@heripo/pdf-parser 0.1.48 → 0.2.0

package/dist/index.js

@@ -73,6 +73,8 @@ import {
   hasToolCall,
   tool
 } from "ai";
+import { wrapLanguageModel } from "ai";
+import { wrapLanguageModel as wrapLanguageModel2 } from "ai";
 var ConcurrentPool = class {
   /**
    * Process items concurrently using a worker pool pattern.
@@ -187,9 +189,23 @@ var LLMCaller = class {
         cacheWriteTokens: allWrites - (write1h ?? 0)
       },
       ...write1h !== void 0 && { cacheWrite1hTokens: write1h },
+      ...usage?.outputTokenDetails?.reasoningTokens != null && {
+        reasoningTokens: usage.outputTokenDetails.reasoningTokens
+      },
       metadata: config.metadata
     };
   }
+  /** Finish reason and structured-output attempts, when known. */
+  static callDetails(response) {
+    return {
+      ...response.finishReason !== void 0 && {
+        finishReason: response.finishReason
+      },
+      ...response.structuredAttempts !== void 0 && {
+        structuredAttempts: response.structuredAttempts
+      }
+    };
+  }
   /**
    * Maximum number of retries when structured output generation fails.
    * Total attempts = MAX_STRUCTURED_OUTPUT_RETRIES + 1.
@@ -230,7 +246,9 @@ var LLMCaller = class {
         return {
           output: toolCall.input,
           usage: lastResult.usage,
-          steps: lastResult.steps
+          steps: lastResult.steps,
+          finishReason: lastResult.finishReason,
+          structuredAttempts: attempt + 1
         };
       }
     }
@@ -264,11 +282,18 @@ var LLMCaller = class {
     let lastError;
     for (let attempt = 0; attempt <= this.MAX_STRUCTURED_OUTPUT_RETRIES; attempt++) {
       try {
-        return await generateText({
+        const result = await generateText({
           model,
           output: Output.object({ schema }),
           ...promptParams
         });
+        return {
+          output: result.output,
+          usage: result.usage,
+          steps: result.steps,
+          finishReason: result.finishReason,
+          structuredAttempts: attempt + 1
+        };
       } catch (error) {
         if (NoObjectGeneratedError.isInstance(error)) {
           lastError = error;
@@ -292,7 +317,8 @@ var LLMCaller = class {
       return {
         output: response.output,
         usage: this.buildUsage(config, primaryModelName, response, false),
-        usedFallback: false
+        usedFallback: false,
+        ...this.callDetails(response)
       };
     } catch (primaryError) {
       if (config.abortSignal?.aborted) {
@@ -306,7 +332,8 @@ var LLMCaller = class {
       return {
         output: response.output,
         usage: this.buildUsage(config, fallbackModelName, response, true),
-        usedFallback: true
+        usedFallback: true,
+        ...this.callDetails(response)
       };
     }
   }
@@ -329,8 +356,17 @@ var LLMCaller = class {
     return this.executeWithFallback(
       config,
       (model) => this.generateStructuredOutput(model, config.schema, {
-        instructions: config.systemPrompt,
+        instructions: config.cacheSystemPrompt ? {
+          role: "system",
+          content: config.systemPrompt,
+          providerOptions: {
+            anthropic: { cacheControl: { type: "ephemeral" } }
+          }
+        } : config.systemPrompt,
         prompt: config.userPrompt,
+        ...config.providerOptions && {
+          providerOptions: config.providerOptions
+        },
         temperature: config.temperature,
         maxRetries: config.maxRetries,
         abortSignal: config.abortSignal
@@ -600,6 +636,7 @@ var LLMTokenUsageAggregator = class {
     this.usage = {};
   }
 };
+var encoder = new TextEncoder();
 
 // src/environment/docling-server.ts
 import { spawn as spawn2 } from "child_process";