@claritylabs/cl-sdk 1.0.0 → 1.1.2

package/dist/index.mjs

@@ -29787,6 +29787,56 @@ var DEFAULT_METADATA_PROVIDER_OPTIONS = {
 var DEFAULT_FALLBACK_PROVIDER_OPTIONS = {
   anthropic: { thinking: { type: "enabled", budgetTokens: 4096 } }
 };
+var MAX_RETRIES = 5;
+var BASE_DELAY_MS = 2e3;
+function isRateLimitError(error) {
+  if (error instanceof Error) {
+    const msg = error.message.toLowerCase();
+    if (msg.includes("rate limit") || msg.includes("rate_limit") || msg.includes("too many requests")) {
+      return true;
+    }
+  }
+  if (typeof error === "object" && error !== null) {
+    const status = error.status ?? error.statusCode;
+    if (status === 429) return true;
+  }
+  return false;
+}
+async function withRetry(fn, log) {
+  for (let attempt = 0; ; attempt++) {
+    try {
+      return await fn();
+    } catch (error) {
+      if (!isRateLimitError(error) || attempt >= MAX_RETRIES) {
+        throw error;
+      }
+      const jitter = Math.random() * 1e3;
+      const delay = BASE_DELAY_MS * Math.pow(2, attempt) + jitter;
+      await log?.(`Rate limited, retrying in ${(delay / 1e3).toFixed(1)}s (attempt ${attempt + 1}/${MAX_RETRIES})...`);
+      await new Promise((resolve) => setTimeout(resolve, delay));
+    }
+  }
+}
+function pLimit(concurrency) {
+  let active = 0;
+  const queue = [];
+  function next() {
+    if (queue.length > 0 && active < concurrency) {
+      active++;
+      queue.shift()();
+    }
+  }
+  return (fn) => new Promise((resolve, reject) => {
+    const run = () => {
+      fn().then(resolve, reject).finally(() => {
+        active--;
+        next();
+      });
+    };
+    queue.push(run);
+    next();
+  });
+}
 function stripFences(text) {
   return text.replace(/^```(?:json)?\s*\n?/i, "").replace(/\n?```\s*$/i, "");
 }
@@ -29865,48 +29915,56 @@ function getPageChunks(totalPages, chunkSize = 30) {
   }
   return chunks;
 }
-async function callModel(model, pdfBase64, prompt, maxTokens, providerOptions, log) {
+async function callModel(model, pdfBase64, prompt, maxTokens, providerOptions, log, onTokenUsage) {
   await log?.(`Calling model (max ${maxTokens} tokens)...`);
   const start = Date.now();
-  const { text, usage } = await generateText({
-    model,
-    maxOutputTokens: maxTokens,
-    messages: [{
-      role: "user",
-      content: [
-        { type: "file", data: pdfBase64, mediaType: "application/pdf" },
-        { type: "text", text: prompt }
-      ]
-    }],
-    ...providerOptions ? { providerOptions } : {}
-  });
+  const { text, usage } = await withRetry(
+    () => generateText({
+      model,
+      maxOutputTokens: maxTokens,
+      messages: [{
+        role: "user",
+        content: [
+          { type: "file", data: pdfBase64, mediaType: "application/pdf" },
+          { type: "text", text: prompt }
+        ]
+      }],
+      ...providerOptions ? { providerOptions } : {}
+    }),
+    log
+  );
   const elapsed = ((Date.now() - start) / 1e3).toFixed(1);
   const inputTokens = usage?.inputTokens ?? 0;
   const outputTokens = usage?.outputTokens ?? 0;
   await log?.(`${inputTokens} in / ${outputTokens} out tokens (${elapsed}s)`);
+  onTokenUsage?.({ inputTokens, outputTokens });
   return text || "{}";
 }
-async function callModelText(model, prompt, maxTokens, log) {
+async function callModelText(model, prompt, maxTokens, log, onTokenUsage) {
   await log?.(`Calling model text-only (max ${maxTokens} tokens)...`);
   const start = Date.now();
-  const { text, usage } = await generateText({
-    model,
-    maxOutputTokens: maxTokens,
-    messages: [{
-      role: "user",
-      content: prompt
-    }]
-  });
+  const { text, usage } = await withRetry(
+    () => generateText({
+      model,
+      maxOutputTokens: maxTokens,
+      messages: [{
+        role: "user",
+        content: prompt
+      }]
+    }),
+    log
+  );
   const elapsed = ((Date.now() - start) / 1e3).toFixed(1);
   const inputTokens = usage?.inputTokens ?? 0;
   const outputTokens = usage?.outputTokens ?? 0;
   await log?.(`text: ${inputTokens} in / ${outputTokens} out tokens (${elapsed}s)`);
+  onTokenUsage?.({ inputTokens, outputTokens });
   return text || "{}";
 }
 function resolveModels(models) {
   return models ?? createDefaultModelConfig();
 }
-async function enrichSupplementaryFields(document, models, log) {
+async function enrichSupplementaryFields(document, models, log, onTokenUsage) {
   const fields = {};
   if (document.regulatoryContext?.content) {
     fields.regulatoryContext = document.regulatoryContext.content;
@@ -29928,7 +29986,7 @@ async function enrichSupplementaryFields(document, models, log) {
   try {
     const resolved = resolveModels(models);
     const prompt = buildSupplementaryEnrichmentPrompt(fields);
-    const raw = await callModelText(resolved.enrichment, prompt, MODEL_TOKEN_LIMITS.enrichment, log);
+    const raw = await callModelText(resolved.enrichment, prompt, MODEL_TOKEN_LIMITS.enrichment, log, onTokenUsage);
     const parsed = JSON.parse(stripFences(raw));
     const enriched = { ...document };
     if (parsed.regulatoryContext && enriched.regulatoryContext) {
@@ -29963,7 +30021,7 @@ async function enrichSupplementaryFields(document, models, log) {
   }
 }
 async function classifyDocumentType(pdfBase64, options) {
-  const { log, models } = options ?? {};
+  const { log, models, onTokenUsage } = options ?? {};
   const resolved = resolveModels(models);
   await log?.("Pass 0: Classifying document type...");
   const raw = await callModel(
@@ -29972,7 +30030,8 @@ async function classifyDocumentType(pdfBase64, options) {
     CLASSIFY_DOCUMENT_PROMPT,
     MODEL_TOKEN_LIMITS.classification,
     void 0,
-    log
+    log,
+    onTokenUsage
   );
   try {
     const parsed = JSON.parse(stripFences(raw));
@@ -30052,7 +30111,7 @@ function mergeChunkedQuoteSections(metadataResult, sectionChunks) {
   };
 }
 var CHUNK_SIZES = [15, 10, 5];
-async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, promptBuilder, fallbackProviderOptions, log) {
+async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, promptBuilder, fallbackProviderOptions, log, onTokenUsage, concurrency = 2) {
   await log?.(`Pass 2: Extracting sections pages ${start}\u2013${end}...`);
   const chunkRaw = await callModel(
     models.sections,
@@ -30060,7 +30119,8 @@ async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, p
     promptBuilder(start, end),
     MODEL_TOKEN_LIMITS.sections,
     void 0,
-    log
+    log,
+    onTokenUsage
   );
   try {
     return [JSON.parse(stripFences(chunkRaw))];
@@ -30074,21 +30134,24 @@ async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, p
         const subChunks = getPageChunks(pageSpan, smallerSize).map(
           ([s, e]) => [s + start - 1, e + start - 1]
         );
-        const results = [];
-        for (const [subStart, subEnd] of subChunks) {
-          const subResults = await extractChunkWithRetry(
-            models,
-            pdfBase64,
-            subStart,
-            subEnd,
-            nextSizeIndex,
-            promptBuilder,
-            fallbackProviderOptions,
-            log
-          );
-          results.push(...subResults);
-        }
-        return results;
+        const limit = pLimit(concurrency);
+        const nestedResults = await Promise.all(
+          subChunks.map(
+            ([subStart, subEnd]) => limit(() => extractChunkWithRetry(
+              models,
+              pdfBase64,
+              subStart,
+              subEnd,
+              nextSizeIndex,
+              promptBuilder,
+              fallbackProviderOptions,
+              log,
+              onTokenUsage,
+              concurrency
+            ))
+          )
+        );
+        return nestedResults.flat();
       }
     }
     await log?.(`Sections model exhausted for pages ${start}\u2013${end}, falling back...`);
@@ -30098,7 +30161,8 @@ async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, p
       promptBuilder(start, end),
       MODEL_TOKEN_LIMITS.sectionsFallback,
       fallbackProviderOptions,
-      log
+      log,
+      onTokenUsage
     );
     try {
       return [JSON.parse(stripFences(fallbackRaw))];
@@ -30109,23 +30173,26 @@ async function extractChunkWithRetry(models, pdfBase64, start, end, sizeIndex, p
     }
   }
 }
-async function extractSectionChunks(models, pdfBase64, pageCount, promptBuilder = buildSectionsPrompt, fallbackProviderOptions, log) {
+async function extractSectionChunks(models, pdfBase64, pageCount, promptBuilder = buildSectionsPrompt, fallbackProviderOptions, log, onTokenUsage, concurrency = 2) {
   const chunks = getPageChunks(pageCount, CHUNK_SIZES[0]);
-  const sectionChunks = [];
-  for (const [start, end] of chunks) {
-    const results = await extractChunkWithRetry(
-      models,
-      pdfBase64,
-      start,
-      end,
-      0,
-      promptBuilder,
-      fallbackProviderOptions,
-      log
-    );
-    sectionChunks.push(...results);
-  }
-  return sectionChunks;
+  const limit = pLimit(concurrency);
+  const nestedResults = await Promise.all(
+    chunks.map(
+      ([start, end]) => limit(() => extractChunkWithRetry(
+        models,
+        pdfBase64,
+        start,
+        end,
+        0,
+        promptBuilder,
+        fallbackProviderOptions,
+        log,
+        onTokenUsage,
+        concurrency
+      ))
+    )
+  );
+  return nestedResults.flat();
 }
 async function extractFromPdf(pdfBase64, options) {
   const {
@@ -30133,7 +30200,9 @@ async function extractFromPdf(pdfBase64, options) {
     onMetadata,
     models,
     metadataProviderOptions = DEFAULT_METADATA_PROVIDER_OPTIONS,
-    fallbackProviderOptions = DEFAULT_FALLBACK_PROVIDER_OPTIONS
+    fallbackProviderOptions = DEFAULT_FALLBACK_PROVIDER_OPTIONS,
+    concurrency = 2,
+    onTokenUsage
   } = options ?? {};
   const resolved = resolveModels(models);
   await log?.("Pass 1: Extracting metadata...");
@@ -30143,7 +30212,8 @@ async function extractFromPdf(pdfBase64, options) {
     METADATA_PROMPT,
     MODEL_TOKEN_LIMITS.metadata,
     metadataProviderOptions,
-    log
+    log,
+    onTokenUsage
   );
   let metadataResult;
   try {
@@ -30162,12 +30232,14 @@ async function extractFromPdf(pdfBase64, options) {
     pageCount,
     buildSectionsPrompt,
     fallbackProviderOptions,
-    log
+    log,
+    onTokenUsage,
+    concurrency
   );
   await log?.("Merging extraction results...");
   const merged = mergeChunkedSections(metadataResult, sectionChunks);
   if (merged.document) {
-    merged.document = await enrichSupplementaryFields(merged.document, resolved, log);
+    merged.document = await enrichSupplementaryFields(merged.document, resolved, log, onTokenUsage);
   }
   const mergedRaw = JSON.stringify(merged);
   return { rawText: mergedRaw, extracted: merged };
@@ -30177,7 +30249,9 @@ async function extractSectionsOnly(pdfBase64, metadataRaw, options) {
     log,
     promptBuilder = buildSectionsPrompt,
     models,
-    fallbackProviderOptions = DEFAULT_FALLBACK_PROVIDER_OPTIONS
+    fallbackProviderOptions = DEFAULT_FALLBACK_PROVIDER_OPTIONS,
+    concurrency = 2,
+    onTokenUsage
   } = options ?? {};
   const resolved = resolveModels(models);
   await log?.("Using saved metadata, skipping pass 1...");
@@ -30195,12 +30269,14 @@ async function extractSectionsOnly(pdfBase64, metadataRaw, options) {
     pageCount,
     promptBuilder,
     fallbackProviderOptions,
-    log
+    log,
+    onTokenUsage,
+    concurrency
   );
   await log?.("Merging extraction results...");
   const merged = mergeChunkedSections(metadataResult, sectionChunks);
   if (merged.document) {
-    merged.document = await enrichSupplementaryFields(merged.document, resolved, log);
+    merged.document = await enrichSupplementaryFields(merged.document, resolved, log, onTokenUsage);
   }
   const mergedRaw = JSON.stringify(merged);
   return { rawText: mergedRaw, extracted: merged };
@@ -30211,7 +30287,9 @@ async function extractQuoteFromPdf(pdfBase64, options) {
     onMetadata,
     models,
     metadataProviderOptions = DEFAULT_METADATA_PROVIDER_OPTIONS,
-    fallbackProviderOptions = DEFAULT_FALLBACK_PROVIDER_OPTIONS
+    fallbackProviderOptions = DEFAULT_FALLBACK_PROVIDER_OPTIONS,
+    concurrency = 2,
+    onTokenUsage
   } = options ?? {};
   const resolved = resolveModels(models);
   await log?.("Pass 1: Extracting quote metadata...");
@@ -30221,7 +30299,8 @@ async function extractQuoteFromPdf(pdfBase64, options) {
     QUOTE_METADATA_PROMPT,
     MODEL_TOKEN_LIMITS.metadata,
     metadataProviderOptions,
-    log
+    log,
+    onTokenUsage
   );
   let metadataResult;
   try {
@@ -30240,7 +30319,9 @@ async function extractQuoteFromPdf(pdfBase64, options) {
     pageCount,
     buildQuoteSectionsPrompt,
     fallbackProviderOptions,
-    log
+    log,
+    onTokenUsage,
+    concurrency
   );
   await log?.("Merging quote extraction results...");
   const merged = mergeChunkedQuoteSections(metadataResult, sectionChunks);