@posthog/ai 7.7.0 → 7.8.0

package/dist/index.mjs

@@ -6,7 +6,7 @@ import { uuidv7 } from '@posthog/core';
 import AnthropicOriginal from '@anthropic-ai/sdk';
 import { GoogleGenAI } from '@google/genai';
 
-var version = "7.7.0";
+var version = "7.8.0";
 
 // Type guards for safer type checking
 const isString = value => {
@@ -722,6 +722,7 @@ const sendEventToPosthog = async ({
   input,
   output,
   latency,
+  timeToFirstToken,
   baseURL,
   params,
   httpStatus = 200,
@@ -788,6 +789,9 @@ const sendEventToPosthog = async ({
     } : {}),
     ...additionalTokenValues,
     $ai_latency: latency,
+    ...(timeToFirstToken !== undefined ? {
+      $ai_time_to_first_token: timeToFirstToken
+    } : {}),
     $ai_trace_id: traceId,
     $ai_base_url: baseURL,
     ...params.posthogProperties,
@@ -859,6 +863,14 @@ function formatOpenAIResponsesInput(input, instructions) {
   return messages;
 }
 
+/**
+ * Checks if a ResponseStreamEvent chunk represents the first token/content from the model.
+ * This includes various content types like text, reasoning, audio, and refusals.
+ */
+function isResponseTokenChunk(chunk) {
+  return chunk.type === 'response.output_item.added' || chunk.type === 'response.content_part.added' || chunk.type === 'response.output_text.delta' || chunk.type === 'response.reasoning_text.delta' || chunk.type === 'response.reasoning_summary_text.delta' || chunk.type === 'response.audio.delta' || chunk.type === 'response.audio.transcript.delta' || chunk.type === 'response.refusal.delta';
+}
+
 const Chat = OpenAI.Chat;
 const Completions = Chat.Completions;
 const Responses = OpenAI.Responses;
@@ -908,6 +920,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
               const contentBlocks = [];
               let accumulatedContent = '';
               let modelFromResponse;
+              let firstTokenTime;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0,
@@ -929,11 +942,17 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
                 // Handle text content
                 const deltaContent = choice?.delta?.content;
                 if (deltaContent) {
+                  if (firstTokenTime === undefined) {
+                    firstTokenTime = Date.now();
+                  }
                   accumulatedContent += deltaContent;
                 }
                 // Handle tool calls
                 const deltaToolCalls = choice?.delta?.tool_calls;
                 if (deltaToolCalls && Array.isArray(deltaToolCalls)) {
+                  if (firstTokenTime === undefined) {
+                    firstTokenTime = Date.now();
+                  }
                   for (const toolCall of deltaToolCalls) {
                     const index = toolCall.index;
                     if (index !== undefined) {
@@ -1005,6 +1024,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
                 }]
               }];
               const latency = (Date.now() - startTime) / 1000;
+              const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
               const availableTools = extractAvailableToolCalls('openai', openAIParams);
               await sendEventToPosthog({
                 client: this.phClient,
@@ -1014,6 +1034,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends Completions {
                 input: sanitizeOpenAI(openAIParams.messages),
                 output: formattedOutput,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -1128,6 +1149,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
             try {
               let finalContent = [];
               let modelFromResponse;
+              let firstTokenTime;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0,
@@ -1135,6 +1157,10 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
               };
               let rawUsageData;
               for await (const chunk of stream1) {
+                // Track first token time on content delta events
+                if (firstTokenTime === undefined && isResponseTokenChunk(chunk)) {
+                  firstTokenTime = Date.now();
+                }
                 if ('response' in chunk && chunk.response) {
                   // Extract model from response object in chunk (for stored prompts)
                   if (!modelFromResponse && chunk.response.model) {
@@ -1160,6 +1186,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
                 }
               }
               const latency = (Date.now() - startTime) / 1000;
+              const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
               const availableTools = extractAvailableToolCalls('openai', openAIParams);
               await sendEventToPosthog({
                 client: this.phClient,
@@ -1169,6 +1196,7 @@ let WrappedResponses$1 = class WrappedResponses extends Responses {
                 input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
                 output: finalContent,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -1407,12 +1435,17 @@ class WrappedTranscriptions extends Transcriptions {
           (async () => {
             try {
               let finalContent = '';
+              let firstTokenTime;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0
               };
               const doneEvent = 'transcript.text.done';
               for await (const chunk of stream1) {
+                // Track first token on text delta events
+                if (firstTokenTime === undefined && chunk.type === 'transcript.text.delta') {
+                  firstTokenTime = Date.now();
+                }
                 if (chunk.type === doneEvent && 'text' in chunk && chunk.text && chunk.text.length > 0) {
                   finalContent = chunk.text;
                 }
@@ -1425,6 +1458,7 @@ class WrappedTranscriptions extends Transcriptions {
                 }
               }
               const latency = (Date.now() - startTime) / 1000;
+              const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
               const availableTools = extractAvailableToolCalls('openai', openAIParams);
               await sendEventToPosthog({
                 client: this.phClient,
@@ -1434,6 +1468,7 @@ class WrappedTranscriptions extends Transcriptions {
                 input: openAIParams.prompt,
                 output: finalContent,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -1552,6 +1587,7 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
               const contentBlocks = [];
               let accumulatedContent = '';
               let modelFromResponse;
+              let firstTokenTime;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0
@@ -1567,11 +1603,17 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
                 // Handle text content
                 const deltaContent = choice?.delta?.content;
                 if (deltaContent) {
+                  if (firstTokenTime === undefined) {
+                    firstTokenTime = Date.now();
+                  }
                   accumulatedContent += deltaContent;
                 }
                 // Handle tool calls
                 const deltaToolCalls = choice?.delta?.tool_calls;
                 if (deltaToolCalls && Array.isArray(deltaToolCalls)) {
+                  if (firstTokenTime === undefined) {
+                    firstTokenTime = Date.now();
+                  }
                   for (const toolCall of deltaToolCalls) {
                     const index = toolCall.index;
                     if (index !== undefined) {
@@ -1641,6 +1683,7 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
                 }]
               }];
               const latency = (Date.now() - startTime) / 1000;
+              const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
@@ -1649,6 +1692,7 @@ class WrappedCompletions extends AzureOpenAI.Chat.Completions {
                 input: sanitizeOpenAI(openAIParams.messages),
                 output: formattedOutput,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -1750,11 +1794,16 @@ class WrappedResponses extends AzureOpenAI.Responses {
             try {
               let finalContent = [];
               let modelFromResponse;
+              let firstTokenTime;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0
               };
               for await (const chunk of stream1) {
+                // Track first token time on content delta events
+                if (firstTokenTime === undefined && isResponseTokenChunk(chunk)) {
+                  firstTokenTime = Date.now();
+                }
                 if ('response' in chunk && chunk.response) {
                   // Extract model from response if not in params (for stored prompts)
                   if (!modelFromResponse && chunk.response.model) {
@@ -1774,6 +1823,7 @@ class WrappedResponses extends AzureOpenAI.Responses {
                 }
               }
               const latency = (Date.now() - startTime) / 1000;
+              const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
               await sendEventToPosthog({
                 client: this.phClient,
                 ...posthogParams,
@@ -1782,6 +1832,7 @@ class WrappedResponses extends AzureOpenAI.Responses {
                 input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
                 output: finalContent,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -2351,6 +2402,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
     doStream: {
       value: async params => {
         const startTime = Date.now();
+        let firstTokenTime;
         let generatedText = '';
         let reasoningText = '';
         let usage = {};
@@ -2374,13 +2426,22 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
             transform(chunk, controller) {
               // Handle streaming patterns - compatible with both V2 and V3
               if (chunk.type === 'text-delta') {
+                if (firstTokenTime === undefined) {
+                  firstTokenTime = Date.now();
+                }
                 generatedText += chunk.delta;
               }
               if (chunk.type === 'reasoning-delta') {
+                if (firstTokenTime === undefined) {
+                  firstTokenTime = Date.now();
+                }
                 reasoningText += chunk.delta;
               }
               // Handle tool call chunks
               if (chunk.type === 'tool-input-start') {
+                if (firstTokenTime === undefined) {
+                  firstTokenTime = Date.now();
+                }
                 // Initialize a new tool call
                 toolCallsInProgress.set(chunk.id, {
                   toolCallId: chunk.id,
@@ -2399,6 +2460,9 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
                 // Tool call is complete, keep it in the map for final processing
               }
               if (chunk.type === 'tool-call') {
+                if (firstTokenTime === undefined) {
+                  firstTokenTime = Date.now();
+                }
                 // Direct tool call chunk (complete tool call)
                 toolCallsInProgress.set(chunk.toolCallId, {
                   toolCallId: chunk.toolCallId,
@@ -2422,6 +2486,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
             },
             flush: async () => {
               const latency = (Date.now() - startTime) / 1000;
+              const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
               // Build content array similar to mapVercelOutput structure
               const content = [];
               if (reasoningText) {
@@ -2474,6 +2539,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
                 input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
                 output: output,
                 latency,
+                timeToFirstToken,
                 baseURL,
                 params: mergedParams,
                 httpStatus: 200,
@@ -2548,6 +2614,7 @@ class WrappedMessages extends AnthropicOriginal.Messages {
         const contentBlocks = [];
         const toolsInProgress = new Map();
         let currentTextBlock = null;
+        let firstTokenTime;
         const usage = {
           inputTokens: 0,
           outputTokens: 0,
@@ -2570,6 +2637,9 @@ class WrappedMessages extends AnthropicOriginal.Messages {
                     };
                     contentBlocks.push(currentTextBlock);
                   } else if (chunk.content_block?.type === 'tool_use') {
+                    if (firstTokenTime === undefined) {
+                      firstTokenTime = Date.now();
+                    }
                     const toolBlock = {
                       type: 'function',
                       id: chunk.content_block.id,
@@ -2590,6 +2660,9 @@ class WrappedMessages extends AnthropicOriginal.Messages {
                 if ('delta' in chunk) {
                   if ('text' in chunk.delta) {
                     const delta = chunk.delta.text;
+                    if (firstTokenTime === undefined) {
+                      firstTokenTime = Date.now();
+                    }
                     accumulatedContent += delta;
                     if (currentTextBlock) {
                       currentTextBlock.text += delta;
@@ -2645,6 +2718,7 @@ class WrappedMessages extends AnthropicOriginal.Messages {
               }
               usage.rawUsage = lastRawUsage;
               const latency = (Date.now() - startTime) / 1000;
+              const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
               const availableTools = extractAvailableToolCalls('anthropic', anthropicParams);
               // Format output to match non-streaming version
               const formattedOutput = contentBlocks.length > 0 ? [{
@@ -2665,6 +2739,7 @@ class WrappedMessages extends AnthropicOriginal.Messages {
                 input: sanitizeAnthropic(mergeSystemPrompt(anthropicParams, 'anthropic')),
                 output: formattedOutput,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -2826,6 +2901,7 @@ class WrappedModels {
     } = extractPosthogParams(params);
     const startTime = Date.now();
     const accumulatedContent = [];
+    let firstTokenTime;
     let usage = {
       inputTokens: 0,
       outputTokens: 0,
@@ -2835,6 +2911,10 @@ class WrappedModels {
     try {
       const stream = await this.client.models.generateContentStream(geminiParams);
       for await (const chunk of stream) {
+        // Track first token time when we get text content
+        if (firstTokenTime === undefined && chunk.text) {
+          firstTokenTime = Date.now();
+        }
         const chunkWebSearchCount = calculateGoogleWebSearchCount(chunk);
         if (chunkWebSearchCount > 0 && chunkWebSearchCount > (usage.webSearchCount ?? 0)) {
           usage.webSearchCount = chunkWebSearchCount;
@@ -2865,6 +2945,9 @@ class WrappedModels {
               for (const part of candidate.content.parts) {
                 // Type-safe check for functionCall
                 if ('functionCall' in part) {
+                  if (firstTokenTime === undefined) {
+                    firstTokenTime = Date.now();
+                  }
                   const funcCall = part.functionCall;
                   if (funcCall?.name) {
                     accumulatedContent.push({
@@ -2895,6 +2978,7 @@ class WrappedModels {
         yield chunk;
       }
       const latency = (Date.now() - startTime) / 1000;
+      const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
       const availableTools = extractAvailableToolCalls('gemini', geminiParams);
       // Format output similar to formatResponseGemini
       const output = accumulatedContent.length > 0 ? [{
@@ -2909,6 +2993,7 @@ class WrappedModels {
         input: this.formatInputForPostHog(geminiParams),
         output,
         latency,
+        timeToFirstToken,
         baseURL: 'https://generativelanguage.googleapis.com',
         params: params,
         httpStatus: 200,