@posthog/ai 8.7.0 → 8.8.0

package/dist/openai/index.mjs

@@ -9,6 +9,33 @@ 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 {
@@ -114,7 +141,6 @@ 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);
   }
@@ -158,22 +184,24 @@ 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();
-const sanitizeOpenAI = data => redactor.redact(data);
-const sanitizeOpenAIResponse = data => redactor.redact(data);
+const sanitize = (data, client) => isFullAiCaptureEnabled(client) ? data : redactor.redact(data);
+const sanitizeOpenAI = (data, client) => sanitize(data, client);
+const sanitizeOpenAIResponse = (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';
 
@@ -544,7 +572,7 @@ function formatOpenAIResponsesInput(input, instructions) {
   return messages;
 }
 
-var version = "8.7.0";
+var version = "8.8.0";
 
 const DEFAULT_MAX_DEPTH = 3;
 const MAX_STACK_LINES = 20;
@@ -711,6 +739,10 @@ const captureAiGeneration$1 = 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
@@ -721,6 +753,18 @@ const captureAiGeneration$1 = 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
       } : {}),
@@ -777,9 +821,9 @@ const captureAiGeneration$1 = 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.
@@ -1446,8 +1490,8 @@ let WrappedCompletions$1 = class WrappedCompletions extends AzureOpenAI.Chat.Com
                 ...posthogParams,
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'azure',
-                input: sanitizeOpenAI(openAIParams.messages),
-                output: sanitizeOpenAIResponse(formattedOutput),
+                input: sanitizeOpenAI(openAIParams.messages, this.phClient),
+                output: sanitizeOpenAIResponse(formattedOutput, this.phClient),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -1464,7 +1508,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends AzureOpenAI.Chat.Com
                 ...posthogParams,
                 model: openAIParams.model,
                 provider: 'azure',
-                input: sanitizeOpenAI(openAIParams.messages),
+                input: sanitizeOpenAI(openAIParams.messages, this.phClient),
                 output: [],
                 latency: 0,
                 baseURL: this.baseURL,
@@ -1503,8 +1547,8 @@ let WrappedCompletions$1 = class WrappedCompletions extends AzureOpenAI.Chat.Com
             ...posthogParams,
             model: openAIParams.model ?? result.model,
             provider: 'azure',
-            input: sanitizeOpenAI(openAIParams.messages),
-            output: sanitizeOpenAIResponse(formatResponseOpenAI(result)),
+            input: sanitizeOpenAI(openAIParams.messages, this.phClient),
+            output: sanitizeOpenAIResponse(formatResponseOpenAI(result), this.phClient),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -1530,7 +1574,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends AzureOpenAI.Chat.Com
           ...posthogParams,
           model: openAIParams.model,
           provider: 'azure',
-          input: sanitizeOpenAI(openAIParams.messages),
+          input: sanitizeOpenAI(openAIParams.messages, this.phClient),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -1672,8 +1716,8 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
                 ...posthogParams,
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'azure',
-                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-                output: sanitizeOpenAIResponse(finalContent),
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
+                output: sanitizeOpenAIResponse(finalContent, this.phClient),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -1695,7 +1739,7 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
                 ...posthogParams,
                 model: openAIParams.model,
                 provider: 'azure',
-                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
                 output: [],
                 latency: 0,
                 baseURL: this.baseURL,
@@ -1735,8 +1779,8 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
             ...posthogParams,
             model: openAIParams.model ?? result.model,
             provider: 'azure',
-            input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-            output: sanitizeOpenAIResponse(result.output),
+            input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
+            output: sanitizeOpenAIResponse(result.output, this.phClient),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -1765,7 +1809,7 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
           ...posthogParams,
           model: openAIParams.model,
           provider: 'azure',
-          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -1851,8 +1895,8 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
         ...posthogParams,
         model: openAIParams.model ?? result.model,
         provider: 'azure',
-        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-        output: sanitizeOpenAIResponse(result.output),
+        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
+        output: sanitizeOpenAIResponse(result.output, this.phClient),
         latency,
         baseURL: this.baseURL,
         modelParameters: getModelParams(body, result.service_tier),
@@ -1879,7 +1923,7 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
         ...posthogParams,
         model: openAIParams.model,
         provider: 'azure',
-        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
         output: [],
         latency: 0,
         baseURL: this.baseURL,
@@ -2155,8 +2199,8 @@ class WrappedCompletions extends Completions {
                 ...posthogParams,
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'openai',
-                input: sanitizeOpenAI(openAIParams.messages),
-                output: sanitizeOpenAIResponse(formattedOutput),
+                input: sanitizeOpenAI(openAIParams.messages, this.phClient),
+                output: sanitizeOpenAIResponse(formattedOutput, this.phClient),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -2183,7 +2227,7 @@ class WrappedCompletions extends Completions {
                 ...posthogParams,
                 model: openAIParams.model,
                 provider: 'openai',
-                input: sanitizeOpenAI(openAIParams.messages),
+                input: sanitizeOpenAI(openAIParams.messages, this.phClient),
                 output: [],
                 latency: 0,
                 baseURL: this.baseURL,
@@ -2224,8 +2268,8 @@ class WrappedCompletions extends Completions {
             ...posthogParams,
             model: openAIParams.model ?? result.model,
             provider: 'openai',
-            input: sanitizeOpenAI(openAIParams.messages),
-            output: sanitizeOpenAIResponse(formattedOutput),
+            input: sanitizeOpenAI(openAIParams.messages, this.phClient),
+            output: sanitizeOpenAIResponse(formattedOutput, this.phClient),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -2255,7 +2299,7 @@ class WrappedCompletions extends Completions {
           ...posthogParams,
           model: openAIParams.model,
           provider: 'openai',
-          input: sanitizeOpenAI(openAIParams.messages),
+          input: sanitizeOpenAI(openAIParams.messages, this.phClient),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -2289,7 +2333,7 @@ class WrappedResponses extends Responses {
       ...posthogParams,
       model: openAIParams.model ?? result.model,
       provider: 'openai',
-      input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+      input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
       output: formatResponseOpenAI({
         output: result.output
       }),
@@ -2412,8 +2456,8 @@ class WrappedResponses extends Responses {
                 ...posthogParams,
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'openai',
-                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-                output: sanitizeOpenAIResponse(finalContent),
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
+                output: sanitizeOpenAIResponse(finalContent, this.phClient),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -2444,7 +2488,7 @@ class WrappedResponses extends Responses {
                 ...posthogParams,
                 model: openAIParams.model,
                 provider: 'openai',
-                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
                 output: [],
                 latency: 0,
                 baseURL: this.baseURL,
@@ -2488,8 +2532,8 @@ class WrappedResponses extends Responses {
             ...posthogParams,
             model: openAIParams.model ?? result.model,
             provider: 'openai',
-            input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-            output: sanitizeOpenAIResponse(formattedOutput),
+            input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
+            output: sanitizeOpenAIResponse(formattedOutput, this.phClient),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -2520,7 +2564,7 @@ class WrappedResponses extends Responses {
           ...posthogParams,
           model: openAIParams.model,
           provider: 'openai',
-          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -2606,8 +2650,8 @@ class WrappedResponses extends Responses {
         ...posthogParams,
         model: openAIParams.model ?? result.model,
         provider: 'openai',
-        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-        output: sanitizeOpenAIResponse(result.output),
+        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
+        output: sanitizeOpenAIResponse(result.output, this.phClient),
         latency,
         baseURL: this.baseURL,
         modelParameters: getModelParams(body, result.service_tier),
@@ -2634,7 +2678,7 @@ class WrappedResponses extends Responses {
         ...posthogParams,
         model: openAIParams.model,
         provider: 'openai',
-        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input, this.phClient), openAIParams.instructions),
         output: [],
         latency: 0,
         baseURL: this.baseURL,
@@ -2779,7 +2823,7 @@ class WrappedTranscriptions extends Transcriptions {
                 model: openAIParams.model,
                 provider: 'openai',
                 input: openAIParams.prompt,
-                output: sanitizeOpenAIResponse(finalContent),
+                output: sanitizeOpenAIResponse(finalContent, this.phClient),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -2824,7 +2868,7 @@ class WrappedTranscriptions extends Transcriptions {
             model: openAIParams.model,
             provider: 'openai',
             input: openAIParams.prompt,
-            output: sanitizeOpenAIResponse(result.text),
+            output: sanitizeOpenAIResponse(result.text, this.phClient),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body),