@posthog/ai 8.9.2 → 8.10.0

package/dist/gemini/index.cjs

@@ -280,6 +280,23 @@ const buildInlineDataBlock = (mimeType, data) => {
     }
   };
 };
+const formatInlineDataBlock = (inlineData, client) => {
+  const mimeType = inlineData.mimeType || inlineData.mime_type || 'application/octet-stream';
+  let data = inlineData.data;
+  if (data instanceof Uint8Array) {
+    if (typeof Buffer !== 'undefined') {
+      data = Buffer.from(data).toString('base64');
+    } else {
+      let binary = '';
+      for (let i = 0; i < data.length; i++) {
+        binary += String.fromCharCode(data[i]);
+      }
+      data = btoa(binary);
+    }
+  }
+  data = isFullAiCaptureEnabled(client) ? data : redactBase64DataUrl(data, mimeType);
+  return buildInlineDataBlock(mimeType, String(data ?? ''));
+};
 const formatResponseGemini = (response, client) => {
   const output = [];
   if (response.candidates && Array.isArray(response.candidates)) {
@@ -301,26 +318,7 @@ const formatResponseGemini = (response, client) => {
               }
             });
           } else if (part.inlineData) {
-            // Handle inline data (images, audio, documents)
-            const mimeType = part.inlineData.mimeType || part.inlineData.mime_type || 'application/octet-stream';
-            let data = part.inlineData.data;
-
-            // Handle binary data (Uint8Array/Buffer -> base64)
-            if (data instanceof Uint8Array) {
-              if (typeof Buffer !== 'undefined') {
-                data = Buffer.from(data).toString('base64');
-              } else {
-                let binary = '';
-                for (let i = 0; i < data.length; i++) {
-                  binary += String.fromCharCode(data[i]);
-                }
-                data = btoa(binary);
-              }
-            }
-
-            // Sanitize base64 data for images and other large inline data
-            data = isFullAiCaptureEnabled(client) ? data : redactBase64DataUrl(data, mimeType);
-            content.push(buildInlineDataBlock(mimeType, data));
+            content.push(formatInlineDataBlock(part.inlineData, client));
           }
         }
         if (content.length > 0) {
@@ -422,7 +420,7 @@ function addDefaults(params) {
   };
 }
 
-var version = "8.9.2";
+var version = "8.10.0";
 
 const DEFAULT_MAX_DEPTH = 3;
 const MAX_STACK_LINES = 20;
@@ -652,7 +650,9 @@ const captureAiGeneration = async (client, options) => {
         $ai_time_to_first_token: options.timeToFirstToken
       } : {}),
       $ai_trace_id: traceId,
-      $ai_base_url: options.baseURL ?? '',
+      ...(options.baseURL === null ? {} : {
+        $ai_base_url: options.baseURL ?? ''
+      }),
       ...options.properties,
       $ai_tokens_source: getTokensSource(options.properties),
       ...(options.distinctId ? {} : {
@@ -686,10 +686,33 @@ const captureAiGeneration = async (client, options) => {
     }
   } catch (error) {
     // Telemetry failures must never affect the instrumented provider call.
+    try {
+      options.onError?.(error);
+    } catch {
+      // Error reporting must not affect the instrumented provider call either.
+    }
     console.warn('[PostHog AI] Failed to capture generation telemetry:', error);
   }
 };
 
+/** Map Gemini usage metadata to PostHog's provider-agnostic token fields. */
+function mapGeminiUsage(metadata, additionalUsage = {}) {
+  return {
+    inputTokens: metadata?.promptTokenCount ?? 0,
+    outputTokens: metadata?.candidatesTokenCount ?? 0,
+    reasoningTokens: metadata?.thoughtsTokenCount ?? 0,
+    cacheReadInputTokens: metadata?.cachedContentTokenCount ?? 0,
+    // Gemini counts cachedContentTokenCount inside promptTokenCount, so declare
+    // the accounting model rather than leaving ingestion to infer it. Under
+    // explicit context caching the two measurements can differ by a few percent.
+    ...(metadata?.cachedContentTokenCount ? {
+      cacheReportingExclusive: false
+    } : {}),
+    ...additionalUsage,
+    rawUsage: metadata
+  };
+}
+
 class PostHogGoogleGenAI {
   constructor(config) {
     const {
@@ -728,21 +751,9 @@ class WrappedModels {
         baseURL: 'https://generativelanguage.googleapis.com',
         modelParameters: getModelParams(params),
         httpStatus: 200,
-        usage: {
-          inputTokens: metadata?.promptTokenCount ?? 0,
-          outputTokens: metadata?.candidatesTokenCount ?? 0,
-          reasoningTokens: metadata?.thoughtsTokenCount ?? 0,
-          cacheReadInputTokens: metadata?.cachedContentTokenCount ?? 0,
-          // Gemini counts cachedContentTokenCount inside promptTokenCount, so declare the
-          // accounting model rather than leaving ingestion to infer it. Under explicit
-          // context caching the two counts come from separate measurements and can disagree
-          // by a few percent, which makes inference from the counts alone unreliable.
-          ...(metadata?.cachedContentTokenCount ? {
-            cacheReportingExclusive: false
-          } : {}),
-          webSearchCount: calculateGoogleWebSearchCount(response),
-          rawUsage: metadata
-        },
+        usage: mapGeminiUsage(metadata, {
+          webSearchCount: calculateGoogleWebSearchCount(response)
+        }),
         stopReason: finishReason ?? undefined,
         tools: availableTools
       });
@@ -843,20 +854,9 @@ class WrappedModels {
 
         // Update usage metadata - handle both old and new field names
         if (chunk.usageMetadata) {
-          const metadata = chunk.usageMetadata;
-          usage = {
-            inputTokens: metadata.promptTokenCount ?? 0,
-            outputTokens: metadata.candidatesTokenCount ?? 0,
-            reasoningTokens: metadata.thoughtsTokenCount ?? 0,
-            cacheReadInputTokens: metadata.cachedContentTokenCount ?? 0,
-            // See the non-streaming path: Gemini counts cachedContentTokenCount inside
-            // promptTokenCount, so the accounting model is declared rather than inferred.
-            ...(metadata.cachedContentTokenCount ? {
-              cacheReportingExclusive: false
-            } : {}),
-            webSearchCount: usage.webSearchCount,
-            rawUsage: metadata
-          };
+          usage = mapGeminiUsage(chunk.usageMetadata, {
+            webSearchCount: usage.webSearchCount
+          });
         }
         yield chunk;
       }