@posthog/ai 8.7.0 → 8.8.0

package/dist/vercel/index.mjs

@@ -10,6 +10,33 @@ const isObject = value => {
   return value !== null && typeof value === 'object' && !Array.isArray(value);
 };
 
+/** @internal */
+
+/** @internal */
+
+/** @internal */
+function isFullAiCaptureEnabled(client) {
+  return client?.enableFullAiCapture === true;
+}
+
+/** @internal */
+function captureAiEvent(client, event) {
+  if (isFullAiCaptureEnabled(client) && typeof client.captureAi === 'function') {
+    client.captureAi(event);
+    return;
+  }
+  client.capture(event);
+}
+
+/** @internal */
+async function captureAiEventImmediate(client, event) {
+  if (isFullAiCaptureEnabled(client) && typeof client.captureAiImmediate === 'function') {
+    await client.captureAiImmediate(event);
+    return;
+  }
+  await client.captureImmediate(event);
+}
+
 const DATA_URL_PREFIX_RE = /^data:([^;,\s]+)(?:;[^;,\s]+)*;base64,/i;
 const BASE64_ALPHABET_RE = /^[A-Za-z0-9+/_=-]+$/;
 class Base64Recognizer {
@@ -115,7 +142,6 @@ class BinaryContentRedactor {
     this.recognizer = recognizer;
   }
   redact(value, mediaType) {
-    if (this.isMultimodalEnabled()) return value;
     this.visited = new WeakSet();
     return this.walk(value, mediaType ? new MediaTypeContext(undefined, undefined, mediaType) : MediaTypeContext.EMPTY);
   }
@@ -159,24 +185,26 @@ class BinaryContentRedactor {
     if (mediaType === 'application/octet-stream') return '[base64 file redacted]';
     return `[base64 ${mediaType} redacted]`;
   }
-  isMultimodalEnabled() {
-    const val = process.env._INTERNAL_LLMA_MULTIMODAL || '';
-    return val.toLowerCase() === 'true' || val === '1' || val.toLowerCase() === 'yes';
-  }
 }
 
 const redactor = new BinaryContentRedactor();
 function redactBase64DataUrl(str, mediaType) {
   return redactor.redact(str, mediaType);
 }
-const sanitizeVercel = data => redactor.redact(data);
+const sanitize = (data, client) => isFullAiCaptureEnabled(client) ? data : redactor.redact(data);
+const sanitizeVercel = (data, client) => sanitize(data, client);
 
 const TOKEN_PROPERTY_KEYS = new Set(['$ai_input_tokens', '$ai_output_tokens', '$ai_cache_read_input_tokens', '$ai_cache_creation_input_tokens', '$ai_total_tokens', '$ai_reasoning_tokens']);
+
+/**
+ * Whether the caller supplied their own token counts, which override the ones the SDK
+ * derived from the provider response.
+ */
+function hasTokenOverrides(posthogProperties) {
+  return !!posthogProperties && Object.keys(posthogProperties).some(key => TOKEN_PROPERTY_KEYS.has(key));
+}
 function getTokensSource(posthogProperties) {
-  if (posthogProperties && Object.keys(posthogProperties).some(key => TOKEN_PROPERTY_KEYS.has(key))) {
-    return 'passthrough';
-  }
-  return 'sdk';
+  return hasTokenOverrides(posthogProperties) ? 'passthrough' : 'sdk';
 }
 
 // limit large outputs by truncating to 200kb (approx 200k bytes)
@@ -246,11 +274,14 @@ function toSafeString(input) {
     return '';
   }
 }
-const truncate = input => {
+const truncate = (input, client) => {
   const str = toSafeString(input);
   if (str === '') {
     return '';
   }
+  if (isFullAiCaptureEnabled(client)) {
+    return str;
+  }
 
   // Check if we need to truncate and ensure STRING_FORMAT is respected
   const buffer = sharedTextEncoder.encode(str);
@@ -406,7 +437,7 @@ function sanitizeValues(obj) {
   return jsonSafe;
 }
 
-var version = "8.7.0";
+var version = "8.8.0";
 
 const DEFAULT_MAX_DEPTH = 3;
 const MAX_STACK_LINES = 20;
@@ -573,6 +604,10 @@ const captureAiGeneration = async (client, options) => {
         $ai_total_cost_usd: inputCostUSD + outputCostUSD
       };
     }
+
+    // The caller's own token counts override the SDK-derived ones further down, via the
+    // `options.properties` spread.
+    const tokensOverridden = hasTokenOverrides(options.properties);
     const additionalTokenValues = {
       ...(usage.reasoningTokens ? {
         $ai_reasoning_tokens: usage.reasoningTokens
@@ -583,6 +618,18 @@ const captureAiGeneration = async (client, options) => {
       ...(usage.cacheCreationInputTokens ? {
         $ai_cache_creation_input_tokens: usage.cacheCreationInputTokens
       } : {}),
+      // Checked against undefined rather than truthiness, because false is the meaningful
+      // value here and a truthiness guard would drop it.
+      //
+      // Dropped entirely when the caller overrides the token counts: the flag describes how
+      // the SDK-derived counts relate to each other, so against passthrough counts it can be
+      // wrong in the expensive direction. Declaring inclusive over counts that are actually
+      // exclusive makes ingestion subtract the cache pool that was never in the input. A
+      // caller who knows their own accounting model can still pass
+      // `$ai_cache_reporting_exclusive` themselves, and that value wins.
+      ...(usage.cacheReportingExclusive !== undefined && !tokensOverridden ? {
+        $ai_cache_reporting_exclusive: usage.cacheReportingExclusive
+      } : {}),
       ...(usage.webSearchCount ? {
         $ai_web_search_count: usage.webSearchCount
       } : {}),
@@ -639,9 +686,9 @@ const captureAiGeneration = async (client, options) => {
       groups: options.groups
     };
     if (options.captureImmediate) {
-      await client.captureImmediate(event);
+      await captureAiEventImmediate(client, event);
     } else {
-      client.capture(event);
+      captureAiEvent(client, event);
     }
   } catch (error) {
     // Telemetry failures must never affect the instrumented provider call.
@@ -664,12 +711,12 @@ function getSpecificationVersion(model) {
 
 // Content types for the output array
 
-const redactFileData = (data, mediaType) => {
+const redactFileData = (data, mediaType, client) => {
   if (data instanceof URL) {
-    return redactBase64DataUrl(data.toString(), data.protocol === 'data:' ? mediaType : undefined);
+    return isFullAiCaptureEnabled(client) ? data.toString() : redactBase64DataUrl(data.toString(), data.protocol === 'data:' ? mediaType : undefined);
   }
   if (isString(data)) {
-    return redactBase64DataUrl(data, mediaType);
+    return isFullAiCaptureEnabled(client) ? data : redactBase64DataUrl(data, mediaType);
   }
   return undefined;
 };
@@ -684,7 +731,7 @@ const mapVercelParams = params => {
     stream: params.stream
   };
 };
-const mapVercelPrompt = messages => {
+const mapVercelPrompt = (messages, client) => {
   // Map and truncate individual content
   const inputs = messages.map(message => {
     let content;
@@ -693,7 +740,7 @@ const mapVercelPrompt = messages => {
     if (message.role === 'system') {
       content = [{
         type: 'text',
-        text: truncate(toContentString(message.content))
+        text: truncate(toContentString(message.content), client)
       }];
     } else {
       // Handle other roles which have array content
@@ -702,11 +749,11 @@ const mapVercelPrompt = messages => {
           if (c.type === 'text') {
             return {
               type: 'text',
-              text: truncate(c.text)
+              text: truncate(c.text, client)
             };
           } else if (c.type === 'file') {
             // Redact base64 data URLs and raw base64 to prevent oversized events
-            const fileData = redactFileData(c.data, c.mediaType) ?? 'raw files not supported';
+            const fileData = redactFileData(c.data, c.mediaType, client) ?? 'raw files not supported';
             return {
               type: 'file',
               file: fileData,
@@ -715,7 +762,7 @@ const mapVercelPrompt = messages => {
           } else if (c.type === 'reasoning') {
             return {
               type: 'reasoning',
-              text: truncate(c.text)
+              text: truncate(c.text, client)
             };
           } else if (c.type === 'tool-call') {
             return {
@@ -729,7 +776,7 @@ const mapVercelPrompt = messages => {
               type: 'tool-result',
               toolCallId: c.toolCallId,
               toolName: c.toolName,
-              output: sanitizeVercel(c.output),
+              output: sanitizeVercel(c.output, client),
               isError: c.isError
             };
           }
@@ -742,7 +789,7 @@ const mapVercelPrompt = messages => {
         // Fallback for non-array content
         content = [{
           type: 'text',
-          text: truncate(toContentString(message.content))
+          text: truncate(toContentString(message.content), client)
         }];
       }
     }
@@ -751,6 +798,12 @@ const mapVercelPrompt = messages => {
       content
     };
   });
+
+  // Full AI capture means no truncation of any kind; the aggregate trim below exists
+  // only to keep the default-mode payload under MAX_OUTPUT_SIZE.
+  if (isFullAiCaptureEnabled(client)) {
+    return inputs;
+  }
   try {
     // Trim the inputs array until its serialized JSON size fits within MAX_OUTPUT_SIZE.
     // Pre-compute each message's byte size once so we can shift by accumulated budget
@@ -787,12 +840,12 @@ const mapVercelPrompt = messages => {
   }
   return inputs;
 };
-const mapVercelOutput = result => {
+const mapVercelOutput = (result, client) => {
   const content = result.map(item => {
     if (item.type === 'text') {
       return {
         type: 'text',
-        text: truncate(item.text)
+        text: truncate(item.text, client)
       };
     }
     if (item.type === 'tool-call') {
@@ -810,15 +863,15 @@ const mapVercelOutput = result => {
     if (item.type === 'reasoning') {
       return {
         type: 'reasoning',
-        text: truncate(item.text)
+        text: truncate(item.text, client)
       };
     }
     if (item.type === 'file') {
       // Handle files similar to input mapping - avoid large base64 data
-      let fileData = redactFileData(item.data, item.mediaType) ?? `[binary ${item.mediaType} file]`;
+      let fileData = redactFileData(item.data, item.mediaType, client) ?? `[binary ${item.mediaType} file]`;
 
-      // If not redacted and still large, replace with size indicator
-      if (typeof item.data === 'string' && fileData === item.data && item.data.length > 1000) {
+      // Skipped under full AI capture: media stays untouched, so no placeholder swap either.
+      if (!isFullAiCaptureEnabled(client) && typeof item.data === 'string' && fileData === item.data && item.data.length > 1000) {
         fileData = `[${item.mediaType} file - ${item.data.length} bytes]`;
       }
       return {
@@ -840,7 +893,7 @@ const mapVercelOutput = result => {
     // Fallback for unknown types - try to extract text if possible
     return {
       type: 'text',
-      text: truncate(JSON.stringify(item))
+      text: truncate(JSON.stringify(item), client)
     };
   });
   if (content.length > 0) {
@@ -853,7 +906,7 @@ const mapVercelOutput = result => {
   try {
     const jsonOutput = JSON.stringify(result);
     return [{
-      content: truncate(jsonOutput),
+      content: truncate(jsonOutput, client),
       role: 'assistant'
     }];
   } catch {
@@ -1049,7 +1102,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
           const modelId = mergedOptions.posthogModelOverride ?? (result.response?.modelId ? result.response.modelId : model.modelId);
           const provider = mergedOptions.posthogProviderOverride ?? extractProvider(model);
           // result.content is undefined when the model returns only tool calls with no text output
-          const content = mapVercelOutput(result.content ?? []);
+          const content = mapVercelOutput(result.content ?? [], phClient);
           const latency = (Date.now() - startTime) / 1000;
           const providerMetadata = result.providerMetadata;
           const additionalTokenValues = extractAdditionalTokenValues(providerMetadata, result.usage);
@@ -1092,7 +1145,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
             ...baseOptions,
             model: modelId,
             provider: provider,
-            input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+            input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt, phClient),
             output: content,
             latency,
             baseURL,
@@ -1109,7 +1162,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
             ...baseOptions,
             model: modelId,
             provider: model.provider,
-            input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+            input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt, phClient),
             output: [],
             latency: 0,
             baseURL,
@@ -1243,13 +1296,13 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
               if (reasoningText) {
                 content.push({
                   type: 'reasoning',
-                  text: truncate(reasoningText)
+                  text: truncate(reasoningText, phClient)
                 });
               }
               if (generatedText) {
                 content.push({
                   type: 'text',
-                  text: truncate(generatedText)
+                  text: truncate(generatedText, phClient)
                 });
               }
               for (const toolCall of toolCallsInProgress.values()) {
@@ -1284,7 +1337,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
                 ...baseOptions,
                 model: modelId,
                 provider: provider,
-                input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+                input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt, phClient),
                 output,
                 latency,
                 timeToFirstToken,
@@ -1340,7 +1393,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
             ...baseOptions,
             model: modelId,
             provider: provider,
-            input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt),
+            input: mergedOptions.posthogPrivacyMode ? '' : mapVercelPrompt(params.prompt, phClient),
             output: [],
             latency: 0,
             baseURL,