@posthog/ai 7.2.1 → 7.3.0

package/dist/index.mjs

@@ -2,11 +2,10 @@ import { OpenAI, AzureOpenAI } from 'openai';
 import { Buffer } from 'buffer';
 import * as uuid from 'uuid';
 import { v4 } from 'uuid';
-import { wrapLanguageModel } from 'ai';
 import AnthropicOriginal from '@anthropic-ai/sdk';
 import { GoogleGenAI } from '@google/genai';
 
-var version = "7.2.1";
+var version = "7.3.0";
 
 // Type guards for safer type checking
 const isString = value => {
@@ -113,6 +112,13 @@ const formatResponseOpenAI = response => {
             });
           }
         }
+        // Handle audio output (gpt-4o-audio-preview)
+        if (choice.message.audio) {
+          content.push({
+            type: 'audio',
+            ...choice.message.audio
+          });
+        }
       }
       if (content.length > 0) {
         output.push({
@@ -194,6 +200,19 @@ const formatResponseGemini = response => {
                 arguments: part.functionCall.args
               }
             });
+          } else if (part.inlineData) {
+            // Handle audio/media inline data
+            const mimeType = part.inlineData.mimeType || 'audio/pcm';
+            let data = part.inlineData.data;
+            // Handle binary data (Buffer/Uint8Array -> base64)
+            if (data instanceof Uint8Array || Buffer.isBuffer(data)) {
+              data = Buffer.from(data).toString('base64');
+            }
+            content.push({
+              type: 'audio',
+              mime_type: mimeType,
+              data: data
+            });
           }
         }
         if (content.length > 0) {
@@ -598,6 +617,13 @@ function formatOpenAIResponsesInput(input, instructions) {
 
 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 => {
@@ -622,6 +648,7 @@ const isRawBase64 = str => {
   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)) {
@@ -672,12 +699,21 @@ const sanitizeOpenAIImage = item => {
       }
     };
   }
+  // Handle audio format
+  if (item.type === 'audio' && 'data' in item) {
+    if (isMultimodalEnabled()) return item;
+    return {
+      ...item,
+      data: REDACTED_IMAGE_PLACEHOLDER
+    };
+  }
   return item;
 };
 const sanitizeAnthropicImage = item => {
+  if (isMultimodalEnabled()) return item;
   if (!isObject(item)) return item;
-  // Handle Anthropic's image format
-  if (item.type === 'image' && 'source' in item && isObject(item.source) && item.source.type === 'base64' && 'data' in item.source) {
+  // Handle Anthropic's image and document formats (same structure, different type field)
+  if ((item.type === 'image' || item.type === 'document') && 'source' in item && isObject(item.source) && item.source.type === 'base64' && 'data' in item.source) {
     return {
       ...item,
       source: {
@@ -689,8 +725,9 @@ const sanitizeAnthropicImage = item => {
   return item;
 };
 const sanitizeGeminiPart = part => {
+  if (isMultimodalEnabled()) return part;
   if (!isObject(part)) return part;
-  // Handle Gemini's inline data format
+  // Handle Gemini's inline data format (images, audio, PDFs all use inlineData)
   if ('inlineData' in part && isObject(part.inlineData) && 'data' in part.inlineData) {
     return {
       ...part,
@@ -735,6 +772,7 @@ const sanitizeLangChainImage = item => {
   }
   // Anthropic style
   if (item.type === 'image' && 'source' in item && isObject(item.source) && 'data' in item.source) {
+    if (isMultimodalEnabled()) return item;
     return {
       ...item,
       source: {
@@ -820,6 +858,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
             try {
               const contentBlocks = [];
               let accumulatedContent = '';
+              let modelFromResponse;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0,
@@ -828,6 +867,10 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
               // Map to track in-progress tool calls
               const toolCallsInProgress = new Map();
               for await (const chunk of stream1) {
+                // Extract model from chunk (Chat Completions chunks have model field)
+                if (!modelFromResponse && chunk.model) {
+                  modelFromResponse = chunk.model;
+                }
                 const choice = chunk?.choices?.[0];
                 const chunkWebSearchCount = calculateWebSearchCount(chunk);
                 if (chunkWebSearchCount > 0 && chunkWebSearchCount > (usage.webSearchCount ?? 0)) {
@@ -915,7 +958,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
-                model: openAIParams.model,
+                model: openAIParams.model ?? modelFromResponse,
                 provider: 'openai',
                 input: sanitizeOpenAI(openAIParams.messages),
                 output: formattedOutput,
@@ -968,7 +1011,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
           await sendEventToPosthog({
             client: this.phClient,
             ...posthogParams,
-            model: openAIParams.model,
+            model: openAIParams.model ?? result.model,
             provider: 'openai',
             input: sanitizeOpenAI(openAIParams.messages),
             output: formattedOutput,
@@ -992,7 +1035,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
         await sendEventToPosthog({
           client: this.phClient,
           ...posthogParams,
-          model: String(openAIParams.model ?? ''),
+          model: openAIParams.model,
           provider: 'openai',
           input: sanitizeOpenAI(openAIParams.messages),
           output: [],
@@ -1034,6 +1077,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
           (async () => {
             try {
               let finalContent = [];
+              let modelFromResponse;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0,
@@ -1041,6 +1085,10 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
               };
               for await (const chunk of stream1) {
                 if ('response' in chunk && chunk.response) {
+                  // Extract model from response object in chunk (for stored prompts)
+                  if (!modelFromResponse && chunk.response.model) {
+                    modelFromResponse = chunk.response.model;
+                  }
                   const chunkWebSearchCount = calculateWebSearchCount(chunk.response);
                   if (chunkWebSearchCount > 0 && chunkWebSearchCount > (usage.webSearchCount ?? 0)) {
                     usage.webSearchCount = chunkWebSearchCount;
@@ -1064,8 +1112,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
-                //@ts-expect-error
-                model: openAIParams.model,
+                model: openAIParams.model ?? modelFromResponse,
                 provider: 'openai',
                 input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
                 output: finalContent,
@@ -1087,7 +1134,6 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
-                //@ts-expect-error
                 model: openAIParams.model,
                 provider: 'openai',
                 input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
@@ -1120,8 +1166,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
           await sendEventToPosthog({
             client: this.phClient,
             ...posthogParams,
-            //@ts-expect-error
-            model: openAIParams.model,
+            model: openAIParams.model ?? result.model,
             provider: 'openai',
             input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
             output: formattedOutput,
@@ -1145,7 +1190,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
         await sendEventToPosthog({
           client: this.phClient,
           ...posthogParams,
-          model: String(openAIParams.model ?? ''),
+          model: openAIParams.model,
           provider: 'openai',
           input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
           output: [],
@@ -1182,7 +1227,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
         await sendEventToPosthog({
           client: this.phClient,
           ...posthogParams,
-          model: String(openAIParams.model ?? ''),
+          model: openAIParams.model ?? result.model,
           provider: 'openai',
           input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
           output: result.output,
@@ -1203,7 +1248,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
         await sendEventToPosthog({
           client: this.phClient,
           ...posthogParams,
-          model: String(openAIParams.model ?? ''),
+          model: openAIParams.model,
           provider: 'openai',
           input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
           output: [],
@@ -1378,7 +1423,7 @@ class WrappedTranscriptions extends Transcriptions {
           await sendEventToPosthog({
             client: this.phClient,
             ...posthogParams,
-            model: String(openAIParams.model ?? ''),
+            model: openAIParams.model,
             provider: 'openai',
             input: openAIParams.prompt,
             output: result.text,
@@ -1398,7 +1443,7 @@ class WrappedTranscriptions extends Transcriptions {
         await sendEventToPosthog({
           client: this.phClient,
           ...posthogParams,
-          model: String(openAIParams.model ?? ''),
+          model: openAIParams.model,
           provider: 'openai',
           input: openAIParams.prompt,
           output: [],
@@ -1460,6 +1505,7 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
             try {
               const contentBlocks = [];
               let accumulatedContent = '';
+              let modelFromResponse;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0
@@ -1467,6 +1513,10 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
               // Map to track in-progress tool calls
               const toolCallsInProgress = new Map();
               for await (const chunk of stream1) {
+                // Extract model from response if not in params
+                if (!modelFromResponse && chunk.model) {
+                  modelFromResponse = chunk.model;
+                }
                 const choice = chunk?.choices?.[0];
                 // Handle text content
                 const deltaContent = choice?.delta?.content;
@@ -1548,7 +1598,7 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
-                model: openAIParams.model,
+                model: openAIParams.model ?? modelFromResponse,
                 provider: 'azure',
                 input: sanitizeOpenAI(openAIParams.messages),
                 output: formattedOutput,
@@ -1592,7 +1642,7 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
           await sendEventToPosthog({
             client: this.phClient,
             ...posthogParams,
-            model: openAIParams.model,
+            model: openAIParams.model ?? result.model,
             provider: 'azure',
             input: openAIParams.messages,
             output: formatResponseOpenAI(result),
@@ -1656,11 +1706,18 @@ class WrappedResponses extends AzureOpenAI.Responses {
           (async () => {
             try {
               let finalContent = [];
+              let modelFromResponse;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0
               };
               for await (const chunk of stream1) {
+                if ('response' in chunk && chunk.response) {
+                  // Extract model from response if not in params (for stored prompts)
+                  if (!modelFromResponse && chunk.response.model) {
+                    modelFromResponse = chunk.response.model;
+                  }
+                }
                 if (chunk.type === 'response.completed' && 'response' in chunk && chunk.response?.output && chunk.response.output.length > 0) {
                   finalContent = chunk.response.output;
                 }
@@ -1677,10 +1734,9 @@ class WrappedResponses extends AzureOpenAI.Responses {
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
-                //@ts-expect-error
-                model: openAIParams.model,
+                model: openAIParams.model ?? modelFromResponse,
                 provider: 'azure',
-                input: openAIParams.input,
+                input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
                 output: finalContent,
                 latency,
                 baseURL: this.baseURL,
@@ -1693,10 +1749,9 @@ class WrappedResponses extends AzureOpenAI.Responses {
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
-                //@ts-expect-error
                 model: openAIParams.model,
                 provider: 'azure',
-                input: openAIParams.input,
+                input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
                 output: [],
                 latency: 0,
                 baseURL: this.baseURL,
@@ -1722,10 +1777,9 @@ class WrappedResponses extends AzureOpenAI.Responses {
           await sendEventToPosthog({
             client: this.phClient,
             ...posthogParams,
-            //@ts-expect-error
-            model: openAIParams.model,
+            model: openAIParams.model ?? result.model,
             provider: 'azure',
-            input: openAIParams.input,
+            input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
             output: result.output,
             latency,
             baseURL: this.baseURL,
@@ -1745,10 +1799,9 @@ class WrappedResponses extends AzureOpenAI.Responses {
         await sendEventToPosthog({
           client: this.phClient,
           ...posthogParams,
-          //@ts-expect-error
           model: openAIParams.model,
           provider: 'azure',
-          input: openAIParams.input,
+          input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -1778,9 +1831,9 @@ class WrappedResponses extends AzureOpenAI.Responses {
       await sendEventToPosthog({
         client: this.phClient,
         ...posthogParams,
-        model: String(openAIParams.model ?? ''),
+        model: openAIParams.model ?? result.model,
         provider: 'azure',
-        input: openAIParams.input,
+        input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
         output: result.output,
         latency,
         baseURL: this.baseURL,
@@ -1798,9 +1851,9 @@ class WrappedResponses extends AzureOpenAI.Responses {
       await sendEventToPosthog({
         client: this.phClient,
         ...posthogParams,
-        model: String(openAIParams.model ?? ''),
+        model: openAIParams.model,
         provider: 'azure',
-        input: openAIParams.input,
+        input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
         output: [],
         latency: 0,
         baseURL: this.baseURL,
@@ -2073,67 +2126,117 @@ const extractProvider = model => {
   const providerName = provider.split('.')[0];
   return providerName;
 };
-const createInstrumentationMiddleware = (phClient, model, options) => {
-  const middleware = {
-    wrapGenerate: async ({
-      doGenerate,
-      params
-    }) => {
+// Extract web search count from provider metadata (works for both V2 and V3)
+const extractWebSearchCount = (providerMetadata, usage) => {
+  // Try Anthropic-specific extraction
+  if (providerMetadata && typeof providerMetadata === 'object' && 'anthropic' in providerMetadata && providerMetadata.anthropic && typeof providerMetadata.anthropic === 'object' && 'server_tool_use' in providerMetadata.anthropic) {
+    const serverToolUse = providerMetadata.anthropic.server_tool_use;
+    if (serverToolUse && typeof serverToolUse === 'object' && 'web_search_requests' in serverToolUse && typeof serverToolUse.web_search_requests === 'number') {
+      return serverToolUse.web_search_requests;
+    }
+  }
+  // Fall back to generic calculation
+  return calculateWebSearchCount({
+    usage,
+    providerMetadata
+  });
+};
+// Extract additional token values from provider metadata
+const extractAdditionalTokenValues = providerMetadata => {
+  if (providerMetadata && typeof providerMetadata === 'object' && 'anthropic' in providerMetadata && providerMetadata.anthropic && typeof providerMetadata.anthropic === 'object' && 'cacheCreationInputTokens' in providerMetadata.anthropic) {
+    return {
+      cacheCreationInputTokens: providerMetadata.anthropic.cacheCreationInputTokens
+    };
+  }
+  return {};
+};
+// Helper to extract numeric token value from V2 (number) or V3 (object with .total) usage formats
+const extractTokenCount = value => {
+  if (typeof value === 'number') {
+    return value;
+  }
+  if (value && typeof value === 'object' && 'total' in value && typeof value.total === 'number') {
+    return value.total;
+  }
+  return undefined;
+};
+// Helper to extract reasoning tokens from V2 (usage.reasoningTokens) or V3 (usage.outputTokens.reasoning)
+const extractReasoningTokens = usage => {
+  // V2 style: top-level reasoningTokens
+  if ('reasoningTokens' in usage) {
+    return usage.reasoningTokens;
+  }
+  // V3 style: nested in outputTokens.reasoning
+  if ('outputTokens' in usage && usage.outputTokens && typeof usage.outputTokens === 'object' && 'reasoning' in usage.outputTokens) {
+    return usage.outputTokens.reasoning;
+  }
+  return undefined;
+};
+// Helper to extract cached input tokens from V2 (usage.cachedInputTokens) or V3 (usage.inputTokens.cacheRead)
+const extractCacheReadTokens = usage => {
+  // V2 style: top-level cachedInputTokens
+  if ('cachedInputTokens' in usage) {
+    return usage.cachedInputTokens;
+  }
+  // V3 style: nested in inputTokens.cacheRead
+  if ('inputTokens' in usage && usage.inputTokens && typeof usage.inputTokens === 'object' && 'cacheRead' in usage.inputTokens) {
+    return usage.inputTokens.cacheRead;
+  }
+  return undefined;
+};
+/**
+ * Wraps a Vercel AI SDK language model (V2 or V3) with PostHog tracing.
+ * Automatically detects the model version and applies appropriate instrumentation.
+ */
+const wrapVercelLanguageModel = (model, phClient, options) => {
+  const traceId = options.posthogTraceId ?? v4();
+  const mergedOptions = {
+    ...options,
+    posthogTraceId: traceId,
+    posthogDistinctId: options.posthogDistinctId,
+    posthogProperties: {
+      ...options.posthogProperties,
+      $ai_framework: 'vercel',
+      $ai_framework_version: model.specificationVersion === 'v3' ? '6' : '5'
+    }
+  };
+  // Create wrapped model that preserves the original type
+  const wrappedModel = {
+    ...model,
+    doGenerate: async params => {
       const startTime = Date.now();
       const mergedParams = {
-        ...options,
-        ...mapVercelParams(params),
-        posthogProperties: {
-          ...options.posthogProperties,
-          $ai_framework: 'vercel'
-        }
+        ...mergedOptions,
+        ...mapVercelParams(params)
       };
       const availableTools = extractAvailableToolCalls('vercel', params);
       try {
-        const result = await doGenerate();
-        const modelId = options.posthogModelOverride ?? (result.response?.modelId ? result.response.modelId : model.modelId);
-        const provider = options.posthogProviderOverride ?? extractProvider(model);
+        const result = await model.doGenerate(params);
+        const modelId = mergedOptions.posthogModelOverride ?? (result.response?.modelId ? result.response.modelId : model.modelId);
+        const provider = mergedOptions.posthogProviderOverride ?? extractProvider(model);
         const baseURL = ''; // cannot currently get baseURL from vercel
         const content = mapVercelOutput(result.content);
         const latency = (Date.now() - startTime) / 1000;
         const providerMetadata = result.providerMetadata;
-        const additionalTokenValues = {
-          ...(providerMetadata?.anthropic ? {
-            cacheCreationInputTokens: providerMetadata.anthropic.cacheCreationInputTokens
-          } : {})
-        };
-        // Calculate web search count based on provider
-        let webSearchCount = 0;
-        if (providerMetadata?.anthropic && typeof providerMetadata.anthropic === 'object' && 'server_tool_use' in providerMetadata.anthropic) {
-          // Anthropic-specific extraction
-          const serverToolUse = providerMetadata.anthropic.server_tool_use;
-          if (serverToolUse && typeof serverToolUse === 'object' && 'web_search_requests' in serverToolUse && typeof serverToolUse.web_search_requests === 'number') {
-            webSearchCount = serverToolUse.web_search_requests;
-          }
-        } else {
-          // For other providers through Vercel, pass available metadata to helper
-          // Note: Vercel abstracts provider responses, so we may not have access to
-          // raw citations/annotations unless Vercel exposes them in usage/metadata
-          webSearchCount = calculateWebSearchCount({
-            usage: result.usage,
-            providerMetadata: providerMetadata
-          });
-        }
+        const additionalTokenValues = extractAdditionalTokenValues(providerMetadata);
+        const webSearchCount = extractWebSearchCount(providerMetadata, result.usage);
+        // V2 usage has simple numbers, V3 has objects with .total - normalize both
+        const usageObj = result.usage;
         const usage = {
-          inputTokens: result.usage.inputTokens,
-          outputTokens: result.usage.outputTokens,
-          reasoningTokens: result.usage.reasoningTokens,
-          cacheReadInputTokens: result.usage.cachedInputTokens,
+          inputTokens: extractTokenCount(result.usage.inputTokens),
+          outputTokens: extractTokenCount(result.usage.outputTokens),
+          reasoningTokens: extractReasoningTokens(usageObj),
+          cacheReadInputTokens: extractCacheReadTokens(usageObj),
           webSearchCount,
           ...additionalTokenValues
         };
         await sendEventToPosthog({
           client: phClient,
-          distinctId: options.posthogDistinctId,
-          traceId: options.posthogTraceId ?? v4(),
+          distinctId: mergedOptions.posthogDistinctId,
+          traceId: mergedOptions.posthogTraceId ?? v4(),
           model: modelId,
           provider: provider,
-          input: options.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+          input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
           output: content,
           latency,
           baseURL,
@@ -2141,18 +2244,18 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
           httpStatus: 200,
           usage,
           tools: availableTools,
-          captureImmediate: options.posthogCaptureImmediate
+          captureImmediate: mergedOptions.posthogCaptureImmediate
         });
         return result;
       } catch (error) {
         const modelId = model.modelId;
         await sendEventToPosthog({
           client: phClient,
-          distinctId: options.posthogDistinctId,
-          traceId: options.posthogTraceId ?? v4(),
+          distinctId: mergedOptions.posthogDistinctId,
+          traceId: mergedOptions.posthogTraceId ?? v4(),
           model: modelId,
           provider: model.provider,
-          input: options.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+          input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
           output: [],
           latency: 0,
           baseURL: '',
@@ -2165,30 +2268,23 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
           isError: true,
           error: truncate(JSON.stringify(error)),
           tools: availableTools,
-          captureImmediate: options.posthogCaptureImmediate
+          captureImmediate: mergedOptions.posthogCaptureImmediate
         });
         throw error;
       }
     },
-    wrapStream: async ({
-      doStream,
-      params
-    }) => {
+    doStream: async params => {
       const startTime = Date.now();
       let generatedText = '';
       let reasoningText = '';
       let usage = {};
       let providerMetadata = undefined;
       const mergedParams = {
-        ...options,
-        ...mapVercelParams(params),
-        posthogProperties: {
-          ...options.posthogProperties,
-          $ai_framework: 'vercel'
-        }
+        ...mergedOptions,
+        ...mapVercelParams(params)
       };
-      const modelId = options.posthogModelOverride ?? model.modelId;
-      const provider = options.posthogProviderOverride ?? extractProvider(model);
+      const modelId = mergedOptions.posthogModelOverride ?? model.modelId;
+      const provider = mergedOptions.posthogProviderOverride ?? extractProvider(model);
       const availableTools = extractAvailableToolCalls('vercel', params);
       const baseURL = ''; // cannot currently get baseURL from vercel
       // Map to track in-progress tool calls
@@ -2197,15 +2293,15 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
         const {
           stream,
           ...rest
-        } = await doStream();
+        } = await model.doStream(params);
         const transformStream = new TransformStream({
           transform(chunk, controller) {
-            // Handle new v5 streaming patterns
+            // Handle streaming patterns - compatible with both V2 and V3
             if (chunk.type === 'text-delta') {
               generatedText += chunk.delta;
             }
             if (chunk.type === 'reasoning-delta') {
-              reasoningText += chunk.delta; // New in v5
+              reasoningText += chunk.delta;
             }
             // Handle tool call chunks
             if (chunk.type === 'tool-input-start') {
@@ -2225,7 +2321,6 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
             }
             if (chunk.type === 'tool-input-end') {
               // Tool call is complete, keep it in the map for final processing
-              // Nothing specific to do here, the tool call is already complete
             }
             if (chunk.type === 'tool-call') {
               // Direct tool call chunk (complete tool call)
@@ -2237,14 +2332,13 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
             }
             if (chunk.type === 'finish') {
               providerMetadata = chunk.providerMetadata;
-              const additionalTokenValues = providerMetadata && typeof providerMetadata === 'object' && 'anthropic' in providerMetadata && providerMetadata.anthropic && typeof providerMetadata.anthropic === 'object' && 'cacheCreationInputTokens' in providerMetadata.anthropic ? {
-                cacheCreationInputTokens: providerMetadata.anthropic.cacheCreationInputTokens
-              } : {};
+              const additionalTokenValues = extractAdditionalTokenValues(providerMetadata);
+              const chunkUsage = chunk.usage || {};
               usage = {
-                inputTokens: chunk.usage?.inputTokens,
-                outputTokens: chunk.usage?.outputTokens,
-                reasoningTokens: chunk.usage?.reasoningTokens,
-                cacheReadInputTokens: chunk.usage?.cachedInputTokens,
+                inputTokens: extractTokenCount(chunk.usage?.inputTokens),
+                outputTokens: extractTokenCount(chunk.usage?.outputTokens),
+                reasoningTokens: extractReasoningTokens(chunkUsage),
+                cacheReadInputTokens: extractCacheReadTokens(chunkUsage),
                 ...additionalTokenValues
               };
             }
@@ -2284,23 +2378,7 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
               role: 'assistant',
               content: content.length === 1 && content[0].type === 'text' ? content[0].text : content
             }] : [];
-            // Calculate web search count based on provider
-            let webSearchCount = 0;
-            if (providerMetadata && typeof providerMetadata === 'object' && 'anthropic' in providerMetadata && providerMetadata.anthropic && typeof providerMetadata.anthropic === 'object' && 'server_tool_use' in providerMetadata.anthropic) {
-              // Anthropic-specific extraction
-              const serverToolUse = providerMetadata.anthropic.server_tool_use;
-              if (serverToolUse && typeof serverToolUse === 'object' && 'web_search_requests' in serverToolUse && typeof serverToolUse.web_search_requests === 'number') {
-                webSearchCount = serverToolUse.web_search_requests;
-              }
-            } else {
-              // For other providers through Vercel, pass available metadata to helper
-              // Note: Vercel abstracts provider responses, so we may not have access to
-              // raw citations/annotations unless Vercel exposes them in usage/metadata
-              webSearchCount = calculateWebSearchCount({
-                usage: usage,
-                providerMetadata: providerMetadata
-              });
-            }
+            const webSearchCount = extractWebSearchCount(providerMetadata, usage);
             // Update usage with web search count
             const finalUsage = {
               ...usage,
@@ -2308,11 +2386,11 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
             };
             await sendEventToPosthog({
               client: phClient,
-              distinctId: options.posthogDistinctId,
-              traceId: options.posthogTraceId ?? v4(),
+              distinctId: mergedOptions.posthogDistinctId,
+              traceId: mergedOptions.posthogTraceId ?? v4(),
               model: modelId,
               provider: provider,
-              input: options.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+              input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
               output: output,
               latency,
               baseURL,
@@ -2320,7 +2398,7 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
               httpStatus: 200,
               usage: finalUsage,
               tools: availableTools,
-              captureImmediate: options.posthogCaptureImmediate
+              captureImmediate: mergedOptions.posthogCaptureImmediate
             });
           }
         });
@@ -2331,11 +2409,11 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
       } catch (error) {
         await sendEventToPosthog({
           client: phClient,
-          distinctId: options.posthogDistinctId,
-          traceId: options.posthogTraceId ?? v4(),
+          distinctId: mergedOptions.posthogDistinctId,
+          traceId: mergedOptions.posthogTraceId ?? v4(),
           model: modelId,
           provider: provider,
-          input: options.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+          input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
           output: [],
           latency: 0,
           baseURL: '',
@@ -2348,25 +2426,12 @@ const createInstrumentationMiddleware = (phClient, model, options) => {
           isError: true,
           error: truncate(JSON.stringify(error)),
           tools: availableTools,
-          captureImmediate: options.posthogCaptureImmediate
+          captureImmediate: mergedOptions.posthogCaptureImmediate
         });
         throw error;
       }
     }
   };
-  return middleware;
-};
-const wrapVercelLanguageModel = (model, phClient, options) => {
-  const traceId = options.posthogTraceId ?? v4();
-  const middleware = createInstrumentationMiddleware(phClient, model, {
-    ...options,
-    posthogTraceId: traceId,
-    posthogDistinctId: options.posthogDistinctId
-  });
-  const wrappedModel = wrapLanguageModel({
-    model,
-    middleware
-  });
   return wrappedModel;
 };
 
@@ -2790,6 +2855,39 @@ class WrappedModels {
       throw error;
     }
   }
+  formatPartsAsContentBlocks(parts) {
+    const blocks = [];
+    for (const part of parts) {
+      // Handle dict/object with text field
+      if (part && typeof part === 'object' && 'text' in part && part.text) {
+        blocks.push({
+          type: 'text',
+          text: String(part.text)
+        });
+      }
+      // Handle string parts
+      else if (typeof part === 'string') {
+        blocks.push({
+          type: 'text',
+          text: part
+        });
+      }
+      // Handle inlineData (images, audio, PDFs)
+      else if (part && typeof part === 'object' && 'inlineData' in part) {
+        const inlineData = part.inlineData;
+        const mimeType = inlineData.mimeType || inlineData.mime_type || '';
+        const contentType = mimeType.startsWith('image/') ? 'image' : 'document';
+        blocks.push({
+          type: contentType,
+          inline_data: {
+            data: inlineData.data,
+            mime_type: mimeType
+          }
+        });
+      }
+    }
+    return blocks;
+  }
   formatInput(contents) {
     if (typeof contents === 'string') {
       return [{
@@ -2814,20 +2912,24 @@ class WrappedModels {
             };
           }
           if ('content' in obj && obj.content) {
+            // If content is a list, format it as content blocks
+            if (Array.isArray(obj.content)) {
+              const contentBlocks = this.formatPartsAsContentBlocks(obj.content);
+              return {
+                role: isString(obj.role) ? obj.role : 'user',
+                content: contentBlocks
+              };
+            }
             return {
               role: isString(obj.role) ? obj.role : 'user',
               content: obj.content
             };
           }
           if ('parts' in obj && Array.isArray(obj.parts)) {
+            const contentBlocks = this.formatPartsAsContentBlocks(obj.parts);
             return {
               role: isString(obj.role) ? obj.role : 'user',
-              content: obj.parts.map(part => {
-                if (part && typeof part === 'object' && 'text' in part) {
-                  return part.text;
-                }
-                return part;
-              })
+              content: contentBlocks
             };
           }
         }
@@ -3405,7 +3507,7 @@ var BaseCallbackHandler = class extends BaseCallbackHandlerMethodsClass {
 	}
 	static fromMethods(methods) {
 		class Handler extends BaseCallbackHandler {
-			name = uuid.v4();
+			name = uuid.v7();
 			constructor() {
 				super();
 				Object.assign(this, methods);