@claritylabs/cl-sdk 1.1.2 → 1.2.0

package/dist/index.mjs

@@ -29779,6 +29779,132 @@ var AGENT_TOOLS = [
 
 // src/extraction/pipeline.ts
 import { generateText } from "ai";
+
+// src/extraction/pdf.ts
+import {
+  PDFDocument,
+  PDFTextField,
+  PDFCheckBox,
+  PDFDropdown,
+  PDFRadioGroup,
+  StandardFonts,
+  rgb
+} from "pdf-lib";
+async function extractPageRange(pdfBase64, startPage, endPage) {
+  const srcBytes = typeof Buffer !== "undefined" ? Buffer.from(pdfBase64, "base64") : Uint8Array.from(atob(pdfBase64), (c) => c.charCodeAt(0));
+  const srcDoc = await PDFDocument.load(srcBytes, { ignoreEncryption: true });
+  const totalPages = srcDoc.getPageCount();
+  const start = Math.max(startPage - 1, 0);
+  const end = Math.min(endPage, totalPages) - 1;
+  if (start === 0 && end >= totalPages - 1) {
+    return pdfBase64;
+  }
+  const newDoc = await PDFDocument.create();
+  const indices = Array.from({ length: end - start + 1 }, (_, i) => start + i);
+  const pages = await newDoc.copyPages(srcDoc, indices);
+  pages.forEach((page) => newDoc.addPage(page));
+  const bytes = await newDoc.save();
+  if (typeof Buffer !== "undefined") {
+    return Buffer.from(bytes).toString("base64");
+  }
+  let binary = "";
+  const uint8 = new Uint8Array(bytes);
+  for (let i = 0; i < uint8.length; i++) {
+    binary += String.fromCharCode(uint8[i]);
+  }
+  return btoa(binary);
+}
+async function getPdfPageCount(pdfBase64) {
+  const srcBytes = typeof Buffer !== "undefined" ? Buffer.from(pdfBase64, "base64") : Uint8Array.from(atob(pdfBase64), (c) => c.charCodeAt(0));
+  const doc = await PDFDocument.load(srcBytes, { ignoreEncryption: true });
+  return doc.getPageCount();
+}
+function getAcroFormFields(pdfDoc) {
+  const form = pdfDoc.getForm();
+  const fields = form.getFields();
+  if (fields.length === 0) return [];
+  return fields.map((field) => {
+    const name = field.getName();
+    if (field instanceof PDFTextField) {
+      return { name, type: "text" };
+    }
+    if (field instanceof PDFCheckBox) {
+      return { name, type: "checkbox" };
+    }
+    if (field instanceof PDFDropdown) {
+      return { name, type: "dropdown", options: field.getOptions() };
+    }
+    if (field instanceof PDFRadioGroup) {
+      return { name, type: "radio", options: field.getOptions() };
+    }
+    return { name, type: "text" };
+  });
+}
+async function fillAcroForm(pdfBytes, mappings) {
+  const pdfDoc = await PDFDocument.load(pdfBytes, { ignoreEncryption: true });
+  const form = pdfDoc.getForm();
+  for (const { acroFormName, value } of mappings) {
+    try {
+      const field = form.getField(acroFormName);
+      if (field instanceof PDFTextField) {
+        field.setText(value);
+      } else if (field instanceof PDFCheckBox) {
+        const lower = value.toLowerCase();
+        if (["yes", "true", "x", "checked", "on"].includes(lower)) {
+          field.check();
+        } else {
+          field.uncheck();
+        }
+      } else if (field instanceof PDFDropdown) {
+        try {
+          field.select(value);
+        } catch {
+        }
+      } else if (field instanceof PDFRadioGroup) {
+        try {
+          field.select(value);
+        } catch {
+        }
+      }
+    } catch {
+    }
+  }
+  form.flatten();
+  return await pdfDoc.save();
+}
+async function overlayTextOnPdf(pdfBytes, overlays) {
+  const pdfDoc = await PDFDocument.load(pdfBytes, { ignoreEncryption: true });
+  const font = await pdfDoc.embedFont(StandardFonts.Helvetica);
+  const pageCount = pdfDoc.getPageCount();
+  for (const overlay of overlays) {
+    if (overlay.page < 0 || overlay.page >= pageCount) continue;
+    const page = pdfDoc.getPage(overlay.page);
+    const { width, height } = page.getSize();
+    const fontSize = overlay.fontSize ?? 10;
+    const x = overlay.x / 100 * width;
+    const y = height - overlay.y / 100 * height - fontSize;
+    if (overlay.isCheckmark) {
+      page.drawText("X", {
+        x,
+        y,
+        size: fontSize,
+        font,
+        color: rgb(0, 0, 0)
+      });
+    } else {
+      page.drawText(overlay.text, {
+        x,
+        y,
+        size: fontSize,
+        font,
+        color: rgb(0, 0, 0)
+      });
+    }
+  }
+  return await pdfDoc.save();
+}
+
+// src/extraction/pipeline.ts
 var SONNET_MODEL = "claude-sonnet-4-6";
 var HAIKU_MODEL = "claude-haiku-4-5-20251001";
 var DEFAULT_METADATA_PROVIDER_OPTIONS = {
@@ -29915,8 +30041,10 @@ function getPageChunks(totalPages, chunkSize = 30) {
   }
   return chunks;
 }
-async function callModel(model, pdfBase64, prompt, maxTokens, providerOptions, log, onTokenUsage) {
-  await log?.(`Calling model (max ${maxTokens} tokens)...`);
+async function callModel(model, pdfBase64, prompt, maxTokens, providerOptions, log, onTokenUsage, pageRange) {
+  const pdfToSend = pageRange ? await extractPageRange(pdfBase64, pageRange[0], pageRange[1]) : pdfBase64;
+  const rangeLabel = pageRange ? ` [pages ${pageRange[0]}\u2013${pageRange[1]}]` : "";
+  await log?.(`Calling model (max ${maxTokens} tokens)${rangeLabel}...`);
   const start = Date.now();
   const { text, usage } = await withRetry(
     () => generateText({
@@ -29925,7 +30053,7 @@ async function callModel(model, pdfBase64, prompt, maxTokens, providerOptions, l
       messages: [{
         role: "user",
         content: [
-          { type: "file", data: pdfBase64, mediaType: "application/pdf" },
+          { type: "file", data: pdfToSend, mediaType: "application/pdf" },
           { type: "text", text: prompt }
         ]
       }],
@@ -30031,7 +30159,9 @@ async function classifyDocumentType(pdfBase64, options) {
     MODEL_TOKEN_LIMITS.classification,
     void 0,
     log,
-    onTokenUsage
+    onTokenUsage,
+    [1, 3]
+    // Only need first 3 pages for classification
   );
   try {
     const parsed = JSON.parse(stripFences(raw));
@@ -30120,7 +30250,9 @@ async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, p
     MODEL_TOKEN_LIMITS.sections,
     void 0,
     log,
-    onTokenUsage
+    onTokenUsage,
+    [start, end]
+    // Only send this chunk's pages
   );
   try {
     return [JSON.parse(stripFences(chunkRaw))];
@@ -30162,7 +30294,9 @@ async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, p
       MODEL_TOKEN_LIMITS.sectionsFallback,
       fallbackProviderOptions,
       log,
-      onTokenUsage
+      onTokenUsage,
+      [start, end]
+      // Only send this chunk's pages
     );
     try {
       return [JSON.parse(stripFences(fallbackRaw))];
@@ -30205,7 +30339,9 @@ async function extractFromPdf(pdfBase64, options) {
     onTokenUsage
   } = options ?? {};
   const resolved = resolveModels(models);
+  const actualPageCount = await getPdfPageCount(pdfBase64);
   await log?.("Pass 1: Extracting metadata...");
+  const metadataPageRange = [1, Math.min(10, actualPageCount)];
   const metadataRaw = await callModel(
     resolved.metadata,
     pdfBase64,
@@ -30213,7 +30349,8 @@ async function extractFromPdf(pdfBase64, options) {
     MODEL_TOKEN_LIMITS.metadata,
     metadataProviderOptions,
     log,
-    onTokenUsage
+    onTokenUsage,
+    metadataPageRange
   );
   let metadataResult;
   try {
@@ -30224,7 +30361,7 @@ async function extractFromPdf(pdfBase64, options) {
     throw new Error(`Metadata JSON parse failed: ${e.message}`);
   }
   await onMetadata?.(metadataRaw);
-  const pageCount = metadataResult.totalPages || 1;
+  const pageCount = actualPageCount;
   await log?.(`Document: ${pageCount} page(s)`);
   const sectionChunks = await extractSectionChunks(
     resolved,
@@ -30292,7 +30429,9 @@ async function extractQuoteFromPdf(pdfBase64, options) {
     onTokenUsage
   } = options ?? {};
   const resolved = resolveModels(models);
+  const actualPageCount = await getPdfPageCount(pdfBase64);
   await log?.("Pass 1: Extracting quote metadata...");
+  const metadataPageRange = [1, Math.min(10, actualPageCount)];
   const metadataRaw = await callModel(
     resolved.metadata,
     pdfBase64,
@@ -30300,7 +30439,8 @@ async function extractQuoteFromPdf(pdfBase64, options) {
     MODEL_TOKEN_LIMITS.metadata,
     metadataProviderOptions,
     log,
-    onTokenUsage
+    onTokenUsage,
+    metadataPageRange
   );
   let metadataResult;
   try {
@@ -30311,7 +30451,7 @@ async function extractQuoteFromPdf(pdfBase64, options) {
     throw new Error(`Quote metadata JSON parse failed: ${e.message}`);
   }
   await onMetadata?.(metadataRaw);
-  const pageCount = metadataResult.totalPages || 1;
+  const pageCount = actualPageCount;
   await log?.(`Quote document: ${pageCount} page(s)`);
   const sectionChunks = await extractSectionChunks(
     resolved,
@@ -30328,101 +30468,6 @@ async function extractQuoteFromPdf(pdfBase64, options) {
   const mergedRaw = JSON.stringify(merged);
   return { rawText: mergedRaw, extracted: merged };
 }
-
-// src/extraction/pdf.ts
-import {
-  PDFDocument,
-  PDFTextField,
-  PDFCheckBox,
-  PDFDropdown,
-  PDFRadioGroup,
-  StandardFonts,
-  rgb
-} from "pdf-lib";
-function getAcroFormFields(pdfDoc) {
-  const form = pdfDoc.getForm();
-  const fields = form.getFields();
-  if (fields.length === 0) return [];
-  return fields.map((field) => {
-    const name = field.getName();
-    if (field instanceof PDFTextField) {
-      return { name, type: "text" };
-    }
-    if (field instanceof PDFCheckBox) {
-      return { name, type: "checkbox" };
-    }
-    if (field instanceof PDFDropdown) {
-      return { name, type: "dropdown", options: field.getOptions() };
-    }
-    if (field instanceof PDFRadioGroup) {
-      return { name, type: "radio", options: field.getOptions() };
-    }
-    return { name, type: "text" };
-  });
-}
-async function fillAcroForm(pdfBytes, mappings) {
-  const pdfDoc = await PDFDocument.load(pdfBytes, { ignoreEncryption: true });
-  const form = pdfDoc.getForm();
-  for (const { acroFormName, value } of mappings) {
-    try {
-      const field = form.getField(acroFormName);
-      if (field instanceof PDFTextField) {
-        field.setText(value);
-      } else if (field instanceof PDFCheckBox) {
-        const lower = value.toLowerCase();
-        if (["yes", "true", "x", "checked", "on"].includes(lower)) {
-          field.check();
-        } else {
-          field.uncheck();
-        }
-      } else if (field instanceof PDFDropdown) {
-        try {
-          field.select(value);
-        } catch {
-        }
-      } else if (field instanceof PDFRadioGroup) {
-        try {
-          field.select(value);
-        } catch {
-        }
-      }
-    } catch {
-    }
-  }
-  form.flatten();
-  return await pdfDoc.save();
-}
-async function overlayTextOnPdf(pdfBytes, overlays) {
-  const pdfDoc = await PDFDocument.load(pdfBytes, { ignoreEncryption: true });
-  const font = await pdfDoc.embedFont(StandardFonts.Helvetica);
-  const pageCount = pdfDoc.getPageCount();
-  for (const overlay of overlays) {
-    if (overlay.page < 0 || overlay.page >= pageCount) continue;
-    const page = pdfDoc.getPage(overlay.page);
-    const { width, height } = page.getSize();
-    const fontSize = overlay.fontSize ?? 10;
-    const x = overlay.x / 100 * width;
-    const y = height - overlay.y / 100 * height - fontSize;
-    if (overlay.isCheckmark) {
-      page.drawText("X", {
-        x,
-        y,
-        size: fontSize,
-        font,
-        color: rgb(0, 0, 0)
-      });
-    } else {
-      page.drawText(overlay.text, {
-        x,
-        y,
-        size: fontSize,
-        font,
-        color: rgb(0, 0, 0)
-      });
-    }
-  }
-  return await pdfDoc.save();
-}
 export {
   AGENT_TOOLS,
   APPLICATION_CLASSIFY_PROMPT,
@@ -30474,11 +30519,13 @@ export {
   createUniformModelConfig,
   enrichSupplementaryFields,
   extractFromPdf,
+  extractPageRange,
   extractQuoteFromPdf,
   extractSectionsOnly,
   fillAcroForm,
   getAcroFormFields,
   getPageChunks,
+  getPdfPageCount,
   mergeChunkedQuoteSections,
   mergeChunkedSections,
   overlayTextOnPdf,