@claritylabs/cl-sdk 0.6.0 → 0.7.1

package/dist/index.mjs

@@ -71,6 +71,69 @@ function sanitizeNulls(obj) {
   return obj;
 }
 
+// src/core/safe-generate.ts
+async function safeGenerateObject(generateObject, params, options) {
+  const maxRetries = options?.maxRetries ?? 1;
+  let lastError;
+  for (let attempt = 0; attempt <= maxRetries; attempt++) {
+    try {
+      const result = await withRetry(
+        () => generateObject(params),
+        options?.log
+      );
+      return result;
+    } catch (error) {
+      lastError = error;
+      options?.onError?.(error, attempt);
+      await options?.log?.(
+        `safeGenerateObject attempt ${attempt + 1}/${maxRetries + 1} failed: ${error instanceof Error ? error.message : String(error)}`
+      );
+      if (attempt < maxRetries) {
+        await new Promise((resolve) => setTimeout(resolve, 1e3));
+      }
+    }
+  }
+  if (options?.fallback !== void 0) {
+    await options?.log?.(
+      `safeGenerateObject: all retries exhausted, returning fallback`
+    );
+    return { object: options.fallback };
+  }
+  throw lastError;
+}
+
+// src/core/pipeline.ts
+function createPipelineContext(opts) {
+  let latest = opts.resumeFrom;
+  const completedPhases = /* @__PURE__ */ new Set();
+  if (opts.resumeFrom) {
+    completedPhases.add(opts.resumeFrom.phase);
+  }
+  return {
+    id: opts.id,
+    async save(phase, state) {
+      const checkpoint = {
+        phase,
+        state,
+        timestamp: Date.now()
+      };
+      latest = checkpoint;
+      completedPhases.add(phase);
+      await opts.onSave?.(checkpoint);
+    },
+    getCheckpoint() {
+      return latest;
+    },
+    isPhaseComplete(phase) {
+      return completedPhases.has(phase);
+    },
+    clear() {
+      latest = void 0;
+      completedPhases.clear();
+    }
+  };
+}
+
 // src/schemas/enums.ts
 import { z } from "zod";
 var PolicyTypeSchema = z.enum([
@@ -471,11 +534,15 @@ var ExclusionSchema = z5.object({
 
 // src/schemas/condition.ts
 import { z as z6 } from "zod";
+var ConditionKeyValueSchema = z6.object({
+  key: z6.string(),
+  value: z6.string()
+});
 var PolicyConditionSchema = z6.object({
   name: z6.string(),
   conditionType: ConditionTypeSchema,
   content: z6.string(),
-  keyValues: z6.record(z6.string(), z6.string()).optional(),
+  keyValues: z6.array(ConditionKeyValueSchema).optional(),
   pageNumber: z6.number().optional()
 });
 
@@ -1652,21 +1719,33 @@ async function formatDocumentContent(doc, generateText, options) {
   for (let i = 0; i < entries.length; i += MAX_ENTRIES_PER_BATCH) {
     batches.push(entries.slice(i, i + MAX_ENTRIES_PER_BATCH));
   }
-  for (const batch of batches) {
-    const prompt = buildFormatPrompt(batch.map((e) => ({ id: e.id, text: e.text })));
-    const result = await withRetry(
-      () => generateText({
-        prompt,
-        maxTokens: 16384,
-        providerOptions: options?.providerOptions
-      })
-    );
-    if (result.usage) {
-      totalUsage.inputTokens += result.usage.inputTokens;
-      totalUsage.outputTokens += result.usage.outputTokens;
+  for (let batchIdx = 0; batchIdx < batches.length; batchIdx++) {
+    const batch = batches[batchIdx];
+    try {
+      const prompt = buildFormatPrompt(batch.map((e) => ({ id: e.id, text: e.text })));
+      const result = await withRetry(
+        () => generateText({
+          prompt,
+          maxTokens: 16384,
+          providerOptions: options?.providerOptions
+        })
+      );
+      if (result.usage) {
+        totalUsage.inputTokens += result.usage.inputTokens;
+        totalUsage.outputTokens += result.usage.outputTokens;
+      }
+      const formatted = parseFormatResponse(result.text);
+      if (formatted.size < batch.length) {
+        await options?.log?.(
+          `Format batch ${batchIdx + 1}/${batches.length}: model returned ${formatted.size}/${batch.length} entries \u2014 unformatted entries will keep original content`
+        );
+      }
+      applyFormattedContent(doc, batch, formatted);
+    } catch (error) {
+      await options?.log?.(
+        `Format batch ${batchIdx + 1}/${batches.length} failed, keeping original content: ${error instanceof Error ? error.message : String(error)}`
+      );
     }
-    const formatted = parseFormatResponse(result.text);
-    applyFormattedContent(doc, batch, formatted);
   }
   return { document: doc, usage: totalUsage };
 }
@@ -2507,9 +2586,13 @@ var ExtractionTaskSchema = z18.object({
   endPage: z18.number(),
   description: z18.string()
 });
+var PageMapEntrySchema = z18.object({
+  section: z18.string(),
+  pages: z18.string()
+});
 var ExtractionPlanSchema = z18.object({
   tasks: z18.array(ExtractionTaskSchema),
-  pageMap: z18.record(z18.string(), z18.string()).optional()
+  pageMap: z18.array(PageMapEntrySchema).optional()
 });
 function buildPlanPrompt(templateHints) {
   return `You are planning the extraction of an insurance document. You have already classified this document. Now scan the full document and create a page map + extraction plan.
@@ -2538,7 +2621,10 @@ Return JSON:
     { "extractorName": "carrier_info", "startPage": 1, "endPage": 2, "description": "Extract carrier details from declarations page" },
     ...
   ],
-  "pageMap": { "declarations": "pages 1-3", "endorsements": "pages 15-22", ... }
+  "pageMap": [
+    { "section": "declarations", "pages": "pages 1-3" },
+    { "section": "endorsements", "pages": "pages 15-22" }
+  ]
 }
 
 Create tasks that cover the entire document. Prefer specific extractors over generic "sections" where possible. Keep page ranges tight \u2014 only include pages relevant to each extractor.
@@ -2855,9 +2941,14 @@ Return JSON only.`;
 
 // src/prompts/extractors/declarations.ts
 import { z as z27 } from "zod";
-var DeclarationsExtractSchema = z27.record(z27.string(), z27.unknown()).describe(
-  "Flexible declarations data \u2014 structure varies by line of business. Keys are descriptive field names, values are the extracted data."
-);
+var DeclarationsFieldSchema = z27.object({
+  field: z27.string().describe("Descriptive field name (e.g. 'policyNumber', 'effectiveDate', 'coverageALimit')"),
+  value: z27.string().describe("Extracted value exactly as it appears in the document"),
+  section: z27.string().optional().describe("Section or grouping this field belongs to (e.g. 'Coverage Limits', 'Vehicle Schedule')")
+});
+var DeclarationsExtractSchema = z27.object({
+  fields: z27.array(DeclarationsFieldSchema).describe("All declarations page fields extracted as key-value pairs. Structure varies by line of business.")
+});
 function buildDeclarationsPrompt() {
   return `You are an expert insurance document analyst. Extract all declarations page data from this document into a flexible key-value structure.
 
@@ -2881,9 +2972,18 @@ For PERSONAL LINES declarations:
 - Flood (NFIP): flood zone, community number, building/contents coverage
 - Personal Articles: scheduled items list with appraised values
 
-Use descriptive field names as keys. Preserve original values exactly as they appear.
+Return each field as an object with "field" (descriptive name), "value" (exact text from document), and optional "section" (grouping).
 
-Return JSON only.`;
+Example output:
+{
+  "fields": [
+    { "field": "policyNumber", "value": "GL-2025-78432", "section": "Policy Info" },
+    { "field": "effectiveDate", "value": "04/10/2025", "section": "Policy Info" },
+    { "field": "eachOccurrenceLimit", "value": "$1,000,000", "section": "Coverage Limits" }
+  ]
+}
+
+Preserve original values exactly as they appear. Return JSON only.`;
 }
 
 // src/prompts/extractors/loss-history.ts
@@ -3031,7 +3131,8 @@ function createExtractor(config) {
     onTokenUsage,
     onProgress,
     log,
-    providerOptions
+    providerOptions,
+    onCheckpointSave
   } = config;
   const limit = pLimit(concurrency);
   let totalUsage = { inputTokens: 0, outputTokens: 0 };
@@ -3042,100 +3143,106 @@ function createExtractor(config) {
       onTokenUsage?.(usage);
     }
   }
-  async function extract(pdfBase64, documentId) {
+  async function extract(pdfBase64, documentId, options) {
     const id = documentId ?? `doc-${Date.now()}`;
     const memory = /* @__PURE__ */ new Map();
     totalUsage = { inputTokens: 0, outputTokens: 0 };
-    onProgress?.("Classifying document...");
-    const pageCount = await getPdfPageCount(pdfBase64);
-    const classifyResult = await withRetry(
-      () => generateObject({
-        prompt: buildClassifyPrompt(),
-        schema: ClassifyResultSchema,
-        maxTokens: 512,
-        providerOptions
-      })
-    );
-    trackUsage(classifyResult.usage);
-    memory.set("classify", classifyResult.object);
-    const { documentType, policyTypes } = classifyResult.object;
-    const primaryType = policyTypes[0] ?? "other";
-    const template = getTemplate(primaryType);
-    onProgress?.(`Planning extraction for ${primaryType} ${documentType}...`);
-    const templateHints = [
-      `Document type: ${primaryType} ${documentType}`,
-      `Expected sections: ${template.expectedSections.join(", ")}`,
-      `Page hints: ${Object.entries(template.pageHints).map(([k, v]) => `${k}: ${v}`).join("; ")}`,
-      `Total pages: ${pageCount}`
-    ].join("\n");
-    const planResult = await withRetry(
-      () => generateObject({
-        prompt: buildPlanPrompt(templateHints),
-        schema: ExtractionPlanSchema,
-        maxTokens: 2048,
-        providerOptions
-      })
-    );
-    trackUsage(planResult.usage);
-    const tasks = planResult.object.tasks;
-    onProgress?.(`Dispatching ${tasks.length} extractors...`);
-    const extractorResults = await Promise.all(
-      tasks.map(
-        (task) => limit(async () => {
-          const ext = getExtractor(task.extractorName);
-          if (!ext) {
-            await log?.(`Unknown extractor: ${task.extractorName}, skipping`);
-            return null;
-          }
-          onProgress?.(`Extracting ${task.extractorName} (pages ${task.startPage}-${task.endPage})...`);
-          try {
-            const result = await runExtractor({
-              name: task.extractorName,
-              prompt: ext.buildPrompt(),
-              schema: ext.schema,
-              pdfBase64,
-              startPage: task.startPage,
-              endPage: task.endPage,
-              generateObject,
-              convertPdfToImages,
-              maxTokens: ext.maxTokens ?? 4096,
-              providerOptions
-            });
-            trackUsage(result.usage);
-            return result;
-          } catch (error) {
-            await log?.(`Extractor ${task.extractorName} failed: ${error}`);
-            return null;
-          }
-        })
-      )
-    );
-    for (const result of extractorResults) {
-      if (result) {
-        memory.set(result.name, result.data);
+    const pipelineCtx = createPipelineContext({
+      id,
+      onSave: onCheckpointSave,
+      resumeFrom: options?.resumeFrom
+    });
+    const resumed = pipelineCtx.getCheckpoint()?.state;
+    if (resumed?.memory) {
+      for (const [k, v] of Object.entries(resumed.memory)) {
+        memory.set(k, v);
       }
     }
-    for (let round = 0; round < maxReviewRounds; round++) {
-      const extractedKeys = [...memory.keys()].filter((k) => k !== "classify");
-      const reviewResult = await withRetry(
-        () => generateObject({
-          prompt: buildReviewPrompt(template.required, extractedKeys),
-          schema: ReviewResultSchema,
-          maxTokens: 1024,
+    let classifyResult;
+    if (resumed?.classifyResult && pipelineCtx.isPhaseComplete("classify")) {
+      classifyResult = resumed.classifyResult;
+      onProgress?.("Resuming from checkpoint (classify complete)...");
+    } else {
+      onProgress?.("Classifying document...");
+      const pageCount2 = await getPdfPageCount(pdfBase64);
+      const classifyResponse = await safeGenerateObject(
+        generateObject,
+        {
+          prompt: buildClassifyPrompt(),
+          schema: ClassifyResultSchema,
+          maxTokens: 512,
           providerOptions
-        })
+        },
+        {
+          fallback: { documentType: "policy", policyTypes: ["other"], confidence: 0 },
+          log,
+          onError: (err, attempt) => log?.(`Classify attempt ${attempt + 1} failed: ${err}`)
+        }
       );
-      trackUsage(reviewResult.usage);
-      if (reviewResult.object.complete || reviewResult.object.additionalTasks.length === 0) {
-        onProgress?.("Extraction complete.");
-        break;
-      }
-      onProgress?.(`Review round ${round + 1}: dispatching ${reviewResult.object.additionalTasks.length} follow-up extractors...`);
-      const followUpResults = await Promise.all(
-        reviewResult.object.additionalTasks.map(
+      trackUsage(classifyResponse.usage);
+      classifyResult = classifyResponse.object;
+      memory.set("classify", classifyResult);
+      await pipelineCtx.save("classify", {
+        id,
+        pageCount: pageCount2,
+        classifyResult,
+        memory: Object.fromEntries(memory)
+      });
+    }
+    const { documentType, policyTypes } = classifyResult;
+    const primaryType = policyTypes[0] ?? "other";
+    const template = getTemplate(primaryType);
+    const pageCount = resumed?.pageCount ?? await getPdfPageCount(pdfBase64);
+    let plan;
+    if (resumed?.plan && pipelineCtx.isPhaseComplete("plan")) {
+      plan = resumed.plan;
+      onProgress?.("Resuming from checkpoint (plan complete)...");
+    } else {
+      onProgress?.(`Planning extraction for ${primaryType} ${documentType}...`);
+      const templateHints = [
+        `Document type: ${primaryType} ${documentType}`,
+        `Expected sections: ${template.expectedSections.join(", ")}`,
+        `Page hints: ${Object.entries(template.pageHints).map(([k, v]) => `${k}: ${v}`).join("; ")}`,
+        `Total pages: ${pageCount}`
+      ].join("\n");
+      const planResponse = await safeGenerateObject(
+        generateObject,
+        {
+          prompt: buildPlanPrompt(templateHints),
+          schema: ExtractionPlanSchema,
+          maxTokens: 2048,
+          providerOptions
+        },
+        {
+          fallback: {
+            tasks: [{ extractorName: "sections", startPage: 1, endPage: pageCount, description: "Full document fallback extraction" }]
+          },
+          log,
+          onError: (err, attempt) => log?.(`Plan attempt ${attempt + 1} failed: ${err}`)
+        }
+      );
+      trackUsage(planResponse.usage);
+      plan = planResponse.object;
+      await pipelineCtx.save("plan", {
+        id,
+        pageCount,
+        classifyResult,
+        plan,
+        memory: Object.fromEntries(memory)
+      });
+    }
+    if (!pipelineCtx.isPhaseComplete("extract")) {
+      const tasks = plan.tasks;
+      onProgress?.(`Dispatching ${tasks.length} extractors...`);
+      const extractorResults = await Promise.all(
+        tasks.map(
           (task) => limit(async () => {
             const ext = getExtractor(task.extractorName);
-            if (!ext) return null;
+            if (!ext) {
+              await log?.(`Unknown extractor: ${task.extractorName}, skipping`);
+              return null;
+            }
+            onProgress?.(`Extracting ${task.extractorName} (pages ${task.startPage}-${task.endPage})...`);
             try {
               const result = await runExtractor({
                 name: task.extractorName,
@@ -3152,28 +3259,114 @@ function createExtractor(config) {
               trackUsage(result.usage);
               return result;
             } catch (error) {
-              await log?.(`Follow-up extractor ${task.extractorName} failed: ${error}`);
+              await log?.(`Extractor ${task.extractorName} failed: ${error}`);
               return null;
             }
           })
         )
       );
-      for (const result of followUpResults) {
+      for (const result of extractorResults) {
         if (result) {
           memory.set(result.name, result.data);
         }
       }
+      await pipelineCtx.save("extract", {
+        id,
+        pageCount,
+        classifyResult,
+        plan,
+        memory: Object.fromEntries(memory)
+      });
+    }
+    if (!pipelineCtx.isPhaseComplete("review")) {
+      for (let round = 0; round < maxReviewRounds; round++) {
+        const extractedKeys = [...memory.keys()].filter((k) => k !== "classify");
+        const reviewResponse = await safeGenerateObject(
+          generateObject,
+          {
+            prompt: buildReviewPrompt(template.required, extractedKeys),
+            schema: ReviewResultSchema,
+            maxTokens: 1024,
+            providerOptions
+          },
+          {
+            fallback: { complete: true, missingFields: [], additionalTasks: [] },
+            log,
+            onError: (err, attempt) => log?.(`Review round ${round + 1} attempt ${attempt + 1} failed: ${err}`)
+          }
+        );
+        trackUsage(reviewResponse.usage);
+        if (reviewResponse.object.complete || reviewResponse.object.additionalTasks.length === 0) {
+          onProgress?.("Extraction complete.");
+          break;
+        }
+        onProgress?.(`Review round ${round + 1}: dispatching ${reviewResponse.object.additionalTasks.length} follow-up extractors...`);
+        const followUpResults = await Promise.all(
+          reviewResponse.object.additionalTasks.map(
+            (task) => limit(async () => {
+              const ext = getExtractor(task.extractorName);
+              if (!ext) return null;
+              try {
+                const result = await runExtractor({
+                  name: task.extractorName,
+                  prompt: ext.buildPrompt(),
+                  schema: ext.schema,
+                  pdfBase64,
+                  startPage: task.startPage,
+                  endPage: task.endPage,
+                  generateObject,
+                  convertPdfToImages,
+                  maxTokens: ext.maxTokens ?? 4096,
+                  providerOptions
+                });
+                trackUsage(result.usage);
+                return result;
+              } catch (error) {
+                await log?.(`Follow-up extractor ${task.extractorName} failed: ${error}`);
+                return null;
+              }
+            })
+          )
+        );
+        for (const result of followUpResults) {
+          if (result) {
+            memory.set(result.name, result.data);
+          }
+        }
+      }
+      await pipelineCtx.save("review", {
+        id,
+        pageCount,
+        classifyResult,
+        plan,
+        memory: Object.fromEntries(memory)
+      });
     }
     onProgress?.("Assembling document...");
     const document = assembleDocument(id, documentType, memory);
+    await pipelineCtx.save("assemble", {
+      id,
+      pageCount,
+      classifyResult,
+      plan,
+      memory: Object.fromEntries(memory),
+      document
+    });
     onProgress?.("Formatting extracted content...");
     const formatResult = await formatDocumentContent(document, generateText, {
       providerOptions,
-      onProgress
+      onProgress,
+      log
     });
     trackUsage(formatResult.usage);
     const chunks = chunkDocument(formatResult.document);
-    return { document: formatResult.document, chunks, tokenUsage: totalUsage };
+    const finalCheckpoint = pipelineCtx.getCheckpoint();
+    return {
+      document: formatResult.document,
+      chunks,
+      tokenUsage: totalUsage,
+      checkpoint: finalCheckpoint
+    };
   }
   return { extract };
 }
@@ -4036,7 +4229,6 @@ function createApplicationPipeline(config) {
     let state = {
       id,
       pdfBase64: void 0,
-      // Don't persist the full PDF in state
       title: void 0,
       applicationType: null,
       fields: [],
@@ -4047,13 +4239,20 @@ function createApplicationPipeline(config) {
       updatedAt: now
     };
     onProgress?.("Classifying document...");
-    const { result: classifyResult, usage: classifyUsage } = await classifyApplication(
-      pdfBase64.slice(0, 2e3),
-      // Send truncated content for classification
-      generateObject,
-      providerOptions
-    );
-    trackUsage(classifyUsage);
+    await applicationStore?.save(state);
+    let classifyResult;
+    try {
+      const { result, usage: classifyUsage } = await classifyApplication(
+        pdfBase64.slice(0, 2e3),
+        generateObject,
+        providerOptions
+      );
+      trackUsage(classifyUsage);
+      classifyResult = result;
+    } catch (error) {
+      await log?.(`Classification failed, treating as non-application: ${error instanceof Error ? error.message : String(error)}`);
+      classifyResult = { isApplication: false, confidence: 0, applicationType: null };
+    }
     if (!classifyResult.isApplication) {
       state.status = "complete";
       state.updatedAt = Date.now();
@@ -4063,13 +4262,28 @@ function createApplicationPipeline(config) {
     state.applicationType = classifyResult.applicationType;
     state.status = "extracting";
     state.updatedAt = Date.now();
+    await applicationStore?.save(state);
     onProgress?.("Extracting form fields...");
-    const { fields, usage: extractUsage } = await extractFields(
-      pdfBase64,
-      generateObject,
-      providerOptions
-    );
-    trackUsage(extractUsage);
+    let fields;
+    try {
+      const { fields: extractedFields, usage: extractUsage } = await extractFields(
+        pdfBase64,
+        generateObject,
+        providerOptions
+      );
+      trackUsage(extractUsage);
+      fields = extractedFields;
+    } catch (error) {
+      await log?.(`Field extraction failed: ${error instanceof Error ? error.message : String(error)}`);
+      fields = [];
+    }
+    if (fields.length === 0) {
+      await log?.("No fields extracted, completing pipeline with empty result");
+      state.status = "complete";
+      state.updatedAt = Date.now();
+      await applicationStore?.save(state);
+      return { state, tokenUsage: totalUsage };
+    }
     state.fields = fields;
     state.title = classifyResult.applicationType ?? void 0;
     state.status = "auto_filling";
@@ -4101,20 +4315,24 @@ function createApplicationPipeline(config) {
         limit(async () => {
           const unfilledFields2 = state.fields.filter((f) => !f.value);
           if (unfilledFields2.length === 0) return;
-          const { result: autoFillResult, usage: afUsage } = await autoFillFromContext(
-            unfilledFields2,
-            orgContext,
-            generateObject,
-            providerOptions
-          );
-          trackUsage(afUsage);
-          for (const match of autoFillResult.matches) {
-            const field = state.fields.find((f) => f.id === match.fieldId);
-            if (field && !field.value) {
-              field.value = match.value;
-              field.source = `auto-fill: ${match.contextKey}`;
-              field.confidence = match.confidence;
+          try {
+            const { result: autoFillResult, usage: afUsage } = await autoFillFromContext(
+              unfilledFields2,
+              orgContext,
+              generateObject,
+              providerOptions
+            );
+            trackUsage(afUsage);
+            for (const match of autoFillResult.matches) {
+              const field = state.fields.find((f) => f.id === match.fieldId);
+              if (field && !field.value) {
+                field.value = match.value;
+                field.source = `auto-fill: ${match.contextKey}`;
+                field.confidence = match.confidence;
+              }
             }
+          } catch (e) {
+            await log?.(`Auto-fill from context failed: ${e instanceof Error ? e.message : String(e)}`);
           }
         })
       );
@@ -4147,13 +4365,18 @@ function createApplicationPipeline(config) {
     if (unfilledFields.length > 0) {
       onProgress?.(`Batching ${unfilledFields.length} remaining questions...`);
       state.status = "batching";
-      const { result: batchResult, usage: batchUsage } = await batchQuestions(
-        unfilledFields,
-        generateObject,
-        providerOptions
-      );
-      trackUsage(batchUsage);
-      state.batches = batchResult.batches;
+      try {
+        const { result: batchResult, usage: batchUsage } = await batchQuestions(
+          unfilledFields,
+          generateObject,
+          providerOptions
+        );
+        trackUsage(batchUsage);
+        state.batches = batchResult.batches;
+      } catch (error) {
+        await log?.(`Batching failed, using single-batch fallback: ${error instanceof Error ? error.message : String(error)}`);
+        state.batches = [unfilledFields.map((f) => f.id)];
+      }
       state.currentBatchIndex = 0;
       state.status = "collecting";
     } else {
@@ -4180,32 +4403,49 @@ function createApplicationPipeline(config) {
       (f) => currentBatchFieldIds.includes(f.id)
     );
     onProgress?.("Classifying reply...");
-    const { intent, usage: intentUsage } = await classifyReplyIntent(
-      currentBatchFields,
-      replyText,
-      generateObject,
-      providerOptions
-    );
-    trackUsage(intentUsage);
-    let fieldsFilled = 0;
-    let responseText;
-    if (intent.hasAnswers) {
-      onProgress?.("Parsing answers...");
-      const { result: parseResult, usage: parseUsage } = await parseAnswers(
+    let intent;
+    try {
+      const { intent: classifiedIntent, usage: intentUsage } = await classifyReplyIntent(
         currentBatchFields,
         replyText,
         generateObject,
         providerOptions
       );
-      trackUsage(parseUsage);
-      for (const answer of parseResult.answers) {
-        const field = state.fields.find((f) => f.id === answer.fieldId);
-        if (field) {
-          field.value = answer.value;
-          field.source = "user";
-          field.confidence = "confirmed";
-          fieldsFilled++;
+      trackUsage(intentUsage);
+      intent = classifiedIntent;
+    } catch (error) {
+      await log?.(`Reply intent classification failed, defaulting to answers_only: ${error instanceof Error ? error.message : String(error)}`);
+      intent = {
+        primaryIntent: "answers_only",
+        hasAnswers: true,
+        questionText: void 0,
+        questionFieldIds: void 0,
+        lookupRequests: void 0
+      };
+    }
+    let fieldsFilled = 0;
+    let responseText;
+    if (intent.hasAnswers) {
+      onProgress?.("Parsing answers...");
+      try {
+        const { result: parseResult, usage: parseUsage } = await parseAnswers(
+          currentBatchFields,
+          replyText,
+          generateObject,
+          providerOptions
+        );
+        trackUsage(parseUsage);
+        for (const answer of parseResult.answers) {
+          const field = state.fields.find((f) => f.id === answer.fieldId);
+          if (field) {
+            field.value = answer.value;
+            field.source = "user";
+            field.confidence = "confirmed";
+            fieldsFilled++;
+          }
         }
+      } catch (error) {
+        await log?.(`Answer parsing failed: ${error instanceof Error ? error.message : String(error)}`);
       }
     }
     if (intent.lookupRequests?.length) {
@@ -4226,36 +4466,45 @@ function createApplicationPipeline(config) {
         const targetFields = state.fields.filter(
           (f) => intent.lookupRequests.some((lr) => lr.targetFieldIds.includes(f.id))
         );
-        const { result: lookupResult, usage: lookupUsage } = await fillFromLookup(
-          intent.lookupRequests,
-          targetFields,
-          availableData,
-          generateObject,
-          providerOptions
-        );
-        trackUsage(lookupUsage);
-        for (const fill of lookupResult.fills) {
-          const field = state.fields.find((f) => f.id === fill.fieldId);
-          if (field) {
-            field.value = fill.value;
-            field.source = `lookup: ${fill.source}`;
-            field.confidence = "high";
-            fieldsFilled++;
+        try {
+          const { result: lookupResult, usage: lookupUsage } = await fillFromLookup(
+            intent.lookupRequests,
+            targetFields,
+            availableData,
+            generateObject,
+            providerOptions
+          );
+          trackUsage(lookupUsage);
+          for (const fill of lookupResult.fills) {
+            const field = state.fields.find((f) => f.id === fill.fieldId);
+            if (field) {
+              field.value = fill.value;
+              field.source = `lookup: ${fill.source}`;
+              field.confidence = "high";
+              fieldsFilled++;
+            }
           }
+        } catch (error) {
+          await log?.(`Lookup fill failed: ${error instanceof Error ? error.message : String(error)}`);
         }
       }
     }
     if (intent.primaryIntent === "question" || intent.primaryIntent === "mixed") {
       if (intent.questionText) {
-        const { text, usage } = await generateText({
-          prompt: `The user is filling out an insurance application and asked: "${intent.questionText}"
+        try {
+          const { text, usage } = await generateText({
+            prompt: `The user is filling out an insurance application and asked: "${intent.questionText}"
 
 Provide a brief, helpful explanation (2-3 sentences). End with "Just reply with the answer when you're ready and I'll fill it in."`,
-          maxTokens: 512,
-          providerOptions
-        });
-        trackUsage(usage);
-        responseText = text;
+            maxTokens: 512,
+            providerOptions
+          });
+          trackUsage(usage);
+          responseText = text;
+        } catch (error) {
+          await log?.(`Question response generation failed: ${error instanceof Error ? error.message : String(error)}`);
+          responseText = `I wasn't able to generate an explanation for your question. Could you rephrase it, or just provide the answer directly?`;
+        }
       }
     }
     const currentBatchComplete = currentBatchFieldIds.every(
@@ -4269,26 +4518,30 @@ Provide a brief, helpful explanation (2-3 sentences). End with "Just reply with
           (f) => nextBatchFieldIds.includes(f.id)
         );
         const filledCount = state.fields.filter((f) => f.value).length;
-        const { text: emailText, usage: emailUsage } = await generateBatchEmail(
-          nextBatchFields,
-          state.currentBatchIndex,
-          state.batches.length,
-          {
-            appTitle: state.title,
-            totalFieldCount: state.fields.length,
-            filledFieldCount: filledCount,
-            companyName: context?.companyName
-          },
-          generateText,
-          providerOptions
-        );
-        trackUsage(emailUsage);
-        if (!responseText) {
-          responseText = emailText;
-        } else {
-          responseText += `
+        try {
+          const { text: emailText, usage: emailUsage } = await generateBatchEmail(
+            nextBatchFields,
+            state.currentBatchIndex,
+            state.batches.length,
+            {
+              appTitle: state.title,
+              totalFieldCount: state.fields.length,
+              filledFieldCount: filledCount,
+              companyName: context?.companyName
+            },
+            generateText,
+            providerOptions
+          );
+          trackUsage(emailUsage);
+          if (!responseText) {
+            responseText = emailText;
+          } else {
+            responseText += `
 
 ${emailText}`;
+          }
+        } catch (error) {
+          await log?.(`Batch email generation failed: ${error instanceof Error ? error.message : String(error)}`);
         }
       } else {
         state.status = "confirming";
@@ -4497,7 +4750,7 @@ var EvidenceItemSchema = z32.object({
   turnId: z32.string().optional(),
   text: z32.string().describe("Text excerpt from the source"),
   relevance: z32.number().min(0).max(1),
-  metadata: z32.record(z32.string(), z32.string()).optional()
+  metadata: z32.array(z32.object({ key: z32.string(), value: z32.string() })).optional()
 });
 var RetrievalResultSchema = z32.object({
   subQuestion: z32.string(),
@@ -4533,6 +4786,9 @@ var QueryResultSchema = z32.object({
 });
 
 // src/query/retriever.ts
+function recordToKVArray(record) {
+  return Object.entries(record).map(([key, value]) => ({ key, value }));
+}
 async function retrieve(subQuestion, conversationId, config) {
   const { documentStore, memoryStore, retrievalLimit, log } = config;
   const evidence = [];
@@ -4559,7 +4815,7 @@ async function retrieve(subQuestion, conversationId, config) {
                 text: chunk.text,
                 relevance: 0.8,
                 // Default — store doesn't expose scores directly
-                metadata: chunk.metadata
+                metadata: recordToKVArray(chunk.metadata)
               });
             }
           }
@@ -4574,7 +4830,7 @@ async function retrieve(subQuestion, conversationId, config) {
               documentId: chunk.documentId,
               text: chunk.text,
               relevance: 0.8,
-              metadata: chunk.metadata
+              metadata: recordToKVArray(chunk.metadata)
             });
           }
         }
@@ -4602,11 +4858,11 @@ async function retrieve(subQuestion, conversationId, config) {
               text: summary,
               relevance: 0.9,
               // Direct lookup is high relevance
-              metadata: {
-                type: doc.type,
-                carrier: doc.carrier ?? "",
-                insuredName: doc.insuredName ?? ""
-              }
+              metadata: [
+                { key: "type", value: doc.type },
+                { key: "carrier", value: doc.carrier ?? "" },
+                { key: "insuredName", value: doc.insuredName ?? "" }
+              ]
             });
           }
         } catch (e) {
@@ -4841,8 +5097,12 @@ function createQueryAgent(config) {
   async function query(input) {
     totalUsage = { inputTokens: 0, outputTokens: 0 };
     const { question, conversationId, context } = input;
+    const pipelineCtx = createPipelineContext({
+      id: `query-${Date.now()}`
+    });
     onProgress?.("Classifying query...");
     const classification = await classify(question, conversationId);
+    await pipelineCtx.save("classify", { classification });
     onProgress?.(`Retrieving evidence for ${classification.subQuestions.length} sub-question(s)...`);
     const retrieverConfig = {
       documentStore,
@@ -4856,9 +5116,10 @@ function createQueryAgent(config) {
       )
     );
     const allEvidence = retrievalResults.flatMap((r) => r.evidence);
+    await pipelineCtx.save("retrieve", { classification, evidence: allEvidence });
     onProgress?.("Reasoning over evidence...");
     const reasonerConfig = { generateObject, providerOptions };
-    let subAnswers = await Promise.all(
+    const reasonResults = await Promise.allSettled(
       classification.subQuestions.map(
         (sq, i) => limit(async () => {
           const { subAnswer, usage } = await reason(
@@ -4872,10 +5133,27 @@ function createQueryAgent(config) {
         })
       )
     );
+    let subAnswers = [];
+    for (let i = 0; i < reasonResults.length; i++) {
+      const result = reasonResults[i];
+      if (result.status === "fulfilled") {
+        subAnswers.push(result.value);
+      } else {
+        await log?.(`Reasoner failed for sub-question "${classification.subQuestions[i].question}": ${result.reason}`);
+        subAnswers.push({
+          subQuestion: classification.subQuestions[i].question,
+          answer: "Unable to answer this part of the question due to a processing error.",
+          citations: [],
+          confidence: 0,
+          needsMoreContext: true
+        });
+      }
+    }
+    await pipelineCtx.save("reason", { classification, evidence: allEvidence, subAnswers });
     onProgress?.("Verifying answer grounding...");
     const verifierConfig = { generateObject, providerOptions };
     for (let round = 0; round < maxVerifyRounds; round++) {
-      const { result: verifyResult, usage } = await verify(
+      const { result: verifyResult, usage } = await safeVerify(
         question,
         subAnswers,
         allEvidence,
@@ -4899,7 +5177,6 @@ function createQueryAgent(config) {
                 () => retrieve(sq, conversationId, {
                   ...retrieverConfig,
                   retrievalLimit: retrievalLimit * 2
-                  // Broader retrieval on retry
                 })
               )
             )
@@ -4907,7 +5184,7 @@ function createQueryAgent(config) {
           for (const r of retryRetrievals) {
             allEvidence.push(...r.evidence);
           }
-          const retrySubAnswers = await Promise.all(
+          const retrySettled = await Promise.allSettled(
             retryQuestions.map(
               (sq, i) => limit(async () => {
                 const { subAnswer, usage: u } = await reason(
@@ -4921,6 +5198,7 @@ function createQueryAgent(config) {
               })
             )
           );
+          const retrySubAnswers = retrySettled.filter((r) => r.status === "fulfilled").map((r) => r.value);
           const retryQSet = new Set(retryQuestions.map((sq) => sq.question));
           subAnswers = subAnswers.map((sa) => {
             if (retryQSet.has(sa.subQuestion)) {
@@ -4973,17 +5251,42 @@ function createQueryAgent(config) {
       }
     }
     const prompt = buildQueryClassifyPrompt(question, conversationContext);
-    const { object, usage } = await withRetry(
-      () => generateObject({
+    const { object, usage } = await safeGenerateObject(
+      generateObject,
+      {
         prompt,
         schema: QueryClassifyResultSchema,
         maxTokens: 2048,
         providerOptions
-      })
+      },
+      {
+        fallback: {
+          intent: "general_knowledge",
+          subQuestions: [
+            {
+              question,
+              intent: "general_knowledge"
+            }
+          ],
+          requiresDocumentLookup: true,
+          requiresChunkSearch: true,
+          requiresConversationHistory: !!conversationId
+        },
+        log,
+        onError: (err, attempt) => log?.(`Query classify attempt ${attempt + 1} failed: ${err}`)
+      }
     );
     trackUsage(usage);
     return object;
   }
+  async function safeVerify(originalQuestion, subAnswers, allEvidence, verifierConfig) {
+    try {
+      return await verify(originalQuestion, subAnswers, allEvidence, verifierConfig);
+    } catch (error) {
+      await log?.(`Verification failed, approving by default: ${error instanceof Error ? error.message : String(error)}`);
+      return { result: { approved: true, issues: [] } };
+    }
+  }
   async function respond(originalQuestion, subAnswers, classification, platform) {
     const subAnswersJson = JSON.stringify(
       subAnswers.map((sa) => ({
@@ -4997,13 +5300,25 @@ function createQueryAgent(config) {
       2
     );
     const prompt = buildRespondPrompt(originalQuestion, subAnswersJson, platform);
-    const { object, usage } = await withRetry(
-      () => generateObject({
+    const { object, usage } = await safeGenerateObject(
+      generateObject,
+      {
         prompt,
         schema: QueryResultSchema,
         maxTokens: 4096,
         providerOptions
-      })
+      },
+      {
+        fallback: {
+          answer: subAnswers.map((sa) => `**${sa.subQuestion}**
+${sa.answer}`).join("\n\n"),
+          citations: subAnswers.flatMap((sa) => sa.citations),
+          intent: classification.intent,
+          confidence: Math.min(...subAnswers.map((sa) => sa.confidence), 1)
+        },
+        log,
+        onError: (err, attempt) => log?.(`Respond attempt ${attempt + 1} failed: ${err}`)
+      }
     );
     trackUsage(usage);
     const result = object;
@@ -5168,6 +5483,7 @@ export {
   CommercialAutoDeclarationsSchema,
   CommercialPropertyDeclarationsSchema,
   CommunicationIntentSchema,
+  ConditionKeyValueSchema,
   ConditionTypeSchema,
   ConstructionTypeSchema,
   ContactSchema,
@@ -5334,6 +5650,7 @@ export {
   chunkDocument,
   createApplicationPipeline,
   createExtractor,
+  createPipelineContext,
   createQueryAgent,
   extractPageRange,
   fillAcroForm,
@@ -5343,6 +5660,7 @@ export {
   getTemplate,
   overlayTextOnPdf,
   pLimit,
+  safeGenerateObject,
   sanitizeNulls,
   stripFences,
   withRetry