@posthog/ai 7.5.3 → 7.6.0

package/dist/openai/index.cjs

@@ -7,7 +7,7 @@ var buffer = require('buffer');
 var uuid = require('uuid');
 var core = require('@posthog/core');
 
-var version = "7.5.3";
+var version = "7.6.0";
 
 // Type guards for safer type checking
 
@@ -18,6 +18,139 @@ const isObject = value => {
   return value !== null && typeof value === 'object' && !Array.isArray(value);
 };
 
+const REDACTED_IMAGE_PLACEHOLDER = '[base64 image redacted]';
+
+// ============================================
+// Multimodal Feature Toggle
+// ============================================
+
+const isMultimodalEnabled = () => {
+  const val = process.env._INTERNAL_LLMA_MULTIMODAL || '';
+  return val.toLowerCase() === 'true' || val === '1' || val.toLowerCase() === 'yes';
+};
+
+// ============================================
+// Base64 Detection Helpers
+// ============================================
+
+const isBase64DataUrl = str => {
+  return /^data:([^;]+);base64,/.test(str);
+};
+const isValidUrl = str => {
+  try {
+    new URL(str);
+    return true;
+  } catch {
+    // Not an absolute URL, check if it's a relative URL or path
+    return str.startsWith('/') || str.startsWith('./') || str.startsWith('../');
+  }
+};
+const isRawBase64 = str => {
+  // Skip if it's a valid URL or path
+  if (isValidUrl(str)) {
+    return false;
+  }
+
+  // Check if it's a valid base64 string
+  // Base64 images are typically at least a few hundred chars, but we'll be conservative
+  return str.length > 20 && /^[A-Za-z0-9+/]+=*$/.test(str);
+};
+function redactBase64DataUrl(str) {
+  if (isMultimodalEnabled()) return str;
+  if (!isString(str)) return str;
+
+  // Check for data URL format
+  if (isBase64DataUrl(str)) {
+    return REDACTED_IMAGE_PLACEHOLDER;
+  }
+
+  // Check for raw base64 (Vercel sends raw base64 for inline images)
+  if (isRawBase64(str)) {
+    return REDACTED_IMAGE_PLACEHOLDER;
+  }
+  return str;
+}
+
+// ============================================
+// Common Message Processing
+// ============================================
+
+const processMessages = (messages, transformContent) => {
+  if (!messages) return messages;
+  const processContent = content => {
+    if (typeof content === 'string') return content;
+    if (!content) return content;
+    if (Array.isArray(content)) {
+      return content.map(transformContent);
+    }
+
+    // Handle single object content
+    return transformContent(content);
+  };
+  const processMessage = msg => {
+    if (!isObject(msg) || !('content' in msg)) return msg;
+    return {
+      ...msg,
+      content: processContent(msg.content)
+    };
+  };
+
+  // Handle both arrays and single messages
+  if (Array.isArray(messages)) {
+    return messages.map(processMessage);
+  }
+  return processMessage(messages);
+};
+
+// ============================================
+// Provider-Specific Image Sanitizers
+// ============================================
+
+const sanitizeOpenAIImage = item => {
+  if (!isObject(item)) return item;
+
+  // Handle image_url format
+  if (item.type === 'image_url' && 'image_url' in item && isObject(item.image_url) && 'url' in item.image_url) {
+    return {
+      ...item,
+      image_url: {
+        ...item.image_url,
+        url: redactBase64DataUrl(item.image_url.url)
+      }
+    };
+  }
+
+  // Handle audio format
+  if (item.type === 'audio' && 'data' in item) {
+    if (isMultimodalEnabled()) return item;
+    return {
+      ...item,
+      data: REDACTED_IMAGE_PLACEHOLDER
+    };
+  }
+  return item;
+};
+const sanitizeOpenAIResponseImage = item => {
+  if (!isObject(item)) return item;
+
+  // Handle input_image format
+  if (item.type === 'input_image' && 'image_url' in item) {
+    return {
+      ...item,
+      image_url: redactBase64DataUrl(item.image_url)
+    };
+  }
+  return item;
+};
+
+// Export individual sanitizers for tree-shaking
+const sanitizeOpenAI = data => {
+  return processMessages(data, sanitizeOpenAIImage);
+};
+const sanitizeOpenAIResponse = data => {
+  return processMessages(data, sanitizeOpenAIResponseImage);
+};
+
 const STRING_FORMAT = 'utf8';
 
 /**
@@ -413,6 +546,9 @@ const sendEventToPosthog = async ({
     } : {}),
     ...(usage.webSearchCount ? {
       $ai_web_search_count: usage.webSearchCount
+    } : {}),
+    ...(usage.rawUsage ? {
+      $ai_usage: usage.rawUsage
     } : {})
   };
   const properties = {
@@ -502,124 +638,6 @@ function formatOpenAIResponsesInput(input, instructions) {
   return messages;
 }
 
-const REDACTED_IMAGE_PLACEHOLDER = '[base64 image redacted]';
-
-// ============================================
-// Multimodal Feature Toggle
-// ============================================
-
-const isMultimodalEnabled = () => {
-  const val = process.env._INTERNAL_LLMA_MULTIMODAL || '';
-  return val.toLowerCase() === 'true' || val === '1' || val.toLowerCase() === 'yes';
-};
-
-// ============================================
-// Base64 Detection Helpers
-// ============================================
-
-const isBase64DataUrl = str => {
-  return /^data:([^;]+);base64,/.test(str);
-};
-const isValidUrl = str => {
-  try {
-    new URL(str);
-    return true;
-  } catch {
-    // Not an absolute URL, check if it's a relative URL or path
-    return str.startsWith('/') || str.startsWith('./') || str.startsWith('../');
-  }
-};
-const isRawBase64 = str => {
-  // Skip if it's a valid URL or path
-  if (isValidUrl(str)) {
-    return false;
-  }
-
-  // Check if it's a valid base64 string
-  // Base64 images are typically at least a few hundred chars, but we'll be conservative
-  return str.length > 20 && /^[A-Za-z0-9+/]+=*$/.test(str);
-};
-function redactBase64DataUrl(str) {
-  if (isMultimodalEnabled()) return str;
-  if (!isString(str)) return str;
-
-  // Check for data URL format
-  if (isBase64DataUrl(str)) {
-    return REDACTED_IMAGE_PLACEHOLDER;
-  }
-
-  // Check for raw base64 (Vercel sends raw base64 for inline images)
-  if (isRawBase64(str)) {
-    return REDACTED_IMAGE_PLACEHOLDER;
-  }
-  return str;
-}
-
-// ============================================
-// Common Message Processing
-// ============================================
-
-const processMessages = (messages, transformContent) => {
-  if (!messages) return messages;
-  const processContent = content => {
-    if (typeof content === 'string') return content;
-    if (!content) return content;
-    if (Array.isArray(content)) {
-      return content.map(transformContent);
-    }
-
-    // Handle single object content
-    return transformContent(content);
-  };
-  const processMessage = msg => {
-    if (!isObject(msg) || !('content' in msg)) return msg;
-    return {
-      ...msg,
-      content: processContent(msg.content)
-    };
-  };
-
-  // Handle both arrays and single messages
-  if (Array.isArray(messages)) {
-    return messages.map(processMessage);
-  }
-  return processMessage(messages);
-};
-
-// ============================================
-// Provider-Specific Image Sanitizers
-// ============================================
-
-const sanitizeOpenAIImage = item => {
-  if (!isObject(item)) return item;
-
-  // Handle image_url format
-  if (item.type === 'image_url' && 'image_url' in item && isObject(item.image_url) && 'url' in item.image_url) {
-    return {
-      ...item,
-      image_url: {
-        ...item.image_url,
-        url: redactBase64DataUrl(item.image_url.url)
-      }
-    };
-  }
-
-  // Handle audio format
-  if (item.type === 'audio' && 'data' in item) {
-    if (isMultimodalEnabled()) return item;
-    return {
-      ...item,
-      data: REDACTED_IMAGE_PLACEHOLDER
-    };
-  }
-  return item;
-};
-
-// Export individual sanitizers for tree-shaking
-const sanitizeOpenAI = data => {
-  return processMessages(data, sanitizeOpenAIImage);
-};
-
 const Chat = openai.OpenAI.Chat;
 const Completions = Chat.Completions;
 const Responses = openai.OpenAI.Responses;
@@ -684,6 +702,7 @@ class WrappedCompletions extends Completions {
 
               // Map to track in-progress tool calls
               const toolCallsInProgress = new Map();
+              let rawUsageData;
               for await (const chunk of stream1) {
                 // Extract model from chunk (Chat Completions chunks have model field)
                 if (!modelFromResponse && chunk.model) {
@@ -734,6 +753,7 @@ class WrappedCompletions extends Completions {
 
                 // Handle usage information
                 if (chunk.usage) {
+                  rawUsageData = chunk.usage;
                   usage = {
                     ...usage,
                     inputTokens: chunk.usage.prompt_tokens ?? 0,
@@ -795,7 +815,8 @@ class WrappedCompletions extends Completions {
                   outputTokens: usage.outputTokens,
                   reasoningTokens: usage.reasoningTokens,
                   cacheReadInputTokens: usage.cacheReadInputTokens,
-                  webSearchCount: usage.webSearchCount
+                  webSearchCount: usage.webSearchCount,
+                  rawUsage: rawUsageData
                 },
                 tools: availableTools
               });
@@ -847,7 +868,8 @@ class WrappedCompletions extends Completions {
               outputTokens: result.usage?.completion_tokens ?? 0,
               reasoningTokens: result.usage?.completion_tokens_details?.reasoning_tokens ?? 0,
               cacheReadInputTokens: result.usage?.prompt_tokens_details?.cached_tokens ?? 0,
-              webSearchCount: calculateWebSearchCount(result)
+              webSearchCount: calculateWebSearchCount(result),
+              rawUsage: result.usage
             },
             tools: availableTools
           });
@@ -912,6 +934,7 @@ class WrappedResponses extends Responses {
                 outputTokens: 0,
                 webSearchCount: 0
               };
+              let rawUsageData;
               for await (const chunk of stream1) {
                 if ('response' in chunk && chunk.response) {
                   // Extract model from response object in chunk (for stored prompts)
@@ -927,6 +950,7 @@ class WrappedResponses extends Responses {
                   finalContent = chunk.response.output;
                 }
                 if ('response' in chunk && chunk.response?.usage) {
+                  rawUsageData = chunk.response.usage;
                   usage = {
                     ...usage,
                     inputTokens: chunk.response.usage.input_tokens ?? 0,
@@ -943,7 +967,7 @@ class WrappedResponses extends Responses {
                 ...posthogParams,
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'openai',
-                input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
                 output: finalContent,
                 latency,
                 baseURL: this.baseURL,
@@ -954,7 +978,8 @@ class WrappedResponses extends Responses {
                   outputTokens: usage.outputTokens,
                   reasoningTokens: usage.reasoningTokens,
                   cacheReadInputTokens: usage.cacheReadInputTokens,
-                  webSearchCount: usage.webSearchCount
+                  webSearchCount: usage.webSearchCount,
+                  rawUsage: rawUsageData
                 },
                 tools: availableTools
               });
@@ -964,7 +989,7 @@ class WrappedResponses extends Responses {
                 ...posthogParams,
                 model: openAIParams.model,
                 provider: 'openai',
-                input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
                 output: [],
                 latency: 0,
                 baseURL: this.baseURL,
@@ -995,7 +1020,7 @@ class WrappedResponses extends Responses {
             ...posthogParams,
             model: openAIParams.model ?? result.model,
             provider: 'openai',
-            input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+            input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
             output: formattedOutput,
             latency,
             baseURL: this.baseURL,
@@ -1006,7 +1031,8 @@ class WrappedResponses extends Responses {
               outputTokens: result.usage?.output_tokens ?? 0,
               reasoningTokens: result.usage?.output_tokens_details?.reasoning_tokens ?? 0,
               cacheReadInputTokens: result.usage?.input_tokens_details?.cached_tokens ?? 0,
-              webSearchCount: calculateWebSearchCount(result)
+              webSearchCount: calculateWebSearchCount(result),
+              rawUsage: result.usage
             },
             tools: availableTools
           });
@@ -1019,7 +1045,7 @@ class WrappedResponses extends Responses {
           ...posthogParams,
           model: openAIParams.model,
           provider: 'openai',
-          input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -1055,7 +1081,7 @@ class WrappedResponses extends Responses {
           ...posthogParams,
           model: openAIParams.model ?? result.model,
           provider: 'openai',
-          input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
           output: result.output,
           latency,
           baseURL: this.baseURL,
@@ -1065,7 +1091,8 @@ class WrappedResponses extends Responses {
             inputTokens: result.usage?.input_tokens ?? 0,
             outputTokens: result.usage?.output_tokens ?? 0,
             reasoningTokens: result.usage?.output_tokens_details?.reasoning_tokens ?? 0,
-            cacheReadInputTokens: result.usage?.input_tokens_details?.cached_tokens ?? 0
+            cacheReadInputTokens: result.usage?.input_tokens_details?.cached_tokens ?? 0,
+            rawUsage: result.usage
           }
         });
         return result;
@@ -1075,7 +1102,7 @@ class WrappedResponses extends Responses {
           ...posthogParams,
           model: openAIParams.model,
           provider: 'openai',
-          input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -1124,7 +1151,8 @@ class WrappedEmbeddings extends Embeddings {
         params: body,
         httpStatus: 200,
         usage: {
-          inputTokens: result.usage?.prompt_tokens ?? 0
+          inputTokens: result.usage?.prompt_tokens ?? 0,
+          rawUsage: result.usage
         }
       });
       return result;
@@ -1207,7 +1235,8 @@ class WrappedTranscriptions extends Transcriptions {
                 if ('usage' in chunk && chunk.usage) {
                   usage = {
                     inputTokens: chunk.usage?.type === 'tokens' ? chunk.usage.input_tokens ?? 0 : 0,
-                    outputTokens: chunk.usage?.type === 'tokens' ? chunk.usage.output_tokens ?? 0 : 0
+                    outputTokens: chunk.usage?.type === 'tokens' ? chunk.usage.output_tokens ?? 0 : 0,
+                    rawUsage: chunk.usage
                   };
                 }
               }
@@ -1268,7 +1297,8 @@ class WrappedTranscriptions extends Transcriptions {
             httpStatus: 200,
             usage: {
               inputTokens: result.usage?.type === 'tokens' ? result.usage.input_tokens ?? 0 : 0,
-              outputTokens: result.usage?.type === 'tokens' ? result.usage.output_tokens ?? 0 : 0
+              outputTokens: result.usage?.type === 'tokens' ? result.usage.output_tokens ?? 0 : 0,
+              rawUsage: result.usage
             }
           });
           return result;