@claritylabs/cl-sdk 0.13.0 → 0.13.1

package/dist/index.js

@@ -4330,9 +4330,15 @@ var SupplementarySchema = import_zod33.z.object({
   nonrenewalNoticeDays: import_zod33.z.number().optional().describe("Required notice period for nonrenewal in days"),
   auxiliaryFacts: import_zod33.z.array(AuxiliaryFactSchema2).optional().describe("Additional retrieval-only facts that do not fit the strict primary schema")
 });
-function buildSupplementaryPrompt() {
-  return `You are an expert insurance document analyst. Extract supplementary, retrieval-only information from this document.
+function buildSupplementaryPrompt(alreadyExtractedSummary) {
+  const exclusionBlock = alreadyExtractedSummary ? `
 
+IMPORTANT \u2014 The following facts have ALREADY been captured by prior extraction passes. Do NOT re-extract any of these. Your job is to find ADDITIONAL information that is missing from this list:
+
+${alreadyExtractedSummary}
+` : "";
+  return `You are an expert insurance document analyst. Extract supplementary, retrieval-only information from this document that is NOT already captured in the structured extraction results.
+${exclusionBlock}
 Focus on:
 - Regulatory contacts: state department of insurance, regulatory bodies, ombudsman offices \u2014 with phone, email, address
 - Claims contacts: how to report claims, claims department contact info, hours of operation
@@ -4346,8 +4352,10 @@ Focus on:
 Look for regulatory notices, complaint contact sections, claims reporting instructions, and cancellation/nonrenewal provisions throughout the document.
 
 For auxiliaryFacts:
+- ONLY capture facts that are NOT already present in the structured extraction results above.
+- Do not duplicate information that has already been extracted \u2014 no policy numbers, insured names, addresses, coverage limits, deductibles, or any other field that appears in the already-extracted data.
 - Capture concrete, policy-specific facts as structured key/value pairs.
-- Prioritize facts that agents may need later but that are often omitted from strict schemas: policyholder names, insured person names, driver names, ages, dates of birth, marital status, garaging information, lienholders, household members, vehicle assignments, schedule row details, and other discrete identifiers.
+- Prioritize facts that agents may need later but that are often omitted from strict schemas: policyholder names, insured person names, driver names, ages, dates of birth, marital status, garaging information, lienholders, household members, vehicle assignments, schedule row details, and other discrete identifiers \u2014 but ONLY if they are not already in the extracted data.
 - Use short normalized keys like "policyholder_name", "policyholder_age", "insured_name", "driver_age", "driver_date_of_birth", "garaging_zip", "vehicle_principal_driver".
 - Use subject when the fact belongs to a specific person, vehicle, property, or scheduled item.
 - Do not invent facts.
@@ -4741,6 +4749,38 @@ function createExtractor(config) {
       sectionCount: Array.isArray(sectionResult?.sections) ? sectionResult.sections.length : 0
     }, null, 2);
   }
+  function buildAlreadyExtractedSummary(memory) {
+    const lines = [];
+    const declarationResult = memory.get("declarations");
+    if (Array.isArray(declarationResult?.fields)) {
+      for (const field of declarationResult.fields) {
+        if (field.key && field.value) {
+          const subject = field.subject ? ` [${field.subject}]` : "";
+          lines.push(`- ${field.key}${subject}: ${field.value}`);
+        }
+      }
+    }
+    const coverageResult = memory.get("coverage_limits");
+    if (Array.isArray(coverageResult?.coverages)) {
+      for (const cov of coverageResult.coverages) {
+        const parts = [cov.name, cov.limit && `limit=${cov.limit}`, cov.deductible && `deductible=${cov.deductible}`].filter(Boolean);
+        if (parts.length > 0) lines.push(`- coverage: ${parts.join(", ")}`);
+      }
+    }
+    const namedInsured = memory.get("named_insured");
+    if (namedInsured) {
+      for (const [key, value] of Object.entries(namedInsured)) {
+        if (value && typeof value === "string") lines.push(`- ${key}: ${value}`);
+      }
+    }
+    const carrierInfo = memory.get("carrier_info");
+    if (carrierInfo) {
+      for (const [key, value] of Object.entries(carrierInfo)) {
+        if (value && typeof value === "string") lines.push(`- ${key}: ${value}`);
+      }
+    }
+    return lines.length > 0 ? lines.join("\n") : "";
+  }
   function formatPageMapSummary(pageAssignments) {
     const extractorPages = /* @__PURE__ */ new Map();
     for (const assignment of pageAssignments) {
@@ -5080,20 +5120,20 @@ function createExtractor(config) {
           mergeMemoryResult(result.name, result.data, memory);
         }
       }
-      const supplementaryExtractor = getExtractor("supplementary");
-      if (supplementaryExtractor) {
+      {
         onProgress?.("Extracting supplementary retrieval facts...");
         try {
+          const alreadyExtractedSummary = buildAlreadyExtractedSummary(memory);
           const supplementaryResult = await runExtractor({
             name: "supplementary",
-            prompt: supplementaryExtractor.buildPrompt(),
-            schema: supplementaryExtractor.schema,
+            prompt: buildSupplementaryPrompt(alreadyExtractedSummary),
+            schema: SupplementarySchema,
             pdfBase64,
             startPage: 1,
             endPage: pageCount,
             generateObject,
             convertPdfToImages,
-            maxTokens: supplementaryExtractor.maxTokens ?? 4096,
+            maxTokens: 4096,
             providerOptions
           });
           trackUsage(supplementaryResult.usage);