@posthog/ai 7.1.0 → 7.1.1

package/dist/index.mjs

@@ -6,7 +6,7 @@ import { wrapLanguageModel } from 'ai';
 import AnthropicOriginal from '@anthropic-ai/sdk';
 import { GoogleGenAI } from '@google/genai';
 
-var version = "7.1.0";
+var version = "7.1.1";
 
 // Type guards for safer type checking
 const isString = value => {
@@ -3711,7 +3711,7 @@ class LangChainCallbackHandler extends BaseCallbackHandler {
       eventProperties['$ai_is_error'] = true;
     } else {
       // Handle token usage
-      const [inputTokens, outputTokens, additionalTokenData] = this.parseUsage(output);
+      const [inputTokens, outputTokens, additionalTokenData] = this.parseUsage(output, run.provider, run.model);
       eventProperties['$ai_input_tokens'] = inputTokens;
       eventProperties['$ai_output_tokens'] = outputTokens;
       // Add additional token data to properties
@@ -3858,7 +3858,7 @@ class LangChainCallbackHandler extends BaseCallbackHandler {
     // Sanitize the message content to redact base64 images
     return sanitizeLangChain(messageDict);
   }
-  _parseUsageModel(usage) {
+  _parseUsageModel(usage, provider, model) {
     const conversionList = [['promptTokens', 'input'], ['completionTokens', 'output'], ['input_tokens', 'input'], ['output_tokens', 'output'], ['prompt_token_count', 'input'], ['candidates_token_count', 'output'], ['inputTokenCount', 'input'], ['outputTokenCount', 'output'], ['input_token_count', 'input'], ['generated_token_count', 'output']];
     const parsedUsage = conversionList.reduce((acc, [modelKey, typeKey]) => {
       const value = usage[modelKey];
@@ -3925,20 +3925,28 @@ class LangChainCallbackHandler extends BaseCallbackHandler {
     if (webSearchCount !== undefined) {
       additionalTokenData.webSearchCount = webSearchCount;
     }
-    // In LangChain, input_tokens is the sum of input and cache read tokens.
-    // Our cost calculation expects them to be separate, for Anthropic.
-    if (parsedUsage.input && additionalTokenData.cacheReadInputTokens) {
+    // For Anthropic providers, LangChain reports input_tokens as the sum of input and cache read tokens.
+    // Our cost calculation expects them to be separate for Anthropic, so we subtract cache tokens.
+    // For other providers (OpenAI, etc.), input_tokens already excludes cache tokens as expected.
+    // Match logic consistent with plugin-server: exact match on provider OR substring match on model
+    let isAnthropic = false;
+    if (provider && provider.toLowerCase() === 'anthropic') {
+      isAnthropic = true;
+    } else if (model && model.toLowerCase().includes('anthropic')) {
+      isAnthropic = true;
+    }
+    if (isAnthropic && parsedUsage.input && additionalTokenData.cacheReadInputTokens) {
       parsedUsage.input = Math.max(parsedUsage.input - additionalTokenData.cacheReadInputTokens, 0);
     }
     return [parsedUsage.input, parsedUsage.output, additionalTokenData];
   }
-  parseUsage(response) {
+  parseUsage(response, provider, model) {
     let llmUsage = [0, 0, {}];
     const llmUsageKeys = ['token_usage', 'usage', 'tokenUsage'];
     if (response.llmOutput != null) {
       const key = llmUsageKeys.find(k => response.llmOutput?.[k] != null);
       if (key) {
-        llmUsage = this._parseUsageModel(response.llmOutput[key]);
+        llmUsage = this._parseUsageModel(response.llmOutput[key], provider, model);
       }
     }
     // If top-level usage info was not found, try checking the generations.
@@ -3947,14 +3955,14 @@ class LangChainCallbackHandler extends BaseCallbackHandler {
         for (const genChunk of generation) {
           // Check other paths for usage information
           if (genChunk.generationInfo?.usage_metadata) {
-            llmUsage = this._parseUsageModel(genChunk.generationInfo.usage_metadata);
+            llmUsage = this._parseUsageModel(genChunk.generationInfo.usage_metadata, provider, model);
             return llmUsage;
           }
           const messageChunk = genChunk.generationInfo ?? {};
           const responseMetadata = messageChunk.response_metadata ?? {};
           const chunkUsage = responseMetadata['usage'] ?? responseMetadata['amazon-bedrock-invocationMetrics'] ?? messageChunk.usage_metadata;
           if (chunkUsage) {
-            llmUsage = this._parseUsageModel(chunkUsage);
+            llmUsage = this._parseUsageModel(chunkUsage, provider, model);
             return llmUsage;
           }
         }