
+  if (typeof confidence !== "number" || confidence < 0 || confidence > 1)
+    errors.push("`confidence` fora de 0..1");
+  if (loinc !== void 0 && loinc !== null && typeof loinc !== "string")
+    errors.push("`loinc` n\xE3o \xE9 string nem null");
+  for (const [key, v] of [
+    ["referenceMax", referenceMax],
+    ["referenceMin", referenceMin]
+  ]) {
+    if (v !== void 0 && v !== null && typeof v !== "number")
+      errors.push(`\`${key}\` n\xE3o \xE9 n\xFAmero nem null`);
+  }
+  return errors;
+}
+function allowedKeys(anchors) {
+  const allowed = /* @__PURE__ */ new Set();
+  for (const match of anchors.matches) {
+    allowed.add(match.code);
+    if (match.loinc) allowed.add(match.loinc);
+  }
+  return allowed;
+}
+function validateExtraction(raw, options = {}) {
+  const { anchors } = options;
+  if (!isRecord(raw)) {
+    return { accepted: [], errors: ["a sa\xEDda n\xE3o \xE9 um objeto JSON"], rejected: [], valid: false };
+  }
+  if (!Array.isArray(raw.biomarkers)) {
+    return {
+      accepted: [],
+      errors: ["`biomarkers` ausente ou n\xE3o \xE9 lista"],
+      rejected: [],
+      valid: false
+    };
+  }
+  const allowed = anchors ? allowedKeys(anchors) : void 0;
+  const accepted = [];
+  const rejected = [];
+  for (const entry of raw.biomarkers) {
+    const problems = schemaErrors(entry);
+    if (problems.length > 0) {
+      rejected.push({ detail: problems.join("; "), raw: entry, reason: "schema" });
+      continue;
+    }
+    const biomarker = entry;
+    if (allowed) {
+      const loinc = _nullishCoalesce(biomarker.loinc, () => ( void 0));
+      if (!loinc) {
+        rejected.push({
+          detail: `"${biomarker.name}" veio sem c\xF3digo LOINC`,
+          raw: entry,
+          reason: "not-anchored"
+        });
+        continue;
+      }
+      const internalCode = _fhir.loincToCode.call(void 0, loinc);
+      const isAnchored = allowed.has(loinc) || internalCode !== void 0 && allowed.has(internalCode);
+      if (!isAnchored) {
+        rejected.push({
+          detail: `${loinc} n\xE3o foi ancorado no texto de origem`,
+          raw: entry,
+          reason: "not-anchored"
+        });
+        continue;
+      }
+    }
+    accepted.push(biomarker);
+  }
+  return { accepted, errors: [], rejected, valid: true };
+}
+function acceptedBiomarkers(raw, options = {}) {
+  return validateExtraction(raw, options).accepted;
+}
+
+// src/llm-client.ts
+function buildPrompt(text, allowed) {
+  return [
+    "Extract the laboratory results from the report below.",
+    "",
+    "Return JSON matching this schema, and nothing else:",
+    JSON.stringify(LAB_EXTRACTION_SCHEMA),
+    "",
+    "Use only LOINC codes from this list:",
+    allowed,
+    "",
+    "REPORT:",
+    text
+  ].join("\n");
+}
+function stripFence(raw) {
+  const fenced = /```(?:json)?\s*([\s\S]*?)```/.exec(raw);
+  return (_nullishCoalesce(_optionalChain([fenced, 'optionalAccess', _ => _[1]]), () => ( raw))).trim();
+}
+async function extractWithModel(text, options) {
+  const { apiKey, baseUrl, model, responseFormat, timeoutMs = 3e5 } = options;
+  const allowed = findBiomarkersInText(text).filteredReference;
+  const startedAt = Date.now();
+  const formatBody = (mode2) => JSON.stringify({
+    messages: [{ content: buildPrompt(text, allowed), role: "user" }],
+    model,
+    ...mode2 === "json_object" ? { response_format: { type: "json_object" } } : {},
+    ...mode2 === "json_schema" ? {
+      response_format: {
+        json_schema: { name: "lab_extraction", schema: LAB_EXTRACTION_SCHEMA, strict: true },
+        type: "json_schema"
+      }
+    } : {},
+    temperature: 0
+  });
+  const post = async (mode2) => fetch(`${baseUrl.replace(/\/$/, "")}/chat/completions`, {
+    body: formatBody(mode2),
+    headers: {
+      "content-type": "application/json",
+      ...apiKey ? { authorization: `Bearer ${apiKey}` } : {}
+    },
+    method: "POST",
+    signal: AbortSignal.timeout(timeoutMs)
+  });
+  const mode = _nullishCoalesce(responseFormat, () => ( "auto"));
+  let response = await post(mode === "auto" ? "json_schema" : mode);
+  if (!response.ok && mode === "auto" && response.status >= 400 && response.status < 500) {
+    response = await post("none");
+  }
+  if (!response.ok) {
+    throw new Error(`${String(response.status)} de ${baseUrl}: ${await response.text()}`);
+  }
+  const body = await response.json();
+  const raw = _nullishCoalesce(_optionalChain([body, 'access', _2 => _2.choices, 'optionalAccess', _3 => _3[0], 'optionalAccess', _4 => _4.message, 'optionalAccess', _5 => _5.content]), () => ( ""));
+  const tookMs = Date.now() - startedAt;
+  try {
+    return { payload: JSON.parse(stripFence(raw)), raw, tookMs };
+  } catch (e) {
+    throw new Error(`O modelo n\xE3o devolveu JSON analis\xE1vel. Resposta crua:
+${raw.slice(0, 500)}`);
+  }
+}
+
+
+
+
+
+
 
 
 
 
 
-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;
+exports.CONFIDENCE_AMBIGUOUS = CONFIDENCE_AMBIGUOUS; exports.CONFIDENCE_NAME_ONLY = CONFIDENCE_NAME_ONLY; exports.CONFIDENCE_VALUE_ADJACENT = CONFIDENCE_VALUE_ADJACENT; exports.LAB_EXTRACTION_SCHEMA = LAB_EXTRACTION_SCHEMA; exports.acceptedBiomarkers = acceptedBiomarkers; exports.extractWithModel = extractWithModel; exports.extractionToLabResult = extractionToLabResult; exports.findBiomarkersInText = findBiomarkersInText; exports.getMatchedCodes = getMatchedCodes; exports.validateExtraction = validateExtraction;
 //# 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;ACaA;AAEE;AACA;AACA;AAAA,2CACK;AAwBA,IAAM,0BAAA,EAA4B,CAAA;AAMlC,IAAM,qBAAA,EAAuB,GAAA;AAM7B,IAAM,qBAAA,EAAuB,GAAA;AAGpC,IAAM,yBAAA,EAA2B,CAAA;AA8BjC,SAAS,SAAA,CAAU,IAAA,EAAsB;AACvC,EAAA,OAAO,IAAA,CACJ,SAAA,CAAU,KAAK,CAAA,CACf,OAAA,CAAQ,kBAAA,EAAoB,EAAE,CAAA,CAC9B,WAAA,CAAY,CAAA,CACZ,OAAA,CAAQ,+BAAA,EAAiC,GAAG,CAAA,CAC5C,OAAA,CAAQ,WAAA,EAAa,GAAG,CAAA;AAC7B;AAEA,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACtC,KAAA;AAAA,EACA,KAAA;AAAA,EACA,IAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA,KAAA;AAAA,EACA;AACF,CAAC,CAAA;AAQD,IAAM,uBAAA,kBAAyB,IAAI,GAAA,CAAI;AAAA,EACrC,UAAA;AAAA;AAAA,EACA,WAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,YAAA;AAAA;AAAA,EACA,KAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,UAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,YAAA;AAAA,EACA,OAAA;AAAA,EACA,SAAA;AAAA,EACA;AACF,CAAC,CAAA;AAOD,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACtC,QAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,OAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,cAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,OAAA;AAAA,EACA,UAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,QAAA;AAAA,EACA,OAAA;AAAA,EACA,cAAA;AAAA,EACA;AACF,CAAC,CAAA;AASD,IAAM,yBAAA,EAAqC;AAAA,EACzC,qBAAA;AAAA;AAAA,EACA,kBAAA;AAAA;AAAA,EACA,iBAAA;AAAA;AAAA,EACA,uBAAA;AAAA;AAAA,EACA,cAAA;AAAA;AAAA,EACA,YAAA;AAAA,EACA,iBAAA;AAAA,EACA,YAAA;AAAA,EACA,cAAA;AAAA,EACA,gBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,gBAAA;AAAA,EACA,mBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,aAAA;AAAA,EACA,WAAA;AAAA,EACA,eAAA;AAAA,EACA,aAAA;AAAA,EACA;AACF,CAAA;AASA,IAAM,oBAAA,kBAAsB,IAAI,GAAA,CAAI;AAAA,EAClC,mBAAA;AAAA,EACA,eAAA;AAAA,EACA,qBAAA;AAAA,EACA,qBAAA;AAAA,EACA,oBAAA;AAAA,EACA,eAAA;AAAA,EACA;AACF,CAAC,CAAA;AAGD,IAAM,uBAAA,EAAmC;AAAA,EACvC,mBAAA;AAAA,EACA,sBAAA;AAAA,EACA,iBAAA;AAAA,EACA;AACF,CAAA;AAOA,IAAM,0BAAA,EAAsC,CAAC,aAAA,EAAe,kBAAA,EAAoB,aAAa,CAAA;AAa7F,IAAM,iBAAA,EAAmB,uBAAA;AAEzB,SAAS,eAAA,CAAgB,IAAA,EAAuB;AAC9C,EAAA,GAAA,CAAI,yBAAA,CAA0B,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,CAAA,EAAG;AACzD,IAAA,OAAO,KAAA;AAAA,EACT;AACA,EAAA,OAAO,sBAAA,CAAuB,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,EAAA,GAAK,gBAAA,CAAiB,IAAA,CAAK,IAAI,CAAA;AACzF;AAEA,IAAM,cAAA,EAAgB,IAAA;AAGtB,IAAI,iBAAA,EAAuC,IAAA;AAE3C,SAAS,aAAA,CAAA,EAA6B;AACpC,EAAA,GAAA,CAAI,CAAC,gBAAA,EAAkB;AACrB,IAAA,iBAAA,EAAmB,IAAI,GAAA;AAAA,MACrB,MAAA,CAAO,IAAA,CAAK,kBAAY,CAAA,CACrB,GAAA,CAAI,CAAC,IAAA,EAAA,GAAS,SAAA,CAAU,IAAI,CAAA,CAAE,IAAA,CAAK,CAAC,CAAA,CACpC,MAAA,CAAO,OAAO;AAAA,IACnB,CAAA;AAAA,EACF;AACA,EAAA,OAAO,gBAAA;AACT;AAqBA,IAAI,eAAA,EAAkD,IAAA;AACtD,IAAI,mBAAA,EAAsD,IAAA;AAE1D,SAAS,WAAA,CAAA,EAAwC;AAC/C,EAAA,GAAA,CAAI,CAAC,cAAA,EAAgB;AACnB,IAAA,eAAA,EAAiB,wCAAA,CAAqB;AAAA,EACxC;AACA,EAAA,OAAO,cAAA;AACT;AAEA,SAAS,YAAA,CAAa,IAAA,EAAsB;AAC1C,EAAA,OAAO,IAAA,CAAK,OAAA,CAAQ,qBAAA,EAAuB,MAAM,CAAA;AACnD;AAkBA,SAAS,gBAAA,CAAiB,cAAA,EAAgC;AACxD,EAAA,MAAM,KAAA,EAAO,cAAA,CAAe,KAAA,CAAM,GAAG,CAAA,CAAE,GAAA,CAAI,YAAY,CAAA,CAAE,IAAA,CAAK,YAAY,CAAA;AAC1E,EAAA,MAAM,OAAA,EAAS,SAAA,CAAU,IAAA,CAAK,cAAc,EAAA,EAAI,KAAA,EAAO,EAAA;AACvD,EAAA,OAAO,IAAI,MAAA,CAAO,CAAA,mBAAA,EAAsB,IAAI,CAAA,EAAA;AAC9C;AAEsE;AAC3B,EAAA;AACD,EAAA;AAC1C;AAOiD;AACb,EAAA;AACzB,IAAA;AACT,EAAA;AACkC,EAAA;AACpC;AAEqD;AAC1B,EAAA;AACkB,IAAA;AACJ,IAAA;AACD,MAAA;AACG,QAAA;AAClB,QAAA;AACf,UAAA;AACF,QAAA;AAC8B,QAAA;AAC5B,UAAA;AACF,QAAA;AAC6B,QAAA;AAClB,QAAA;AACyB,UAAA;AACV,UAAA;AAC1B,QAAA;AACkB,QAAA;AACW,UAAA;AACb,UAAA;AACwB,UAAA;AAC5B,UAAA;AACX,QAAA;AACH,MAAA;AACF,IAAA;AACqB,IAAA;AACvB,EAAA;AACO,EAAA;AACT;AAEuF;AACxC,EAAA;AACA,EAAA;AACP,EAAA;AACxC;AAEkD;AACV,EAAA;AACxC;AAMiD;AACjB,EAAA;AACrB,IAAA;AACT,EAAA;AACiC,EAAA;AACF,EAAA;AACU,IAAA;AAC9B,MAAA;AACT,IAAA;AACF,EAAA;AACO,EAAA;AACT;AASgE;AAC7B,EAAA;AACU,EAAA;AACL,IAAA;AAClC,MAAA;AACF,IAAA;AACoB,IAAA;AACU,IAAA;AACZ,IAAA;AACA,IAAA;AACmB,IAAA;AACE,IAAA;AACA,MAAA;AACtB,MAAA;AACkB,MAAA;AACnC,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAa+D;AAC9B,EAAA;AACU,IAAA;AACzC,EAAA;AAC+B,EAAA;AACC,EAAA;AACW,IAAA;AACd,IAAA;AAEE,MAAA;AAG7B,IAAA;AACgB,IAAA;AACS,MAAA;AACzB,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAUoE;AACvC,EAAA;AACa,EAAA;AACQ,EAAA;AACF,EAAA;AACF,EAAA;AACA,EAAA;AAEJ,EAAA;AACK,IAAA;AAEH,IAAA;AACb,IAAA;AACG,MAAA;AACO,MAAA;AACrC,IAAA;AACa,IAAA;AACX,MAAA;AACF,IAAA;AAEuC,IAAA;AACX,IAAA;AACE,MAAA;AACM,MAAA;AACpC,IAAA;AAEoC,IAAA;AACX,IAAA;AACgB,MAAA;AACR,MAAA;AACjC,IAAA;AAEuC,IAAA;AACH,MAAA;AAChC,QAAA;AACF,MAAA;AAEqC,MAAA;AACnC,QAAA;AACF,MAAA;AAEiB,MAAA;AACI,MAAA;AACN,QAAA;AACM,MAAA;AACN,QAAA;AACf,MAAA;AAEsC,MAAA;AAGvB,MAAA;AAEH,MAAA;AACiB,QAAA;AACb,UAAA;AACZ,UAAA;AACa,UAAA;AACM,UAAA;AACC,UAAA;AACrB,QAAA;AACH,MAAA;AACF,IAAA;AACF,EAAA;AAE6C,EAAA;AACb,EAAA;AAEzB,EAAA;AACc,IAAA;AACnB,IAAA;AACO,IAAA;AACiB,MAAA;AACtB,MAAA;AAC6B,MAAA;AAC/B,IAAA;AACF,EAAA;AACF;AAKgE;AACvB,EAAA;AACzC;ADpO+C;AACA;AACA;AACA;AACA;AACA;AACA","file":"/home/runner/work/fhir-brasil/fhir-brasil/packages/ocr-utils/dist/index.cjs","sourcesContent":[null,"/**\n * OCR Anchor — Biomarker text anchoring\n *\n * Scans OCR text for biomarker names BEFORE sending to LLM.\n * This prevents hallucination by constraining what biomarkers\n * the LLM is allowed to extract.\n *\n * Matching is deliberately conservative: a name only anchors when it appears\n * as a whole token, is not swallowed by a longer biomarker name, is not inside\n * a genetic report line, and — for generic single-word names — sits on a line\n * that actually carries a value.\n */\n\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\n  UNIT_TO_UCUM,\n} from '@precisa-saude/fhir';\n\nexport interface AnchorMatch {\n  code: string;\n  confidence: number;\n  loinc?: string;\n  matchedName: string;\n  position: number;\n}\n\nexport interface AnchorResult {\n  filteredReference: string;\n  matches: AnchorMatch[];\n  stats: {\n    totalPatterns: number;\n    matchedCount: number;\n    scanTimeMs: number;\n  };\n}\n\n/**\n * Confidence assigned to a specific biomarker name found on a line that also\n * carries a value (a number, a unit, or an expected qualitative term).\n */\nexport const CONFIDENCE_VALUE_ADJACENT = 1.0;\n\n/**\n * Confidence assigned to a specific biomarker name with no value evidence\n * nearby — a section heading, or a mention in prose.\n */\nexport const CONFIDENCE_NAME_ONLY = 0.7;\n\n/**\n * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,\n * `Blood`, …) that only anchored because a value was found next to it.\n */\nexport const CONFIDENCE_AMBIGUOUS = 0.4;\n\n/** Cap on how many occurrences of the same name are inspected per document. */\nconst MAX_OCCURRENCES_PER_NAME = 5;\n\n/**\n * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Treats a hyphen that joins words as a space\n * - Collapses horizontal whitespace, but KEEPS line breaks — the line is the\n *   context window used to decide whether a match is a real biomarker mention\n *\n * O hífen entre palavras vira espaço porque o catálogo e o laboratório\n * discordam sobre ele o tempo todo: o catálogo escreve \"Proteína C-Reativa\" e\n * \"High-Density Lipoprotein\", e os laudos imprimem \"Proteína C Reativa\" e\n * \"High Density Lipoprotein\". Sem essa equivalência, 82 dos 159 nomes com\n * hífen deixam de ancorar na grafia que o documento usa.\n *\n * Não era teórico: um GGT de verdade foi descartado como alucinação em 27\n * laudos porque o documento escrevia \"Gama glutamil transferase\" e o catálogo\n * \"Gama-Glutamil Transferase\". Um caractere derrubava o valor antes de\n * qualquer validação.\n *\n * A troca exige **letra antes** do hífen, e por isso não toca em número:\n * o `-2.5` de um T-score e o `0-5` de uma faixa de urina seguem intactos.\n * Trocar sem essa guarda apagaria o sinal de um valor negativo, que é bem\n * pior que o problema original.\n *\n * O pré-filtro de substring do `collectCandidates` compara a chave com o texto\n * já normalizado, então a equivalência precisa nascer aqui: aplicada só na\n * regex, o `includes` descartaria o nome antes de ela rodar.\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/(?<=\\p{L})-(?=[\\p{L}\\p{N}])/gu, ' ')\n    .replace(/[^\\S\\n]+/g, ' ');\n}\n\nconst UNAMBIGUOUS_SHORT_NAMES = new Set([\n  'hdl',\n  'ldl',\n  'lh',\n  'tsh',\n  'crp',\n  'pcr',\n  'ggt',\n  'alt',\n  'ast',\n  'bun',\n  'wbc',\n  'rbc',\n  'mcv',\n  'mch',\n  'rdw',\n  'mpv',\n  'psa',\n  'fsh',\n  'hba1c',\n  'egfr',\n  'acr',\n  'esr',\n  'vhs',\n  'bmc',\n  'bmd',\n  'vat',\n  'dxa',\n  'dmo',\n  'cmo',\n  'ffm',\n  'lbm',\n  'mlg',\n  'tav',\n]);\n\n/**\n * Single-word catalog names that are ordinary words in EN/PT, so seeing them\n * proves nothing on its own. They only anchor when the line also carries a\n * value. Qualitative urine markers (`Color`, `Protein`, `Blood`, …) are\n * detected automatically — see `isQualitativeUrine` — and don't belong here.\n */\nconst CONTEXT_REQUIRED_NAMES = new Set([\n  'bacteria', // Bacteria_Urine — tem unidade, escapa da regra automática\n  'bacterias', // Bacteria_Urine\n  'lead', // Lead — verbo/substantivo comuníssimo em inglês\n  'peso', // TotalMass\n  'saturation', // TransferrinSaturation — \"oxygen saturation\", \"saturation index\"\n  'tap', // ProthrombinTime — \"tap\" em inglês\n  'volume', // VATVolume\n  'weight', // TotalMass\n  // Sítios de dobra pelo nome nu. São partes do corpo antes de serem medidas,\n  // e aparecem em prosa: num laudo de DEXA real, \"hips and thighs\" e\n  // \"abdominal region\" ancoravam dobra cutânea que o documento não tem.\n  // Exigir valor na linha separa a tabela do parágrafo.\n  'abdominal',\n  'chest',\n  'coxa',\n  'peitoral',\n  'subescapular',\n  'subscapular',\n  'suprailiac',\n  'thigh',\n  'triceps',\n  'tricipital',\n]);\n\n/**\n * Qualitative results expected next to a non-numeric biomarker\n * (urine dipstick, sediment, appearance). Normalized, single tokens —\n * \"não reagente\" is covered by `reagente`, \"não detectado\" by `detectado`.\n */\nconst QUALITATIVE_VALUE_TERMS = new Set([\n  'absent',\n  'alguns',\n  'amarela',\n  'amarelo',\n  'anormal',\n  'ausencia',\n  'ausente',\n  'ausentes',\n  'citrino',\n  'claro',\n  'clear',\n  'cloudy',\n  'colorless',\n  'detectado',\n  'detected',\n  'escuro',\n  'incolor',\n  'indetectavel',\n  'limpido',\n  'moderada',\n  'moderado',\n  'negativa',\n  'negative',\n  'negativo',\n  'normais',\n  'normal',\n  'numerosos',\n  'ocasional',\n  'positiva',\n  'positive',\n  'positivo',\n  'present',\n  'presente',\n  'presentes',\n  'raras',\n  'raro',\n  'raros',\n  'reagente',\n  'trace',\n  'traces',\n  'tracos',\n  'turvo',\n  'undetectable',\n  'yellow',\n]);\n\n/**\n * Signals that a line comes from a genetic/molecular report rather than from a\n * panel of measured values. Gene symbols collide with biomarker names (`APOB`\n * the gene vs. `ApoB` the lipoprotein), so the context — not a static HGNC\n * blocklist — is what tells them apart. Blocking the token itself would break\n * real lipid panels.\n */\nconst GENETIC_CONTEXT_PATTERNS: RegExp[] = [\n  /\\b[nx][mrpc]_\\d{6,}/, // RefSeq: NM_000384.2, NP_, NR_, XM_\n  /\\bens[gtp]\\d{6,}/, // Ensembl: ENSG00000084674\n  /\\bp\\.[a-z]{3}\\d/, // HGVS proteína: p.Trp448*\n  /\\bc\\.\\d+[acgt]?[>_+-]/, // HGVS codificante: c.1234A>G, c.76_78del\n  /\\brs\\d{4,}\\b/, // dbSNP\n  /\\bgenes?\\b/,\n  /\\bvariante?s?\\b/,\n  /\\bexons?\\b/,\n  /\\bzygosity\\b/,\n  /\\bzigosidade\\b/,\n  /\\balleles?\\b/,\n  /\\balelos?\\b/,\n  /\\bmutations?\\b/,\n  /\\bmutac(ao|oes)\\b/,\n  /\\bpathogenic/,\n  /\\bpatogenic/,\n  /\\bheterozyg/,\n  /\\bhomozyg/,\n  /\\bheterozigot/,\n  /\\bhomozigot/,\n  /\\bsequence change\\b/,\n];\n\n/**\n * Sítios de dobra cutânea cujo nome nu também nomeia uma circunferência:\n * \"Coxa\" aparece tanto em \"Dobra Cutânea Coxa\" quanto em \"Circunferência da\n * Coxa\". O termo nu precisa existir como alias, porque há laudo que imprime\n * só o sítio na coluna, então a desambiguação tem que vir do contexto da\n * linha, como já se faz com laudo genético.\n */\nconst SKINFOLD_SITE_CODES = new Set([\n  'SkinfoldAbdominal',\n  'SkinfoldChest',\n  'SkinfoldMidaxillary',\n  'SkinfoldSubscapular',\n  'SkinfoldSuprailiac',\n  'SkinfoldThigh',\n  'SkinfoldTriceps',\n]);\n\n/** Uma linha de circunferência ou perímetro não mede dobra. */\nconst GIRTH_CONTEXT_PATTERNS: RegExp[] = [\n  /\\bcircumference\\b/,\n  /\\bcircunferencias?\\b/,\n  /\\bperimetros?\\b/,\n  /\\bgirth\\b/,\n];\n\n/**\n * Só bloqueia quando a linha fala de circunferência e não fala de dobra:\n * \"Dobra Cutânea Coxa\" e \"Thigh Skinfold\" continuam ancorando normalmente,\n * e uma linha que traga as duas palavras é ambígua demais para descartar.\n */\nconst SKINFOLD_CONTEXT_PATTERNS: RegExp[] = [/\\bdobras?\\b/, /\\bskin ?folds?\\b/, /\\bpregas?\\b/];\n\n/**\n * Medida em centímetros numa linha de sítio corporal.\n *\n * Dobra cutânea é em milímetros, sempre: um valor em cm no mesmo sítio é\n * circunferência. É o desambiguador mais forte que existe aqui, porque não\n * depende de a folha escrever a palavra \"circunferência\", e num laudo de\n * antropometria a coluna costuma trazer só o sítio e o número.\n *\n * Rejeita cm em vez de exigir mm: há folha que imprime a unidade no cabeçalho\n * da coluna e não em cada linha, e exigir mm perderia essas.\n */\nconst CENTIMETRE_VALUE = /\\d\\s*(?:,\\d+\\s*)?cm\\b/;\n\nfunction hasGirthContext(line: string): boolean {\n  if (SKINFOLD_CONTEXT_PATTERNS.some((re) => re.test(line))) {\n    return false;\n  }\n  return GIRTH_CONTEXT_PATTERNS.some((re) => re.test(line)) || CENTIMETRE_VALUE.test(line);\n}\n\nconst DIGIT_PATTERN = /\\d/;\n\n/** Unit tokens reused from the core catalog instead of a parallel list. */\nlet cachedUnitTokens: Set<string> | null = null;\n\nfunction getUnitTokens(): Set<string> {\n  if (!cachedUnitTokens) {\n    cachedUnitTokens = new Set(\n      Object.keys(UNIT_TO_UCUM)\n        .map((unit) => normalize(unit).trim())\n        .filter(Boolean),\n    );\n  }\n  return cachedUnitTokens;\n}\n\ninterface PatternEntry {\n  ambiguous: boolean;\n  code: string;\n  loinc?: string;\n  original: string;\n}\n\ninterface NamePattern {\n  entries: PatternEntry[];\n  /** Built on first use — most names never match a given document. */\n  regex: RegExp | null;\n}\n\ninterface Candidate {\n  end: number;\n  entries: PatternEntry[];\n  start: number;\n}\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNamePatterns: Map<string, NamePattern> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&');\n}\n\n/**\n * Build a whole-token matcher for a normalized name.\n *\n * Lookarounds instead of `\\b` because names may start or end with a non-word\n * character (`Lp(a)`), where `\\b` asserts the wrong thing.\n *\n * A multi-word name must sit on a single line: in the column layouts labs\n * print, consecutive lines are separate biomarkers, and allowing a line break\n * inside a name turns \"Colesterol\\nHDL\" into the name \"Colesterol HDL\".\n * A wrapped name still anchors through its head token when that token is a\n * name of its own (\"Colesterol\\nTotal\" → `Cholesterol`).\n *\n * The trailing optional `s` keeps the plurals labs actually print\n * (\"Proteínas\", \"Cetonas\") anchored to the singular catalog name — without\n * letting `proteína` match inside `proteinúria`.\n */\nfunction buildNamePattern(normalizedName: string): RegExp {\n  const body = normalizedName.split(' ').map(escapeRegExp).join('[^\\\\S\\\\n]+');\n  const plural = /\\p{L}$/u.test(normalizedName) ? 's?' : '';\n  return new RegExp(`(?<![\\\\p{L}\\\\p{N}])${body}${plural}(?![\\\\p{L}\\\\p{N}])`, 'gu');\n}\n\nfunction isQualitativeUrine(pattern: BiomarkerSearchPattern): boolean {\n  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];\n  return categories.includes('urina') && !pattern.unit;\n}\n\n/**\n * A name is ambiguous when it is a single token that also reads as ordinary\n * text. Multi-word names (`Occult Blood`, `Urine Protein`) are specific enough\n * on their own.\n */\nfunction isAmbiguousName(normalizedName: string, pattern: BiomarkerSearchPattern): boolean {\n  if (normalizedName.includes(' ')) {\n    return false;\n  }\n  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);\n}\n\nfunction getNamePatterns(): Map<string, NamePattern> {\n  if (!cachedNamePatterns) {\n    const map = new Map<string, NamePattern>();\n    for (const pattern of getPatterns()) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name).trim();\n        if (!normalized) {\n          continue;\n        }\n        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {\n          continue;\n        }\n        let slot = map.get(normalized);\n        if (!slot) {\n          slot = { entries: [], regex: null };\n          map.set(normalized, slot);\n        }\n        slot.entries.push({\n          ambiguous: isAmbiguousName(normalized, pattern),\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n      }\n    }\n    cachedNamePatterns = map;\n  }\n  return cachedNamePatterns;\n}\n\nfunction getLineBounds(text: string, position: number): { end: number; start: number } {\n  const start = text.lastIndexOf('\\n', position) + 1;\n  const nextBreak = text.indexOf('\\n', position);\n  return { end: nextBreak === -1 ? text.length : nextBreak, start };\n}\n\nfunction hasGeneticContext(line: string): boolean {\n  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));\n}\n\n/**\n * Does this line carry something that looks like a measured result?\n * A digit, a known unit, or an expected qualitative term.\n */\nfunction hasValueEvidence(line: string): boolean {\n  if (DIGIT_PATTERN.test(line)) {\n    return true;\n  }\n  const unitTokens = getUnitTokens();\n  for (const token of line.split(/[^\\p{L}\\p{N}%/]+/u)) {\n    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {\n      return true;\n    }\n  }\n  return false;\n}\n\n/**\n * Cheap pre-filter before the (much costlier) boundary regex.\n *\n * Sound because `normalize` collapses horizontal whitespace to a single space\n * and a name never spans a line break: whenever the pattern can match, the\n * literal name is a substring of the text.\n */\nfunction collectCandidates(normalizedText: string): Candidate[] {\n  const candidates: Candidate[] = [];\n  for (const [name, slot] of getNamePatterns()) {\n    if (!normalizedText.includes(name)) {\n      continue;\n    }\n    const { entries } = slot;\n    const regex = (slot.regex ??= buildNamePattern(name));\n    regex.lastIndex = 0;\n    let occurrences = 0;\n    let match = regex.exec(normalizedText);\n    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {\n      candidates.push({ end: match.index + match[0].length, entries, start: match.index });\n      occurrences += 1;\n      match = regex.exec(normalizedText);\n    }\n  }\n  return candidates;\n}\n\n/**\n * Longest match wins: drop a match fully contained in a longer one, so\n * `Cholesterol` doesn't anchor inside `HDL Cholesterol` and `Blood` doesn't\n * anchor inside `Blood Glucose`.\n *\n * Strictly longer, not longer-or-equal: containment plus equal length means an\n * identical span, which only happens when two distinct catalog names match the\n * same text (a singular and its plural form, say). Dropping one of those by\n * catalog order would silently lose a code, and losing an anchor is worse than\n * keeping both — `findBiomarkersInText` dedups per code anyway.\n */\nfunction resolveOverlaps(candidates: Candidate[]): Candidate[] {\n  const sorted = [...candidates].sort(\n    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start,\n  );\n  const accepted: Candidate[] = [];\n  for (const candidate of sorted) {\n    const length = candidate.end - candidate.start;\n    const swallowed = accepted.some(\n      (other) =>\n        other.start <= candidate.start &&\n        candidate.end <= other.end &&\n        other.end - other.start > length,\n    );\n    if (!swallowed) {\n      accepted.push(candidate);\n    }\n  }\n  return accepted;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n *\n * Matching is whole-token, longest-match-wins, and context-aware: matches\n * inside genetic report lines are discarded, and generic names only anchor\n * when a value sits on the same line. Returns one match per biomarker code —\n * the highest-confidence occurrence.\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const bestByCode = new Map<string, AnchorMatch>();\n  const geneticLines = new Map<number, boolean>();\n  const valueLines = new Map<number, boolean>();\n  const girthLines = new Map<number, boolean>();\n\n  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {\n    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);\n\n    let genetic = geneticLines.get(lineStart);\n    if (genetic === undefined) {\n      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));\n      geneticLines.set(lineStart, genetic);\n    }\n    if (genetic) {\n      continue;\n    }\n\n    let hasValue = valueLines.get(lineStart);\n    if (hasValue === undefined) {\n      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));\n      valueLines.set(lineStart, hasValue);\n    }\n\n    let girth = girthLines.get(lineStart);\n    if (girth === undefined) {\n      girth = hasGirthContext(normalizedText.slice(lineStart, lineEnd));\n      girthLines.set(lineStart, girth);\n    }\n\n    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\n        continue;\n      }\n\n      if (girth && SKINFOLD_SITE_CODES.has(entry.code)) {\n        continue;\n      }\n\n      let confidence = CONFIDENCE_NAME_ONLY;\n      if (entry.ambiguous) {\n        confidence = CONFIDENCE_AMBIGUOUS;\n      } else if (hasValue) {\n        confidence = CONFIDENCE_VALUE_ADJACENT;\n      }\n\n      const existing = bestByCode.get(entry.code);\n      const better =\n        !existing ||\n        confidence > existing.confidence ||\n        (confidence === existing.confidence && candidate.start < existing.position);\n      if (better) {\n        bestByCode.set(entry.code, {\n          code: entry.code,\n          confidence,\n          loinc: entry.loinc,\n          matchedName: entry.original,\n          position: candidate.start,\n        });\n      }\n    }\n  }\n\n  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);\n  const scanTimeMs = Date.now() - startTime;\n\n  return {\n    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),\n    matches,\n    stats: {\n      matchedCount: matches.length,\n      scanTimeMs,\n      totalPatterns: getPatterns().length,\n    },\n  };\n}\n\n/**\n * Get the list of matched biomarker codes from an anchor result.\n */\nexport function getMatchedCodes(result: AnchorResult): string[] {\n  return result.matches.map((m) => m.code);\n}\n"]}
+{"version":3,"sources":["/home/runner/work/fhir-brasil/fhir-brasil/packages/ocr-utils/dist/index.cjs","../src/anchor.ts","../src/extraction-schema.ts","../src/extraction-to-lab-result.ts","../src/extraction-validator.ts","../src/llm-client.ts"],"names":[],"mappings":"AAAA;ACaA;AAEE;AACA;AACA;AAAA,2CACK;AAwBA,IAAM,0BAAA,EAA4B,CAAA;AAMlC,IAAM,qBAAA,EAAuB,GAAA;AAM7B,IAAM,qBAAA,EAAuB,GAAA;AAGpC,IAAM,yBAAA,EAA2B,CAAA;AA8BjC,SAAS,SAAA,CAAU,IAAA,EAAsB;AACvC,EAAA,OAAO,IAAA,CACJ,SAAA,CAAU,KAAK,CAAA,CACf,OAAA,CAAQ,kBAAA,EAAoB,EAAE,CAAA,CAC9B,WAAA,CAAY,CAAA,CACZ,OAAA,CAAQ,+BAAA,EAAiC,GAAG,CAAA,CAC5C,OAAA,CAAQ,WAAA,EAAa,GAAG,CAAA;AAC7B;AAEA,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACt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uBAAA,kBAAyB,IAAI,GAAA,CAAI;AAAA,EACrC,UAAA;AAAA;AAAA,EACA,WAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,MAAA;AAAA;AAAA,EACA,YAAA;AAAA;AAAA,EACA,KAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA,EACA,QAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,UAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,YAAA;AAAA,EACA,OAAA;AAAA,EACA,SAAA;AAAA,EACA;AACF,CAAC,CAAA;AAOD,IAAM,wBAAA,kBAA0B,IAAI,GAAA,CAAI;AAAA,EACtC,QAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,OAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,QAAA;AAAA,EACA,SAAA;AAAA,EACA,cAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,QAAA;AAAA,EACA,WAAA;AAAA,EACA,WAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,UAAA;AAAA,EACA,SAAA;AAAA,EACA,UAAA;AAAA,EACA,WAAA;AAAA,EACA,OAAA;AAAA,EACA,MAAA;AAAA,EACA,OAAA;AAAA,EACA,UAAA;AAAA,EACA,OAAA;AAAA,EACA,QAAA;AAAA,EACA,QAAA;AAAA,EACA,OAAA;AAAA,EACA,cAAA;AAAA,EACA;AACF,CAAC,CAAA;AASD,IAAM,yBAAA,EAAqC;AAAA,EACzC,qBAAA;AAAA;AAAA,EACA,kBAAA;AAAA;AAAA,EACA,iBAAA;AAAA;AAAA,EACA,uBAAA;AAAA;AAAA,EACA,cAAA;AAAA;AAAA,EACA,YAAA;AAAA,EACA,iBAAA;AAAA,EACA,YAAA;AAAA,EACA,cAAA;AAAA,EACA,gBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,gBAAA;AAAA,EACA,mBAAA;AAAA,EACA,cAAA;AAAA,EACA,aAAA;AAAA,EACA,aAAA;AAAA,EACA,WAAA;AAAA,EACA,eAAA;AAAA,EACA,aAAA;AAAA,EACA;AACF,CAAA;AASA,IAAM,oBAAA,kBAAsB,IAAI,GAAA,CAAI;AAAA,EAClC,mBAAA;AAAA,EACA,eAAA;AAAA,EACA,qBAAA;AAAA,EACA,qBAAA;AAAA,EACA,oBAAA;AAAA,EACA,eAAA;AAAA,EACA;AACF,CAAC,CAAA;AAGD,IAAM,uBAAA,EAAmC;AAAA,EACvC,mBAAA;AAAA,EACA,sBAAA;AAAA,EACA,iBAAA;AAAA,EACA;AACF,CAAA;AAOA,IAAM,0BAAA,EAAsC,CAAC,aAAA,EAAe,kBAAA,EAAoB,aAAa,CAAA;AAa7F,IAAM,iBAAA,EAAmB,uBAAA;AAEzB,SAAS,eAAA,CAAgB,IAAA,EAAuB;AAC9C,EAAA,GAAA,CAAI,yBAAA,CAA0B,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,CAAA,EAAG;AACzD,IAAA,OAAO,KAAA;AAAA,EACT;AACA,EAAA,OAAO,sBAAA,CAAuB,IAAA,CAAK,CAAC,EAAA,EAAA,GAAO,EAAA,CAAG,IAAA,CAAK,IAAI,CAAC,EAAA,GAAK,gBAAA,CAAiB,IAAA,CAAK,IAAI,CAAA;AACzF;AAEA,IAAM,cAAA,EAAgB,IAAA;AAGtB,IAAI,iBAAA,EAAuC,IAAA;AAE3C,SAAS,aAAA,CAAA,EAA6B;AACpC,EAAA,GAAA,CAAI,CAAC,gBAAA,EAAkB;AACrB,IAAA,iBAAA,EAAmB,IAAI,GAAA;AAAA,MACrB,MAAA,CAAO,IAAA,CAAK,kBAAY,CAAA,CACrB,GAAA,CAAI,CAAC,IAAA,EAAA,GAAS,SAAA,CAAU,IAAI,CAAA,CAAE,IAAA,CAAK,CAAC,CAAA,CACpC,MAAA,CAAO,OAAO;AAAA,IACnB,CAAA;AAAA,EACF;AACA,EAAA,OAAO,gBAAA;AACT;AAqBA,IAAI,eAAA,EAAkD,IAAA;AACtD,IAAI,mBAAA,EAAsD,IAAA;AAE1D,SAAS,WAAA,CAAA,EAAwC;AAC/C,EAAA,GAAA,CAAI,CAAC,cAAA,EAAgB;AACnB,IAAA,eAAA,EAAiB,wCAAA,CAAqB;AAAA,EACxC;AACA,EAAA,OAAO,cAAA;AACT;AAEA,SAAS,YAAA,CAAa,IAAA,EAAsB;AAC1C,EAAA,OAAO,IAAA,CAAK,OAAA,CAAQ,qBAAA,EAAuB,MAAM,CAAA;AACnD;AAkBA,SAAS,gBAAA,CAAiB,cAAA,EAAgC;AACxD,EAAA,MAAM,KAAA,EAAO,cAAA,CAAe,KAAA,CAAM,GAAG,CAAA,CAAE,GAAA,CAAI,YAAY,CAAA,CAAE,IAAA,CAAK,YAAY,CAAA;AAC1E,EAAA,MAAM,OAAA,EAAS,SAAA,CAAU,IAAA,CAAK,cAAc,EAAA,EAAI,KAAA,EAAO,EAAA;AACvD,EAAA,OAAO,IAAI,MAAA,CAAO,CAAA,mBAAA,EAAsB,IAAI,CAAA,EAAA;AAC9C;AAEsE;AAC3B,EAAA;AACD,EAAA;AAC1C;AAOiD;AACb,EAAA;AACzB,IAAA;AACT,EAAA;AACkC,EAAA;AACpC;AAEqD;AAC1B,EAAA;AACkB,IAAA;AACJ,IAAA;AACD,MAAA;AACG,QAAA;AAClB,QAAA;AACf,UAAA;AACF,QAAA;AAC8B,QAAA;AAC5B,UAAA;AACF,QAAA;AAC6B,QAAA;AAClB,QAAA;AACyB,UAAA;AACV,UAAA;AAC1B,QAAA;AACkB,QAAA;AACW,UAAA;AACb,UAAA;AACwB,UAAA;AAC5B,UAAA;AACX,QAAA;AACH,MAAA;AACF,IAAA;AACqB,IAAA;AACvB,EAAA;AACO,EAAA;AACT;AAEuF;AACxC,EAAA;AACA,EAAA;AACP,EAAA;AACxC;AAEkD;AACV,EAAA;AACxC;AAMiD;AACjB,EAAA;AACrB,IAAA;AACT,EAAA;AACiC,EAAA;AACF,EAAA;AACU,IAAA;AAC9B,MAAA;AACT,IAAA;AACF,EAAA;AACO,EAAA;AACT;AASgE;AAC7B,EAAA;AACU,EAAA;AACL,IAAA;AAClC,MAAA;AACF,IAAA;AACoB,IAAA;AACU,IAAA;AACZ,IAAA;AACA,IAAA;AACmB,IAAA;AACE,IAAA;AACA,MAAA;AACtB,MAAA;AACkB,MAAA;AACnC,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAa+D;AAC9B,EAAA;AACU,IAAA;AACzC,EAAA;AAC+B,EAAA;AACC,EAAA;AACW,IAAA;AACd,IAAA;AAEE,MAAA;AAG7B,IAAA;AACgB,IAAA;AACS,MAAA;AACzB,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAUoE;AACvC,EAAA;AACa,EAAA;AACQ,EAAA;AACF,EAAA;AACF,EAAA;AACA,EAAA;AAEJ,EAAA;AACK,IAAA;AAEH,IAAA;AACb,IAAA;AACG,MAAA;AACO,MAAA;AACrC,IAAA;AACa,IAAA;AACX,MAAA;AACF,IAAA;AAEuC,IAAA;AACX,IAAA;AACE,MAAA;AACM,MAAA;AACpC,IAAA;AAEoC,IAAA;AACX,IAAA;AACgB,MAAA;AACR,MAAA;AACjC,IAAA;AAEuC,IAAA;AACH,MAAA;AAChC,QAAA;AACF,MAAA;AAEqC,MAAA;AACnC,QAAA;AACF,MAAA;AAEiB,MAAA;AACI,MAAA;AACN,QAAA;AACM,MAAA;AACN,QAAA;AACf,MAAA;AAEsC,MAAA;AAGvB,MAAA;AAEH,MAAA;AACiB,QAAA;AACb,UAAA;AACZ,UAAA;AACa,UAAA;AACM,UAAA;AACC,UAAA;AACrB,QAAA;AACH,MAAA;AACF,IAAA;AACF,EAAA;AAE6C,EAAA;AACb,EAAA;AAEzB,EAAA;AACc,IAAA;AACnB,IAAA;AACO,IAAA;AACiB,MAAA;AACtB,MAAA;AAC6B,MAAA;AAC/B,IAAA;AACF,EAAA;AACF;AAKgE;AACvB,EAAA;AACzC;ADpO+C;AACA;AEnVV;AAC9B,EAAA;AACI,EAAA;AACa,EAAA;AACV,EAAA;AACE,IAAA;AACG,MAAA;AACN,MAAA;AACiB,QAAA;AACV,QAAA;AACE,UAAA;AACG,YAAA;AACJ,YAAA;AACA,YAAA;AACH,YAAA;AACR,UAAA;AACO,UAAA;AACyB,YAAA;AACjB,YAAA;AACf,UAAA;AACM,UAAA;AACS,YAAA;AACP,YAAA;AACR,UAAA;AACc,UAAA;AACkB,YAAA;AACjB,YAAA;AACf,UAAA;AACc,UAAA;AACkB,YAAA;AACjB,YAAA;AACf,UAAA;AACY,UAAA;AACG,YAAA;AACP,YAAA;AACR,UAAA;AACM,UAAA;AACS,YAAA;AACP,YAAA;AACR,UAAA;AACO,UAAA;AACyB,YAAA;AACjB,YAAA;AACf,UAAA;AACF,QAAA;AACoC,QAAA;AAC9B,QAAA;AACR,MAAA;AACM,MAAA;AACR,IAAA;AACF,EAAA;AACuB,EAAA;AAChB,EAAA;AACD,EAAA;AACR;AFqV+C;AACA;AGvanB;AAyC4B;AACd,EAAA;AACI,EAAA;AACA,EAAA;AACrC,EAAA;AACT;AAIE;AAEqC,EAAA;AACJ,EAAA;AACF,EAAA;AAEU,EAAA;AAIG,IAAA;AACvB,IAAA;AAEZ,IAAA;AACL,MAAA;AACiB,QAAA;AACE,QAAA;AACF,QAAA;AACe,QAAA;AACA,QAAA;AAC9B,QAAA;AACQ,QAAA;AACC,QAAA;AACX,MAAA;AACF,IAAA;AACD,EAAA;AAEM,EAAA;AACL,IAAA;AACiB,IAAA;AACT,IAAA;AACN,MAAA;AAC4B,MAAA;AACY,MAAA;AACtB,MAAA;AAClB,MAAA;AACA,MAAA;AACF,IAAA;AACF,EAAA;AACF;AHuX+C;AACA;AIjdnB;AAwDD;AAGmB;AAChB,EAAA;AAEF,EAAA;AACO,EAAA;AAEI,EAAA;AACC,EAAA;AACO,IAAA;AACH,EAAA;AACF,EAAA;AACF,EAAA;AACG,IAAA;AAEI,EAAA;AAC/B,IAAA;AAES,EAAA;AACQ,IAAA;AACA,IAAA;AACnB,EAAA;AACkC,IAAA;AACtB,MAAA;AACxB,EAAA;AAEO,EAAA;AACT;AAOyD;AACvB,EAAA;AACK,EAAA;AACb,IAAA;AACkB,IAAA;AAC1C,EAAA;AACO,EAAA;AACT;AAS8B;AACR,EAAA;AAEA,EAAA;AACc,IAAA;AAClC,EAAA;AACoC,EAAA;AAC3B,IAAA;AACM,MAAA;AACF,MAAA;AACE,MAAA;AACJ,MAAA;AACT,IAAA;AACF,EAAA;AAE6C,EAAA;AACL,EAAA;AACD,EAAA;AAEH,EAAA;AACC,IAAA;AACV,IAAA;AACe,MAAA;AACtC,MAAA;AACF,IAAA;AAEkB,IAAA;AAEL,IAAA;AACsB,MAAA;AAGrB,MAAA;AACI,QAAA;AACc,UAAA;AACrB,UAAA;AACG,UAAA;AACT,QAAA;AACD,QAAA;AACF,MAAA;AAIsC,MAAA;AAEb,MAAA;AACR,MAAA;AACD,QAAA;AACI,UAAA;AACX,UAAA;AACG,UAAA;AACT,QAAA;AACD,QAAA;AACF,MAAA;AACF,IAAA;AAEuB,IAAA;AACzB,EAAA;AAEyC,EAAA;AAC3C;AAMwB;AACkB,EAAA;AAC1C;AJqX+C;AACA;AK7ea;AACnD,EAAA;AACL,IAAA;AACA,IAAA;AACA,IAAA;AACoC,IAAA;AACpC,IAAA;AACA,IAAA;AACA,IAAA;AACA,IAAA;AACA,IAAA;AACA,IAAA;AACS,EAAA;AACb;AAGyC;AACxB,EAAA;AACkB,EAAA;AACnC;AAY0B;AACQ,EAAA;AAIW,EAAA;AAEhB,EAAA;AAGV,EAAA;AAC2B,IAAA;AACxC,IAAA;AAC+B,IAAA;AAE3B,IAAA;AACmB,MAAA;AACwB,QAAA;AACjC,QAAA;AACR,MAAA;AAED,IAAA;AACQ,IAAA;AACd,EAAA;AAGQ,EAAA;AACc,IAAA;AACZ,IAAA;AACS,MAAA;AACwB,MAAA;AAC1C,IAAA;AACQ,IAAA;AAC6B,IAAA;AACtC,EAAA;AAE4B,EAAA;AACa,EAAA;AAIL,EAAA;AACT,IAAA;AAC9B,EAAA;AAEkB,EAAA;AAC0B,IAAA;AAC5C,EAAA;AAEkC,EAAA;AAGM,EAAA;AACZ,EAAA;AAExB,EAAA;AACyC,IAAA;AACrC,EAAA;AACU,IAAA;AAA6E;AAC/F,EAAA;AACF;ALgd+C;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA","file":"/home/runner/work/fhir-brasil/fhir-brasil/packages/ocr-utils/dist/index.cjs","sourcesContent":[null,"/**\n * OCR Anchor — Biomarker text anchoring\n *\n * Scans OCR text for biomarker names BEFORE sending to LLM.\n * This prevents hallucination by constraining what biomarkers\n * the LLM is allowed to extract.\n *\n * Matching is deliberately conservative: a name only anchors when it appears\n * as a whole token, is not swallowed by a longer biomarker name, is not inside\n * a genetic report line, and — for generic single-word names — sits on a line\n * that actually carries a value.\n */\n\nimport {\n  type BiomarkerSearchPattern,\n  generateFilteredLLMReference,\n  getAllSearchPatterns,\n  UNIT_TO_UCUM,\n} from '@precisa-saude/fhir';\n\nexport interface AnchorMatch {\n  code: string;\n  confidence: number;\n  loinc?: string;\n  matchedName: string;\n  position: number;\n}\n\nexport interface AnchorResult {\n  filteredReference: string;\n  matches: AnchorMatch[];\n  stats: {\n    totalPatterns: number;\n    matchedCount: number;\n    scanTimeMs: number;\n  };\n}\n\n/**\n * Confidence assigned to a specific biomarker name found on a line that also\n * carries a value (a number, a unit, or an expected qualitative term).\n */\nexport const CONFIDENCE_VALUE_ADJACENT = 1.0;\n\n/**\n * Confidence assigned to a specific biomarker name with no value evidence\n * nearby — a section heading, or a mention in prose.\n */\nexport const CONFIDENCE_NAME_ONLY = 0.7;\n\n/**\n * Confidence assigned to a generic/ambiguous name (`Color`, `Protein`,\n * `Blood`, …) that only anchored because a value was found next to it.\n */\nexport const CONFIDENCE_AMBIGUOUS = 0.4;\n\n/** Cap on how many occurrences of the same name are inspected per document. */\nconst MAX_OCCURRENCES_PER_NAME = 5;\n\n/**\n * Normalize text for comparison:\n * - Removes diacritics (ã→a, ç→c, é→e)\n * - Converts to lowercase\n * - Treats a hyphen that joins words as a space\n * - Collapses horizontal whitespace, but KEEPS line breaks — the line is the\n *   context window used to decide whether a match is a real biomarker mention\n *\n * O hífen entre palavras vira espaço porque o catálogo e o laboratório\n * discordam sobre ele o tempo todo: o catálogo escreve \"Proteína C-Reativa\" e\n * \"High-Density Lipoprotein\", e os laudos imprimem \"Proteína C Reativa\" e\n * \"High Density Lipoprotein\". Sem essa equivalência, 82 dos 159 nomes com\n * hífen deixam de ancorar na grafia que o documento usa.\n *\n * Não era teórico: um GGT de verdade foi descartado como alucinação em 27\n * laudos porque o documento escrevia \"Gama glutamil transferase\" e o catálogo\n * \"Gama-Glutamil Transferase\". Um caractere derrubava o valor antes de\n * qualquer validação.\n *\n * A troca exige **letra antes** do hífen, e por isso não toca em número:\n * o `-2.5` de um T-score e o `0-5` de uma faixa de urina seguem intactos.\n * Trocar sem essa guarda apagaria o sinal de um valor negativo, que é bem\n * pior que o problema original.\n *\n * O pré-filtro de substring do `collectCandidates` compara a chave com o texto\n * já normalizado, então a equivalência precisa nascer aqui: aplicada só na\n * regex, o `includes` descartaria o nome antes de ela rodar.\n */\nfunction normalize(text: string): string {\n  return text\n    .normalize('NFD')\n    .replace(/[\\u0300-\\u036f]/g, '')\n    .toLowerCase()\n    .replace(/(?<=\\p{L})-(?=[\\p{L}\\p{N}])/gu, ' ')\n    .replace(/[^\\S\\n]+/g, ' ');\n}\n\nconst UNAMBIGUOUS_SHORT_NAMES = new Set([\n  'hdl',\n  'ldl',\n  'lh',\n  'tsh',\n  'crp',\n  'pcr',\n  'ggt',\n  'alt',\n  'ast',\n  'bun',\n  'wbc',\n  'rbc',\n  'mcv',\n  'mch',\n  'rdw',\n  'mpv',\n  'psa',\n  'fsh',\n  'hba1c',\n  'egfr',\n  'acr',\n  'esr',\n  'vhs',\n  'bmc',\n  'bmd',\n  'vat',\n  'dxa',\n  'dmo',\n  'cmo',\n  'ffm',\n  'lbm',\n  'mlg',\n  'tav',\n]);\n\n/**\n * Single-word catalog names that are ordinary words in EN/PT, so seeing them\n * proves nothing on its own. They only anchor when the line also carries a\n * value. Qualitative urine markers (`Color`, `Protein`, `Blood`, …) are\n * detected automatically — see `isQualitativeUrine` — and don't belong here.\n */\nconst CONTEXT_REQUIRED_NAMES = new Set([\n  'bacteria', // Bacteria_Urine — tem unidade, escapa da regra automática\n  'bacterias', // Bacteria_Urine\n  'lead', // Lead — verbo/substantivo comuníssimo em inglês\n  'peso', // TotalMass\n  'saturation', // TransferrinSaturation — \"oxygen saturation\", \"saturation index\"\n  'tap', // ProthrombinTime — \"tap\" em inglês\n  'volume', // VATVolume\n  'weight', // TotalMass\n  // Sítios de dobra pelo nome nu. São partes do corpo antes de serem medidas,\n  // e aparecem em prosa: num laudo de DEXA real, \"hips and thighs\" e\n  // \"abdominal region\" ancoravam dobra cutânea que o documento não tem.\n  // Exigir valor na linha separa a tabela do parágrafo.\n  'abdominal',\n  'chest',\n  'coxa',\n  'peitoral',\n  'subescapular',\n  'subscapular',\n  'suprailiac',\n  'thigh',\n  'triceps',\n  'tricipital',\n]);\n\n/**\n * Qualitative results expected next to a non-numeric biomarker\n * (urine dipstick, sediment, appearance). Normalized, single tokens —\n * \"não reagente\" is covered by `reagente`, \"não detectado\" by `detectado`.\n */\nconst QUALITATIVE_VALUE_TERMS = new Set([\n  'absent',\n  'alguns',\n  'amarela',\n  'amarelo',\n  'anormal',\n  'ausencia',\n  'ausente',\n  'ausentes',\n  'citrino',\n  'claro',\n  'clear',\n  'cloudy',\n  'colorless',\n  'detectado',\n  'detected',\n  'escuro',\n  'incolor',\n  'indetectavel',\n  'limpido',\n  'moderada',\n  'moderado',\n  'negativa',\n  'negative',\n  'negativo',\n  'normais',\n  'normal',\n  'numerosos',\n  'ocasional',\n  'positiva',\n  'positive',\n  'positivo',\n  'present',\n  'presente',\n  'presentes',\n  'raras',\n  'raro',\n  'raros',\n  'reagente',\n  'trace',\n  'traces',\n  'tracos',\n  'turvo',\n  'undetectable',\n  'yellow',\n]);\n\n/**\n * Signals that a line comes from a genetic/molecular report rather than from a\n * panel of measured values. Gene symbols collide with biomarker names (`APOB`\n * the gene vs. `ApoB` the lipoprotein), so the context — not a static HGNC\n * blocklist — is what tells them apart. Blocking the token itself would break\n * real lipid panels.\n */\nconst GENETIC_CONTEXT_PATTERNS: RegExp[] = [\n  /\\b[nx][mrpc]_\\d{6,}/, // RefSeq: NM_000384.2, NP_, NR_, XM_\n  /\\bens[gtp]\\d{6,}/, // Ensembl: ENSG00000084674\n  /\\bp\\.[a-z]{3}\\d/, // HGVS proteína: p.Trp448*\n  /\\bc\\.\\d+[acgt]?[>_+-]/, // HGVS codificante: c.1234A>G, c.76_78del\n  /\\brs\\d{4,}\\b/, // dbSNP\n  /\\bgenes?\\b/,\n  /\\bvariante?s?\\b/,\n  /\\bexons?\\b/,\n  /\\bzygosity\\b/,\n  /\\bzigosidade\\b/,\n  /\\balleles?\\b/,\n  /\\balelos?\\b/,\n  /\\bmutations?\\b/,\n  /\\bmutac(ao|oes)\\b/,\n  /\\bpathogenic/,\n  /\\bpatogenic/,\n  /\\bheterozyg/,\n  /\\bhomozyg/,\n  /\\bheterozigot/,\n  /\\bhomozigot/,\n  /\\bsequence change\\b/,\n];\n\n/**\n * Sítios de dobra cutânea cujo nome nu também nomeia uma circunferência:\n * \"Coxa\" aparece tanto em \"Dobra Cutânea Coxa\" quanto em \"Circunferência da\n * Coxa\". O termo nu precisa existir como alias, porque há laudo que imprime\n * só o sítio na coluna, então a desambiguação tem que vir do contexto da\n * linha, como já se faz com laudo genético.\n */\nconst SKINFOLD_SITE_CODES = new Set([\n  'SkinfoldAbdominal',\n  'SkinfoldChest',\n  'SkinfoldMidaxillary',\n  'SkinfoldSubscapular',\n  'SkinfoldSuprailiac',\n  'SkinfoldThigh',\n  'SkinfoldTriceps',\n]);\n\n/** Uma linha de circunferência ou perímetro não mede dobra. */\nconst GIRTH_CONTEXT_PATTERNS: RegExp[] = [\n  /\\bcircumference\\b/,\n  /\\bcircunferencias?\\b/,\n  /\\bperimetros?\\b/,\n  /\\bgirth\\b/,\n];\n\n/**\n * Só bloqueia quando a linha fala de circunferência e não fala de dobra:\n * \"Dobra Cutânea Coxa\" e \"Thigh Skinfold\" continuam ancorando normalmente,\n * e uma linha que traga as duas palavras é ambígua demais para descartar.\n */\nconst SKINFOLD_CONTEXT_PATTERNS: RegExp[] = [/\\bdobras?\\b/, /\\bskin ?folds?\\b/, /\\bpregas?\\b/];\n\n/**\n * Medida em centímetros numa linha de sítio corporal.\n *\n * Dobra cutânea é em milímetros, sempre: um valor em cm no mesmo sítio é\n * circunferência. É o desambiguador mais forte que existe aqui, porque não\n * depende de a folha escrever a palavra \"circunferência\", e num laudo de\n * antropometria a coluna costuma trazer só o sítio e o número.\n *\n * Rejeita cm em vez de exigir mm: há folha que imprime a unidade no cabeçalho\n * da coluna e não em cada linha, e exigir mm perderia essas.\n */\nconst CENTIMETRE_VALUE = /\\d\\s*(?:,\\d+\\s*)?cm\\b/;\n\nfunction hasGirthContext(line: string): boolean {\n  if (SKINFOLD_CONTEXT_PATTERNS.some((re) => re.test(line))) {\n    return false;\n  }\n  return GIRTH_CONTEXT_PATTERNS.some((re) => re.test(line)) || CENTIMETRE_VALUE.test(line);\n}\n\nconst DIGIT_PATTERN = /\\d/;\n\n/** Unit tokens reused from the core catalog instead of a parallel list. */\nlet cachedUnitTokens: Set<string> | null = null;\n\nfunction getUnitTokens(): Set<string> {\n  if (!cachedUnitTokens) {\n    cachedUnitTokens = new Set(\n      Object.keys(UNIT_TO_UCUM)\n        .map((unit) => normalize(unit).trim())\n        .filter(Boolean),\n    );\n  }\n  return cachedUnitTokens;\n}\n\ninterface PatternEntry {\n  ambiguous: boolean;\n  code: string;\n  loinc?: string;\n  original: string;\n}\n\ninterface NamePattern {\n  entries: PatternEntry[];\n  /** Built on first use — most names never match a given document. */\n  regex: RegExp | null;\n}\n\ninterface Candidate {\n  end: number;\n  entries: PatternEntry[];\n  start: number;\n}\n\nlet cachedPatterns: BiomarkerSearchPattern[] | null = null;\nlet cachedNamePatterns: Map<string, NamePattern> | null = null;\n\nfunction getPatterns(): BiomarkerSearchPattern[] {\n  if (!cachedPatterns) {\n    cachedPatterns = getAllSearchPatterns();\n  }\n  return cachedPatterns;\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&');\n}\n\n/**\n * Build a whole-token matcher for a normalized name.\n *\n * Lookarounds instead of `\\b` because names may start or end with a non-word\n * character (`Lp(a)`), where `\\b` asserts the wrong thing.\n *\n * A multi-word name must sit on a single line: in the column layouts labs\n * print, consecutive lines are separate biomarkers, and allowing a line break\n * inside a name turns \"Colesterol\\nHDL\" into the name \"Colesterol HDL\".\n * A wrapped name still anchors through its head token when that token is a\n * name of its own (\"Colesterol\\nTotal\" → `Cholesterol`).\n *\n * The trailing optional `s` keeps the plurals labs actually print\n * (\"Proteínas\", \"Cetonas\") anchored to the singular catalog name — without\n * letting `proteína` match inside `proteinúria`.\n */\nfunction buildNamePattern(normalizedName: string): RegExp {\n  const body = normalizedName.split(' ').map(escapeRegExp).join('[^\\\\S\\\\n]+');\n  const plural = /\\p{L}$/u.test(normalizedName) ? 's?' : '';\n  return new RegExp(`(?<![\\\\p{L}\\\\p{N}])${body}${plural}(?![\\\\p{L}\\\\p{N}])`, 'gu');\n}\n\nfunction isQualitativeUrine(pattern: BiomarkerSearchPattern): boolean {\n  const categories = Array.isArray(pattern.category) ? pattern.category : [pattern.category];\n  return categories.includes('urina') && !pattern.unit;\n}\n\n/**\n * A name is ambiguous when it is a single token that also reads as ordinary\n * text. Multi-word names (`Occult Blood`, `Urine Protein`) are specific enough\n * on their own.\n */\nfunction isAmbiguousName(normalizedName: string, pattern: BiomarkerSearchPattern): boolean {\n  if (normalizedName.includes(' ')) {\n    return false;\n  }\n  return CONTEXT_REQUIRED_NAMES.has(normalizedName) || isQualitativeUrine(pattern);\n}\n\nfunction getNamePatterns(): Map<string, NamePattern> {\n  if (!cachedNamePatterns) {\n    const map = new Map<string, NamePattern>();\n    for (const pattern of getPatterns()) {\n      for (const name of pattern.names) {\n        const normalized = normalize(name).trim();\n        if (!normalized) {\n          continue;\n        }\n        if (normalized.length < 3 && !UNAMBIGUOUS_SHORT_NAMES.has(normalized)) {\n          continue;\n        }\n        let slot = map.get(normalized);\n        if (!slot) {\n          slot = { entries: [], regex: null };\n          map.set(normalized, slot);\n        }\n        slot.entries.push({\n          ambiguous: isAmbiguousName(normalized, pattern),\n          code: pattern.code,\n          ...(pattern.loinc && { loinc: pattern.loinc }),\n          original: name,\n        });\n      }\n    }\n    cachedNamePatterns = map;\n  }\n  return cachedNamePatterns;\n}\n\nfunction getLineBounds(text: string, position: number): { end: number; start: number } {\n  const start = text.lastIndexOf('\\n', position) + 1;\n  const nextBreak = text.indexOf('\\n', position);\n  return { end: nextBreak === -1 ? text.length : nextBreak, start };\n}\n\nfunction hasGeneticContext(line: string): boolean {\n  return GENETIC_CONTEXT_PATTERNS.some((pattern) => pattern.test(line));\n}\n\n/**\n * Does this line carry something that looks like a measured result?\n * A digit, a known unit, or an expected qualitative term.\n */\nfunction hasValueEvidence(line: string): boolean {\n  if (DIGIT_PATTERN.test(line)) {\n    return true;\n  }\n  const unitTokens = getUnitTokens();\n  for (const token of line.split(/[^\\p{L}\\p{N}%/]+/u)) {\n    if (token && (unitTokens.has(token) || QUALITATIVE_VALUE_TERMS.has(token))) {\n      return true;\n    }\n  }\n  return false;\n}\n\n/**\n * Cheap pre-filter before the (much costlier) boundary regex.\n *\n * Sound because `normalize` collapses horizontal whitespace to a single space\n * and a name never spans a line break: whenever the pattern can match, the\n * literal name is a substring of the text.\n */\nfunction collectCandidates(normalizedText: string): Candidate[] {\n  const candidates: Candidate[] = [];\n  for (const [name, slot] of getNamePatterns()) {\n    if (!normalizedText.includes(name)) {\n      continue;\n    }\n    const { entries } = slot;\n    const regex = (slot.regex ??= buildNamePattern(name));\n    regex.lastIndex = 0;\n    let occurrences = 0;\n    let match = regex.exec(normalizedText);\n    while (match !== null && occurrences < MAX_OCCURRENCES_PER_NAME) {\n      candidates.push({ end: match.index + match[0].length, entries, start: match.index });\n      occurrences += 1;\n      match = regex.exec(normalizedText);\n    }\n  }\n  return candidates;\n}\n\n/**\n * Longest match wins: drop a match fully contained in a longer one, so\n * `Cholesterol` doesn't anchor inside `HDL Cholesterol` and `Blood` doesn't\n * anchor inside `Blood Glucose`.\n *\n * Strictly longer, not longer-or-equal: containment plus equal length means an\n * identical span, which only happens when two distinct catalog names match the\n * same text (a singular and its plural form, say). Dropping one of those by\n * catalog order would silently lose a code, and losing an anchor is worse than\n * keeping both — `findBiomarkersInText` dedups per code anyway.\n */\nfunction resolveOverlaps(candidates: Candidate[]): Candidate[] {\n  const sorted = [...candidates].sort(\n    (a, b) => b.end - b.start - (a.end - a.start) || a.start - b.start,\n  );\n  const accepted: Candidate[] = [];\n  for (const candidate of sorted) {\n    const length = candidate.end - candidate.start;\n    const swallowed = accepted.some(\n      (other) =>\n        other.start <= candidate.start &&\n        candidate.end <= other.end &&\n        other.end - other.start > length,\n    );\n    if (!swallowed) {\n      accepted.push(candidate);\n    }\n  }\n  return accepted;\n}\n\n/**\n * Find all biomarker names present in OCR text.\n *\n * Matching is whole-token, longest-match-wins, and context-aware: matches\n * inside genetic report lines are discarded, and generic names only anchor\n * when a value sits on the same line. Returns one match per biomarker code —\n * the highest-confidence occurrence.\n */\nexport function findBiomarkersInText(ocrText: string): AnchorResult {\n  const startTime = Date.now();\n  const normalizedText = normalize(ocrText);\n  const bestByCode = new Map<string, AnchorMatch>();\n  const geneticLines = new Map<number, boolean>();\n  const valueLines = new Map<number, boolean>();\n  const girthLines = new Map<number, boolean>();\n\n  for (const candidate of resolveOverlaps(collectCandidates(normalizedText))) {\n    const { end: lineEnd, start: lineStart } = getLineBounds(normalizedText, candidate.start);\n\n    let genetic = geneticLines.get(lineStart);\n    if (genetic === undefined) {\n      genetic = hasGeneticContext(normalizedText.slice(lineStart, lineEnd));\n      geneticLines.set(lineStart, genetic);\n    }\n    if (genetic) {\n      continue;\n    }\n\n    let hasValue = valueLines.get(lineStart);\n    if (hasValue === undefined) {\n      hasValue = hasValueEvidence(normalizedText.slice(lineStart, lineEnd));\n      valueLines.set(lineStart, hasValue);\n    }\n\n    let girth = girthLines.get(lineStart);\n    if (girth === undefined) {\n      girth = hasGirthContext(normalizedText.slice(lineStart, lineEnd));\n      girthLines.set(lineStart, girth);\n    }\n\n    for (const entry of candidate.entries) {\n      if (entry.ambiguous && !hasValue) {\n        continue;\n      }\n\n      if (girth && SKINFOLD_SITE_CODES.has(entry.code)) {\n        continue;\n      }\n\n      let confidence = CONFIDENCE_NAME_ONLY;\n      if (entry.ambiguous) {\n        confidence = CONFIDENCE_AMBIGUOUS;\n      } else if (hasValue) {\n        confidence = CONFIDENCE_VALUE_ADJACENT;\n      }\n\n      const existing = bestByCode.get(entry.code);\n      const better =\n        !existing ||\n        confidence > existing.confidence ||\n        (confidence === existing.confidence && candidate.start < existing.position);\n      if (better) {\n        bestByCode.set(entry.code, {\n          code: entry.code,\n          confidence,\n          loinc: entry.loinc,\n          matchedName: entry.original,\n          position: candidate.start,\n        });\n      }\n    }\n  }\n\n  const matches = Array.from(bestByCode.values()).sort((a, b) => a.position - b.position);\n  const scanTimeMs = Date.now() - startTime;\n\n  return {\n    filteredReference: generateFilteredLLMReference(matches.map((m) => m.code)),\n    matches,\n    stats: {\n      matchedCount: matches.length,\n      scanTimeMs,\n      totalPatterns: getPatterns().length,\n    },\n  };\n}\n\n/**\n * Get the list of matched biomarker codes from an anchor result.\n */\nexport function getMatchedCodes(result: AnchorResult): string[] {\n  return result.matches.map((m) => m.code);\n}\n","/**\n * Contrato de saída para extração de laudo por modelo.\n *\n * Este é o **contrato de interoperabilidade**, e não um prompt. Ele descreve a\n * forma do JSON que qualquer modelo precisa devolver para o resto do toolkit\n * conseguir conferir e converter o resultado. Não diz como pedir isso ao\n * modelo, não traz instrução de comportamento e não depende de fornecedor:\n * quem usa liga do jeito que a plataforma dele permitir (saída estruturada,\n * gramática, tool use ou simples prompt com validação por cima).\n *\n * As descrições são deliberadamente neutras. Regra de comportamento (\"nunca\n * infira\", \"copie literalmente\") é ajuste de prompt, muda de modelo para\n * modelo e não pertence a um contrato público.\n *\n * As descrições são as únicas strings em inglês do pacote, e isso é\n * deliberado: o schema é contrato de integração lido por quem consome de fora\n * do Brasil, e uma descrição em pt-BR não ajuda ninguém em Colônia ou Madri.\n * O resto da documentação segue a regra do ecossistema.\n *\n * O campo `sourceText` existe porque é o que torna a conferência possível:\n * sem o trecho que originou o valor não dá para auditar a extração depois.\n *\n * Campo que aceita mais de um tipo usa `anyOf`, e não `type: [...]`. As duas\n * formas são JSON Schema válido, mas decodificador restrito não engole a\n * segunda: o LM Studio recusa a geração com `'type' must be a string`. Como o\n * ponto do contrato é servir a qualquer modelo, vale a forma mais aceita.\n */\nexport const LAB_EXTRACTION_SCHEMA = {\n  $id: 'https://fhir-brasil.dev.br/schemas/lab-extraction.json',\n  $schema: 'https://json-schema.org/draft/2020-12/schema',\n  additionalProperties: false,\n  properties: {\n    biomarkers: {\n      description: 'The measurements read from the report.',\n      items: {\n        additionalProperties: false,\n        properties: {\n          confidence: {\n            description: 'Confidence in this reading, from 0 to 1.',\n            maximum: 1,\n            minimum: 0,\n            type: 'number',\n          },\n          loinc: {\n            anyOf: [{ type: 'string' }, { type: 'null' }],\n            description: 'A LOINC code from the allowed list, or null when none of them applies.',\n          },\n          name: {\n            description: 'The measurement name as the report prints it.',\n            type: 'string',\n          },\n          referenceMax: {\n            anyOf: [{ type: 'number' }, { type: 'null' }],\n            description: 'Upper bound of the range printed on the report, or null.',\n          },\n          referenceMin: {\n            anyOf: [{ type: 'number' }, { type: 'null' }],\n            description: 'Lower bound of the range printed on the report, or null.',\n          },\n          sourceText: {\n            description: 'The snippet of the report carrying this measurement and its value.',\n            type: 'string',\n          },\n          unit: {\n            description: 'Unit as the report prints it. Empty string when there is none.',\n            type: 'string',\n          },\n          value: {\n            anyOf: [{ type: 'number' }, { type: 'string' }],\n            description: 'Numeric value, or text for a qualitative result.',\n          },\n        },\n        required: ['name', 'value', 'unit', 'sourceText', 'confidence'],\n        type: 'object',\n      },\n      type: 'array',\n    },\n  },\n  required: ['biomarkers'],\n  title: 'Laboratory report extraction',\n  type: 'object',\n} as const;\n\n/** Uma grandeza como o modelo devolve, antes de qualquer conferência. */\nexport interface ExtractedBiomarker {\n  confidence: number;\n  loinc?: string | null;\n  name: string;\n  referenceMax?: number | null;\n  referenceMin?: number | null;\n  sourceText: string;\n  unit: string;\n  value: number | string;\n}\n\n/** O objeto inteiro que o modelo devolve. */\nexport interface ExtractionPayload {\n  biomarkers: ExtractedBiomarker[];\n}\n","import { loincToCode } from '@precisa-saude/fhir';\n\nimport type { ExtractedBiomarker } from './extraction-schema.js';\n\n/**\n * Converte grandezas já conferidas no envelope que o `fhir-bio convert` come.\n *\n * O laudo e o paciente não vêm do modelo: o contrato de extração cobre só as\n * grandezas. Os dois saem daqui com valores sintéticos e óbvios, na mesma\n * linha do `fhir-rnds-sandbox`, para a demo rodar de ponta a ponta sem inventar\n * identidade de ninguém. Quem integra de verdade troca os dois pelo que já tem.\n */\nexport interface LabResultEnvelope {\n  observations: {\n    biomarkerCode: string;\n    biomarkerName: string;\n    flag: 'H' | 'L' | '';\n    referenceMax?: number;\n    referenceMin?: number;\n    reportId: string;\n    unit: string;\n    value: number | string;\n  }[];\n  profile: { name: string; userId: string };\n  report: {\n    collectionDate: string;\n    createdAt: string;\n    overallStatus: 'ANORMAL' | 'NORMAL';\n    processingStatus: 'complete';\n    reportId: string;\n    userId: string;\n  };\n}\n\nexport interface ToLabResultOptions {\n  collectionDate?: string;\n  reportId?: string;\n  userId?: string;\n}\n\n/** `H`/`L` só quando o próprio laudo trouxe a faixa. Nunca inferida daqui. */\nfunction flagFor(b: ExtractedBiomarker): 'H' | 'L' | '' {\n  if (typeof b.value !== 'number') return '';\n  if (typeof b.referenceMax === 'number' && b.value > b.referenceMax) return 'H';\n  if (typeof b.referenceMin === 'number' && b.value < b.referenceMin) return 'L';\n  return '';\n}\n\nexport function extractionToLabResult(\n  biomarkers: ExtractedBiomarker[],\n  options: ToLabResultOptions = {},\n): LabResultEnvelope {\n  const reportId = options.reportId ?? 'laudo-demo';\n  const userId = options.userId ?? 'paciente-demo';\n  const collectionDate = options.collectionDate ?? new Date().toISOString().slice(0, 10);\n\n  const observations = biomarkers.flatMap((b) => {\n    // Sem código interno não há como converter, e o LOINC sozinho não basta\n    // para o `convert`. Cai fora em silêncio porque a checagem de ancoragem já\n    // rodou antes: o que chega aqui sem código é grandeza fora do catálogo.\n    const code = b.loinc ? loincToCode(b.loinc) : undefined;\n    if (!code) return [];\n\n    return [\n      {\n        biomarkerCode: code,\n        biomarkerName: b.name,\n        flag: flagFor(b),\n        ...(typeof b.referenceMax === 'number' ? { referenceMax: b.referenceMax } : {}),\n        ...(typeof b.referenceMin === 'number' ? { referenceMin: b.referenceMin } : {}),\n        reportId,\n        unit: b.unit,\n        value: b.value,\n      },\n    ];\n  });\n\n  return {\n    observations,\n    profile: { name: 'Paciente de Demonstração', userId },\n    report: {\n      collectionDate,\n      createdAt: `${collectionDate}T00:00:00Z`,\n      overallStatus: observations.some((o) => o.flag !== '') ? 'ANORMAL' : 'NORMAL',\n      processingStatus: 'complete',\n      reportId,\n      userId,\n    },\n  };\n}\n","import { loincToCode } from '@precisa-saude/fhir';\n\nimport type { AnchorResult } from './anchor.js';\nimport type { ExtractedBiomarker, ExtractionPayload } from './extraction-schema.js';\n\n/**\n * Conferência da saída do modelo contra o contrato e contra a ancoragem.\n *\n * São duas checagens, e as duas são determinísticas:\n *\n * 1. **Forma.** O objeto bate com `LAB_EXTRACTION_SCHEMA`. Modelo que devolve\n *    texto solto, campo faltando ou tipo errado é recusado aqui, o que deixa\n *    a qualidade do modelo virar problema de cobertura e nunca de correção.\n * 2. **Ancoragem.** O código veio da lista que a varredura liberou. Código que\n *    o laudo não mencionou é descartado, que é a falha cara: um valor\n *    plausível pendurado num exame que não estava na página.\n *\n * A validação de citação, a correção de código contra nome impresso e a\n * política de confiança não moram aqui.\n *\n * Sem dependência de runtime além do `@precisa-saude/fhir`: a checagem de\n * forma é escrita à mão porque o schema é pequeno e o pacote não carrega\n * validador de JSON Schema.\n */\n\n/** Por que uma grandeza foi recusada. */\nexport type RejectionReason = 'not-anchored' | 'schema';\n\nexport interface RejectedBiomarker {\n  /** Mensagem legível, já em pt-BR, dizendo o que falhou. */\n  detail: string;\n  /** O que o modelo devolveu, sem alteração, para o consumidor poder logar. */\n  raw: unknown;\n  reason: RejectionReason;\n}\n\nexport interface ExtractionValidationResult {\n  accepted: ExtractedBiomarker[];\n  /** Erros do objeto inteiro, quando nem dá para chegar nas grandezas. */\n  errors: string[];\n  rejected: RejectedBiomarker[];\n  /** `true` quando o objeto tem forma válida, mesmo que toda grandeza caia. */\n  valid: boolean;\n}\n\nexport interface ValidateExtractionOptions {\n  /**\n   * Resultado da ancoragem sobre o mesmo texto que foi ao modelo. Sem ele a\n   * checagem de ancoragem não roda e só a forma é conferida, que é um modo\n   * deliberadamente mais fraco: serve para inspecionar saída de modelo sem o\n   * laudo em mãos.\n   */\n  anchors?: AnchorResult;\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v);\n\n/** Confere uma grandeza contra o schema. Devolve a lista de problemas. */\nfunction schemaErrors(raw: unknown): string[] {\n  if (!isRecord(raw)) return ['não é um objeto'];\n\n  const errors: string[] = [];\n  const { confidence, loinc, name, referenceMax, referenceMin, sourceText, unit, value } = raw;\n\n  if (typeof name !== 'string' || name.length === 0) errors.push('`name` ausente ou vazio');\n  if (typeof sourceText !== 'string' || sourceText.length === 0)\n    errors.push('`sourceText` ausente ou vazio');\n  if (typeof unit !== 'string') errors.push('`unit` ausente');\n  if (typeof value !== 'number' && typeof value !== 'string') errors.push('`value` ausente');\n  if (typeof confidence !== 'number' || confidence < 0 || confidence > 1)\n    errors.push('`confidence` fora de 0..1');\n\n  if (loinc !== undefined && loinc !== null && typeof loinc !== 'string')\n    errors.push('`loinc` não é string nem null');\n\n  for (const [key, v] of [\n    ['referenceMax', referenceMax],\n    ['referenceMin', referenceMin],\n  ] as const) {\n    if (v !== undefined && v !== null && typeof v !== 'number')\n      errors.push(`\\`${key}\\` não é número nem null`);\n  }\n\n  return errors;\n}\n\n/**\n * O conjunto de códigos que a varredura liberou, pelos dois lados: o LOINC e o\n * código interno. O modelo devolve LOINC, mas aceitar o código interno também\n * evita recusar consumidor que prefira trabalhar com ele.\n */\nfunction allowedKeys(anchors: AnchorResult): Set<string> {\n  const allowed = new Set<string>();\n  for (const match of anchors.matches) {\n    allowed.add(match.code);\n    if (match.loinc) allowed.add(match.loinc);\n  }\n  return allowed;\n}\n\n/**\n * Confere a saída de um modelo contra o contrato e, quando a ancoragem é\n * fornecida, contra a lista de códigos que a varredura liberou.\n */\nexport function validateExtraction(\n  raw: unknown,\n  options: ValidateExtractionOptions = {},\n): ExtractionValidationResult {\n  const { anchors } = options;\n\n  if (!isRecord(raw)) {\n    return { accepted: [], errors: ['a saída não é um objeto JSON'], rejected: [], valid: false };\n  }\n  if (!Array.isArray(raw.biomarkers)) {\n    return {\n      accepted: [],\n      errors: ['`biomarkers` ausente ou não é lista'],\n      rejected: [],\n      valid: false,\n    };\n  }\n\n  const allowed = anchors ? allowedKeys(anchors) : undefined;\n  const accepted: ExtractedBiomarker[] = [];\n  const rejected: RejectedBiomarker[] = [];\n\n  for (const entry of raw.biomarkers) {\n    const problems = schemaErrors(entry);\n    if (problems.length > 0) {\n      rejected.push({ detail: problems.join('; '), raw: entry, reason: 'schema' });\n      continue;\n    }\n\n    const biomarker = entry as unknown as ExtractedBiomarker;\n\n    if (allowed) {\n      const loinc = biomarker.loinc ?? undefined;\n      // Sem código não há o que conferir contra a ancoragem, e aceitar assim\n      // deixaria passar justamente o caso que a varredura existe para pegar.\n      if (!loinc) {\n        rejected.push({\n          detail: `\"${biomarker.name}\" veio sem código LOINC`,\n          raw: entry,\n          reason: 'not-anchored',\n        });\n        continue;\n      }\n      // O `?? ''` de antes nunca deixava código inválido passar, porque string\n      // vazia não entra no conjunto de permitidos, mas obrigava quem lê a\n      // provar isso. A forma explícita não precisa de prova.\n      const internalCode = loincToCode(loinc);\n      const isAnchored =\n        allowed.has(loinc) || (internalCode !== undefined && allowed.has(internalCode));\n      if (!isAnchored) {\n        rejected.push({\n          detail: `${loinc} não foi ancorado no texto de origem`,\n          raw: entry,\n          reason: 'not-anchored',\n        });\n        continue;\n      }\n    }\n\n    accepted.push(biomarker);\n  }\n\n  return { accepted, errors: [], rejected, valid: true };\n}\n\n/** Só a lista de grandezas aprovadas, para quem não quer o relatório inteiro. */\nexport function acceptedBiomarkers(\n  raw: unknown,\n  options: ValidateExtractionOptions = {},\n): ExtractedBiomarker[] {\n  return validateExtraction(raw, options).accepted;\n}\n\nexport type { ExtractedBiomarker, ExtractionPayload };\n","import { findBiomarkersInText } from './anchor.js';\nimport { LAB_EXTRACTION_SCHEMA } from './extraction-schema.js';\n\n/**\n * Cliente mínimo para endpoint compatível com OpenAI.\n *\n * Isto é **conveniência, não contrato**. O contrato é o\n * `LAB_EXTRACTION_SCHEMA`, e o toolkit funciona inteiro sem esta função: quem\n * integra chama o próprio modelo do jeito que a plataforma dele permitir e\n * entrega o JSON ao `validateExtraction`. Esta função existe para a demo rodar\n * de uma ponta à outra sem um `curl` no meio.\n *\n * `/v1/chat/completions` é o que praticamente todo mundo fala: LM Studio,\n * Ollama, llama.cpp, vLLM, OpenRouter, OpenAI, e a Anthropic pelo endpoint de\n * compatibilidade. Por isso não há SDK de fornecedor aqui, e por isso o pacote\n * continua sem dependência de runtime: `fetch` é do Node.\n *\n * A chave **nunca** entra por argumento de linha de comando, só por variável de\n * ambiente: argumento fica no histórico do shell e na lista de processos.\n */\nexport interface ExtractOptions {\n  apiKey?: string;\n  baseUrl: string;\n  model: string;\n  /**\n   * Modo de saída estruturada. O padrão é negociar sozinho.\n   *\n   * Aqui é onde a compatibilidade quebra de verdade: o LM Studio recusa\n   * `json_object` com 400 e só aceita `json_schema`, a OpenAI aceita os dois,\n   * e servidor mais simples não conhece o campo. Como nenhum valor serve a\n   * todos, a primeira tentativa vai com `json_schema` e, se o servidor recusar,\n   * a segunda vai sem nada. Quem quiser fixar um modo passa ele aqui.\n   *\n   * Vale lembrar que isto mexe em **aproveitamento**, não em correção: saída\n   * malformada é recusada pela conferência de qualquer jeito.\n   */\n  responseFormat?: 'auto' | 'json_object' | 'json_schema' | 'none';\n  /** Milissegundos até desistir. Modelo local frio demora para carregar. */\n  timeoutMs?: number;\n}\n\nexport interface ExtractResult {\n  /** O JSON que o modelo devolveu, ainda sem conferência nenhuma. */\n  payload: unknown;\n  /** Texto cru da resposta, guardado para quando o parse falha. */\n  raw: string;\n  tookMs: number;\n}\n\n/**\n * O prompt é deliberadamente curto e neutro.\n *\n * Ele diz o que devolver e nada sobre como ler um laudo. Toda a instrução de\n * comportamento que um extrator de produção carrega é ajuste que muda de modelo\n * para modelo, e não pertence a um pacote público. O que sustenta a qualidade\n * aqui não é o prompt: é a conferência que roda depois.\n */\nfunction buildPrompt(text: string, allowed: string): string {\n  return [\n    'Extract the laboratory results from the report below.',\n    '',\n    'Return JSON matching this schema, and nothing else:',\n    JSON.stringify(LAB_EXTRACTION_SCHEMA),\n    '',\n    'Use only LOINC codes from this list:',\n    allowed,\n    '',\n    'REPORT:',\n    text,\n  ].join('\\n');\n}\n\n/** Modelo costuma embrulhar o JSON em cerca de markdown. Tira a cerca. */\nfunction stripFence(raw: string): string {\n  const fenced = /```(?:json)?\\s*([\\s\\S]*?)```/.exec(raw);\n  return (fenced?.[1] ?? raw).trim();\n}\n\n/**\n * Manda o laudo e a lista ancorada ao modelo e devolve o que ele respondeu.\n *\n * Não confere nada: a saída vai para o `validateExtraction`, que é onde a\n * ancoragem é cobrada. Separar os dois é proposital, porque é o que deixa\n * trocar de modelo sem mexer na parte que garante o resultado.\n */\nexport async function extractWithModel(\n  text: string,\n  options: ExtractOptions,\n): Promise<ExtractResult> {\n  const { apiKey, baseUrl, model, responseFormat, timeoutMs = 300_000 } = options;\n\n  // A própria ancoragem já monta a lista de permitidos, então o prompt e a\n  // conferência bebem exatamente da mesma fonte.\n  const allowed = findBiomarkersInText(text).filteredReference;\n\n  const startedAt = Date.now();\n\n  const formatBody = (mode: 'json_object' | 'json_schema' | 'none'): string =>\n    JSON.stringify({\n      messages: [{ content: buildPrompt(text, allowed), role: 'user' }],\n      model,\n      ...(mode === 'json_object' ? { response_format: { type: 'json_object' } } : {}),\n      ...(mode === 'json_schema'\n        ? {\n            response_format: {\n              json_schema: { name: 'lab_extraction', schema: LAB_EXTRACTION_SCHEMA, strict: true },\n              type: 'json_schema',\n            },\n          }\n        : {}),\n      temperature: 0,\n    });\n\n  const post = async (mode: 'json_object' | 'json_schema' | 'none'): Promise<Response> =>\n    fetch(`${baseUrl.replace(/\\/$/, '')}/chat/completions`, {\n      body: formatBody(mode),\n      headers: {\n        'content-type': 'application/json',\n        ...(apiKey ? { authorization: `Bearer ${apiKey}` } : {}),\n      },\n      method: 'POST',\n      signal: AbortSignal.timeout(timeoutMs),\n    });\n\n  const mode = responseFormat ?? 'auto';\n  let response = await post(mode === 'auto' ? 'json_schema' : mode);\n\n  // Servidor que não conhece o modo estrito responde 4xx. A segunda tentativa\n  // vai sem nada, que é o denominador comum, e só então o erro sobe.\n  if (!response.ok && mode === 'auto' && response.status >= 400 && response.status < 500) {\n    response = await post('none');\n  }\n\n  if (!response.ok) {\n    throw new Error(`${String(response.status)} de ${baseUrl}: ${await response.text()}`);\n  }\n\n  const body = (await response.json()) as {\n    choices?: { message?: { content?: string } }[];\n  };\n  const raw = body.choices?.[0]?.message?.content ?? '';\n  const tookMs = Date.now() - startedAt;\n\n  try {\n    return { payload: JSON.parse(stripFence(raw)), raw, tookMs };\n  } catch {\n    throw new Error(`O modelo não devolveu JSON analisável. Resposta crua:\\n${raw.slice(0, 500)}`);\n  }\n}\n"]}