@posthog/ai 8.6.4 → 8.6.5

package/dist/openai/index.mjs

@@ -41,14 +41,16 @@ const FILE_FAMILY_TYPES = new Set(['file', 'input_file', 'document', 'media', 'f
 const KNOWN_AUDIO_FORMATS = new Set(['wav', 'mp3', 'ogg', 'flac', 'm4a', 'aac', 'webm']);
 class MediaTypeContext {
   static EMPTY = new MediaTypeContext(undefined, undefined);
-  constructor(parent, key) {
+  constructor(parent, key, explicitMediaType) {
     this.parent = parent;
     this.key = key;
+    this.explicitMediaType = explicitMediaType;
   }
   inferMediaType() {
     return this.inferFromSiblingMime() ?? this.inferFromSiblingFormat() ?? this.inferFromParentType() ?? this.inferFromKey();
   }
   inferFromSiblingMime() {
+    if (this.explicitMediaType) return this.explicitMediaType;
     if (!this.parent) return undefined;
     for (const hint of MIME_HINT_KEYS) {
       const v = this.parent[hint];
@@ -83,7 +85,13 @@ class MediaTypeContext {
     if (key.includes('file') || key.includes('document')) return 'application/octet-stream';
     return undefined;
   }
+  hasExplicitBinaryMediaType() {
+    if (!this.explicitMediaType && (!this.parent || !this.key || !STRONG_CONTEXT_KEYS.has(this.key))) return false;
+    const mediaType = this.inferFromSiblingMime();
+    return mediaType !== undefined && !mediaType.toLowerCase().startsWith('text/');
+  }
   signalsBinary() {
+    if (this.explicitMediaType) return true;
     if (this.parent) {
       for (const hint of MIME_HINT_KEYS) {
         if (typeof this.parent[hint] === 'string') return true;
@@ -105,10 +113,10 @@ class BinaryContentRedactor {
   constructor(recognizer = new Base64Recognizer()) {
     this.recognizer = recognizer;
   }
-  redact(value) {
+  redact(value, mediaType) {
     if (this.isMultimodalEnabled()) return value;
     this.visited = new WeakSet();
-    return this.walk(value, MediaTypeContext.EMPTY);
+    return this.walk(value, mediaType ? new MediaTypeContext(undefined, undefined, mediaType) : MediaTypeContext.EMPTY);
   }
   walk(value, ctx) {
     if (value === null || value === undefined) return value;
@@ -132,8 +140,10 @@ class BinaryContentRedactor {
     return out;
   }
   redactString(value, ctx) {
-    const minLength = ctx.signalsBinary() ? STRONG_CONTEXT_MIN_LENGTH : WEAK_CONTEXT_MIN_LENGTH;
-    const recognition = this.recognizer.recognize(value, minLength);
+    const hasExplicitBinaryMediaType = ctx.hasExplicitBinaryMediaType();
+    const recognitionValue = hasExplicitBinaryMediaType ? value.replace(/[\r\n]/g, '') : value;
+    const minLength = hasExplicitBinaryMediaType ? Math.min(recognitionValue.length, STRONG_CONTEXT_MIN_LENGTH) : ctx.signalsBinary() ? STRONG_CONTEXT_MIN_LENGTH : WEAK_CONTEXT_MIN_LENGTH;
+    const recognition = this.recognizer.recognize(recognitionValue, minLength);
     switch (recognition.kind) {
       case 'data-url':
         return this.placeholderFor(recognition.mediaType);
@@ -534,7 +544,7 @@ function formatOpenAIResponsesInput(input, instructions) {
   return messages;
 }
 
-var version = "8.6.4";
+var version = "8.6.5";
 
 const DEFAULT_MAX_DEPTH = 3;
 const MAX_STACK_LINES = 20;
@@ -834,7 +844,7 @@ function extractRequestId(result) {
  * Assembles the `$ai_provider_metadata` blob for OpenAI / Azure OpenAI events.
  * Provider-specific fields (system fingerprint, request id) live here rather
  * than in the shared, provider-agnostic `$ai_*` namespace. Only keys with a
- * truthy value are included, and `undefined` is returned when there is nothing
+ * meaningful value are included, and `undefined` is returned when there is nothing
  * to report so the property can be omitted from the event entirely.
  */
 function buildProviderMetadata(fields) {
@@ -845,8 +855,33 @@ function buildProviderMetadata(fields) {
   if (fields.requestId) {
     metadata.request_id = fields.requestId;
   }
+  if (fields.incompleteDetails != null) {
+    metadata.incomplete_details = fields.incompleteDetails;
+  }
   return Object.keys(metadata).length > 0 ? metadata : undefined;
 }
+const TERMINAL_RESPONSE_STATUSES = new Set(['completed', 'failed', 'cancelled', 'incomplete']);
+
+/**
+ * Checks whether a Responses API response has reached a status that should
+ * produce a final `$ai_generation` event.
+ */
+function isTerminalResponse(response) {
+  return !!response?.status && TERMINAL_RESPONSE_STATUSES.has(response.status);
+}
+
+/**
+ * Returns an isolated copy of a failed Responses API error for `$ai_error`, or
+ * creates a fallback error when the provider omitted failure details.
+ */
+function getResponseFailure(response) {
+  if (response?.status !== 'failed') {
+    return undefined;
+  }
+  return response.error ? {
+    ...response.error
+  } : new Error(`OpenAI response ${response.id} failed without error details`);
+}
 
 function addRequestId(result, response, requestIdHeader) {
   if (!result || typeof result !== 'object' || Array.isArray(result)) {
@@ -1326,7 +1361,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends AzureOpenAI.Chat.Com
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'azure',
                 input: sanitizeOpenAI(openAIParams.messages),
-                output: formattedOutput,
+                output: sanitizeOpenAIResponse(formattedOutput),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -1382,8 +1417,8 @@ let WrappedCompletions$1 = class WrappedCompletions extends AzureOpenAI.Chat.Com
             ...posthogParams,
             model: openAIParams.model ?? result.model,
             provider: 'azure',
-            input: openAIParams.messages,
-            output: formatResponseOpenAI(result),
+            input: sanitizeOpenAI(openAIParams.messages),
+            output: sanitizeOpenAIResponse(formatResponseOpenAI(result)),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -1408,7 +1443,7 @@ let WrappedCompletions$1 = class WrappedCompletions extends AzureOpenAI.Chat.Com
           ...posthogParams,
           model: openAIParams.model,
           provider: 'azure',
-          input: openAIParams.messages,
+          input: sanitizeOpenAI(openAIParams.messages),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -1464,6 +1499,7 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
                 inputTokens: 0,
                 outputTokens: 0
               };
+              let terminalResponse;
               for await (const chunk of stream1) {
                 // Track first token time on content delta events
                 if (firstTokenTime === undefined && isResponseTokenChunk(chunk)) {
@@ -1480,9 +1516,10 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
                   if (chunk.response.service_tier != null) {
                     serviceTierFromResponse = chunk.response.service_tier;
                   }
-                }
-                if (chunk.type === 'response.completed' && 'response' in chunk && chunk.response?.output && chunk.response.output.length > 0) {
-                  finalContent = chunk.response.output;
+                  if (isTerminalResponse(chunk.response)) {
+                    terminalResponse = chunk.response;
+                    finalContent = chunk.response.output ?? [];
+                  }
                 }
                 if ('response' in chunk && chunk.response?.usage) {
                   usage = {
@@ -1499,22 +1536,27 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
                 ...posthogParams,
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'azure',
-                input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
-                output: finalContent,
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+                output: sanitizeOpenAIResponse(finalContent),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
                 modelParameters: getModelParams(body, serviceTierFromResponse),
                 httpStatus: 200,
                 usage,
-                completionId: completionIdFromResponse
+                stopReason: terminalResponse?.status ?? undefined,
+                completionId: completionIdFromResponse,
+                providerMetadata: buildProviderMetadata({
+                  incompleteDetails: terminalResponse?.incomplete_details
+                }),
+                error: getResponseFailure(terminalResponse)
               });
             } catch (error) {
               await captureAiGeneration(this.phClient, {
                 ...posthogParams,
                 model: openAIParams.model,
                 provider: 'azure',
-                input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+                input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
                 output: [],
                 latency: 0,
                 baseURL: this.baseURL,
@@ -1547,8 +1589,8 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
             ...posthogParams,
             model: openAIParams.model ?? result.model,
             provider: 'azure',
-            input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
-            output: result.output,
+            input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+            output: sanitizeOpenAIResponse(result.output),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -1559,10 +1601,13 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
               reasoningTokens: result.usage?.output_tokens_details?.reasoning_tokens ?? 0,
               cacheReadInputTokens: result.usage?.input_tokens_details?.cached_tokens ?? 0
             },
+            stopReason: result.status ?? undefined,
             completionId: result.id,
             providerMetadata: buildProviderMetadata({
-              requestId: extractRequestId(result)
-            })
+              requestId: extractRequestId(result),
+              incompleteDetails: result.incomplete_details
+            }),
+            error: getResponseFailure(result)
           });
         }
         return result;
@@ -1572,7 +1617,7 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
           ...posthogParams,
           model: openAIParams.model,
           provider: 'azure',
-          input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+          input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
           output: [],
           latency: 0,
           baseURL: this.baseURL,
@@ -1602,8 +1647,8 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
         ...posthogParams,
         model: openAIParams.model ?? result.model,
         provider: 'azure',
-        input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
-        output: result.output,
+        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
+        output: sanitizeOpenAIResponse(result.output),
         latency,
         baseURL: this.baseURL,
         modelParameters: getModelParams(body, result.service_tier),
@@ -1614,10 +1659,13 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
           reasoningTokens: result.usage?.output_tokens_details?.reasoning_tokens ?? 0,
           cacheReadInputTokens: result.usage?.input_tokens_details?.cached_tokens ?? 0
         },
+        stopReason: result.status ?? undefined,
         completionId: result.id,
         providerMetadata: buildProviderMetadata({
-          requestId: extractRequestId(result)
-        })
+          requestId: extractRequestId(result),
+          incompleteDetails: result.incomplete_details
+        }),
+        error: getResponseFailure(result)
       });
       return result;
     }, async error => {
@@ -1625,7 +1673,7 @@ let WrappedResponses$1 = class WrappedResponses extends AzureOpenAI.Responses {
         ...posthogParams,
         model: openAIParams.model,
         provider: 'azure',
-        input: formatOpenAIResponsesInput(openAIParams.input, openAIParams.instructions),
+        input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
         output: [],
         latency: 0,
         baseURL: this.baseURL,
@@ -1715,9 +1763,8 @@ async function captureAiGenerationAfterSuccess(...args) {
     captureAiGenerationInBackground(...args);
   }
 }
-const TERMINAL_RESPONSE_STATUSES = new Set(['completed', 'failed', 'cancelled', 'incomplete']);
 function isPendingBackgroundResponse(params, response) {
-  return params.background === true && !response.usage && !!response.status && !TERMINAL_RESPONSE_STATUSES.has(response.status);
+  return params.background === true && !response.usage && !!response.status && !isTerminalResponse(response);
 }
 class PostHogOpenAI extends OpenAI {
   constructor(config) {
@@ -1905,7 +1952,7 @@ class WrappedCompletions extends Completions {
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'openai',
                 input: sanitizeOpenAI(openAIParams.messages),
-                output: formattedOutput,
+                output: sanitizeOpenAIResponse(formattedOutput),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -1973,7 +2020,7 @@ class WrappedCompletions extends Completions {
             model: openAIParams.model ?? result.model,
             provider: 'openai',
             input: sanitizeOpenAI(openAIParams.messages),
-            output: formattedOutput,
+            output: sanitizeOpenAIResponse(formattedOutput),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -2061,6 +2108,7 @@ class WrappedResponses extends Responses {
                 webSearchCount: 0
               };
               let rawUsageData;
+              let terminalResponse;
               for await (const chunk of stream1) {
                 // Track first token time on content delta events
                 if (firstTokenTime === undefined && isResponseTokenChunk(chunk)) {
@@ -2081,10 +2129,9 @@ class WrappedResponses extends Responses {
                   if (chunkWebSearchCount > 0 && chunkWebSearchCount > (usage.webSearchCount ?? 0)) {
                     usage.webSearchCount = chunkWebSearchCount;
                   }
-                }
-                if (chunk.type === 'response.completed' && 'response' in chunk && chunk.response?.output && chunk.response.output.length > 0) {
-                  finalContent = chunk.response.output;
-                  if (chunk.response.status) {
+                  if (isTerminalResponse(chunk.response)) {
+                    terminalResponse = chunk.response;
+                    finalContent = chunk.response.output ?? [];
                     stopReason = chunk.response.status;
                   }
                 }
@@ -2107,7 +2154,7 @@ class WrappedResponses extends Responses {
                 model: openAIParams.model ?? modelFromResponse,
                 provider: 'openai',
                 input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-                output: finalContent,
+                output: sanitizeOpenAIResponse(finalContent),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -2123,7 +2170,11 @@ class WrappedResponses extends Responses {
                 },
                 stopReason,
                 tools: availableTools,
-                completionId: completionIdFromResponse
+                completionId: completionIdFromResponse,
+                providerMetadata: buildProviderMetadata({
+                  incompleteDetails: terminalResponse?.incomplete_details
+                }),
+                error: getResponseFailure(terminalResponse)
               });
             } catch (error) {
               await captureAiGeneration(this.phClient, {
@@ -2171,7 +2222,7 @@ class WrappedResponses extends Responses {
             model: openAIParams.model ?? result.model,
             provider: 'openai',
             input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-            output: formattedOutput,
+            output: sanitizeOpenAIResponse(formattedOutput),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body, result.service_tier),
@@ -2188,8 +2239,10 @@ class WrappedResponses extends Responses {
             tools: availableTools,
             completionId: result.id,
             providerMetadata: buildProviderMetadata({
-              requestId: extractRequestId(result)
-            })
+              requestId: extractRequestId(result),
+              incompleteDetails: result.incomplete_details
+            }),
+            error: getResponseFailure(result)
           });
         }
         return result;
@@ -2233,7 +2286,7 @@ class WrappedResponses extends Responses {
         model: openAIParams.model ?? result.model,
         provider: 'openai',
         input: formatOpenAIResponsesInput(sanitizeOpenAIResponse(openAIParams.input), openAIParams.instructions),
-        output: result.output,
+        output: sanitizeOpenAIResponse(result.output),
         latency,
         baseURL: this.baseURL,
         modelParameters: getModelParams(body, result.service_tier),
@@ -2248,8 +2301,10 @@ class WrappedResponses extends Responses {
         stopReason: result.status ?? undefined,
         completionId: result.id,
         providerMetadata: buildProviderMetadata({
-          requestId: extractRequestId(result)
-        })
+          requestId: extractRequestId(result),
+          incompleteDetails: result.incomplete_details
+        }),
+        error: getResponseFailure(result)
       });
       return result;
     }, async error => {
@@ -2402,7 +2457,7 @@ class WrappedTranscriptions extends Transcriptions {
                 model: openAIParams.model,
                 provider: 'openai',
                 input: openAIParams.prompt,
-                output: finalContent,
+                output: sanitizeOpenAIResponse(finalContent),
                 latency,
                 timeToFirstToken,
                 baseURL: this.baseURL,
@@ -2447,7 +2502,7 @@ class WrappedTranscriptions extends Transcriptions {
             model: openAIParams.model,
             provider: 'openai',
             input: openAIParams.prompt,
-            output: result.text,
+            output: sanitizeOpenAIResponse(result.text),
             latency,
             baseURL: this.baseURL,
             modelParameters: getModelParams(body),