@posthog/ai 7.7.0 → 7.8.0

package/dist/openai/index.mjs

@@ -3,7 +3,7 @@ import { Buffer } from 'buffer';
 import { v4 } from 'uuid';
 import { uuidv7 } from '@posthog/core';
 
-var version = "7.7.0";
+var version = "7.8.0";
 
 // Type guards for safer type checking
 
@@ -496,6 +496,7 @@ const sendEventToPosthog = async ({
   input,
   output,
   latency,
+  timeToFirstToken,
   baseURL,
   params,
   httpStatus = 200,
@@ -562,6 +563,9 @@ 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,
@@ -634,6 +638,14 @@ 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;
@@ -690,6 +702,7 @@ class WrappedCompletions extends Completions {
               const contentBlocks = [];
               let accumulatedContent = '';
               let modelFromResponse;
+              let firstTokenTime;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0,
@@ -713,12 +726,18 @@ 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) {
@@ -794,6 +813,7 @@ 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,
@@ -803,6 +823,7 @@ class WrappedCompletions extends Completions {
                 input: sanitizeOpenAI(openAIParams.messages),
                 output: formattedOutput,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -925,6 +946,7 @@ class WrappedResponses extends Responses {
             try {
               let finalContent = [];
               let modelFromResponse;
+              let firstTokenTime;
               let usage = {
                 inputTokens: 0,
                 outputTokens: 0,
@@ -932,6 +954,10 @@ 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) {
@@ -957,6 +983,7 @@ 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,
@@ -966,6 +993,7 @@ class WrappedResponses extends Responses {
                 input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
                 output: finalContent,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,
@@ -1219,12 +1247,17 @@ 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;
                 }
@@ -1237,6 +1270,7 @@ 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,
@@ -1246,6 +1280,7 @@ class WrappedTranscriptions extends Transcriptions {
                 input: openAIParams.prompt,
                 output: finalContent,
                 latency,
+                timeToFirstToken,
                 baseURL: this.baseURL,
                 params: body,
                 httpStatus: 200,