@posthog/ai 8.7.0 → 8.8.0

package/dist/gemini/index.mjs

@@ -8,6 +8,33 @@ const isString = value => {
   return typeof value === 'string';
 };
 
+/** @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 {
@@ -113,7 +140,6 @@ 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);
   }
@@ -157,24 +183,26 @@ 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 sanitizeGemini = data => redactor.redact(data);
+const sanitize = (data, client) => isFullAiCaptureEnabled(client) ? data : redactor.redact(data);
+const sanitizeGemini = (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';
 }
 const STRING_FORMAT = 'utf8';
 
@@ -248,7 +276,7 @@ const buildInlineDataBlock = (mimeType, data) => {
     }
   };
 };
-const formatResponseGemini = response => {
+const formatResponseGemini = (response, client) => {
   const output = [];
   if (response.candidates && Array.isArray(response.candidates)) {
     for (const candidate of response.candidates) {
@@ -287,7 +315,7 @@ const formatResponseGemini = response => {
             }
 
             // Sanitize base64 data for images and other large inline data
-            data = redactBase64DataUrl(data, mimeType);
+            data = isFullAiCaptureEnabled(client) ? data : redactBase64DataUrl(data, mimeType);
             content.push(buildInlineDataBlock(mimeType, data));
           }
         }
@@ -390,7 +418,7 @@ function addDefaults(params) {
   };
 }
 
-var version = "8.7.0";
+var version = "8.8.0";
 
 const DEFAULT_MAX_DEPTH = 3;
 const MAX_STACK_LINES = 20;
@@ -557,6 +585,10 @@ 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
@@ -567,6 +599,18 @@ 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
       } : {}),
@@ -623,9 +667,9 @@ 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.
@@ -666,7 +710,7 @@ class WrappedModels {
         model: geminiParams.model,
         provider: 'gemini',
         input: this.formatInputForPostHog(geminiParams),
-        output: formatResponseGemini(response),
+        output: formatResponseGemini(response, this.phClient),
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
         modelParameters: getModelParams(params),
@@ -676,6 +720,13 @@ class WrappedModels {
           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
         },
@@ -789,6 +840,11 @@ class WrappedModels {
             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
           };
@@ -1021,7 +1077,7 @@ class WrappedModels {
     return null;
   }
   formatInputForPostHog(params) {
-    const sanitized = sanitizeGemini(params.contents);
+    const sanitized = sanitizeGemini(params.contents, this.phClient);
     const messages = this.formatInput(sanitized);
     const systemInstruction = this.extractSystemInstruction(params);
     if (systemInstruction) {