@posthog/ai 7.16.15 → 7.17.1

package/dist/gemini/index.cjs

@@ -6,8 +6,6 @@ var genai = require('@google/genai');
 var uuid = require('uuid');
 var core = require('@posthog/core');
 
-var version = "7.16.15";
-
 // Type guards for safer type checking
 
 const isString = value => {
@@ -115,6 +113,14 @@ function getTokensSource(posthogProperties) {
   }
   return 'sdk';
 }
+const STRING_FORMAT = 'utf8';
+
+// Reused across calls to avoid per-invocation allocation; truncate() runs
+// hundreds of times for prompts with many parts.
+new TextEncoder();
+new TextDecoder(STRING_FORMAT, {
+  fatal: false
+});
 
 /**
  * Safely converts content to a string, preserving structure for objects/arrays.
@@ -299,73 +305,69 @@ function addDefaults(params) {
     traceId: params.traceId ?? uuid.v4()
   };
 }
-const sendEventWithErrorToPosthog = async ({
-  client,
-  traceId,
-  error,
-  ...args
-}) => {
-  const httpStatus = error && typeof error === 'object' && 'status' in error ? error.status ?? 500 : 500;
-  const properties = {
-    client,
-    traceId,
-    httpStatus,
-    error: JSON.stringify(error),
-    ...args
-  };
-  const enrichedError = error;
-  if (client.options?.enableExceptionAutocapture) {
-    // assign a uuid that can be used to link the trace and exception events
-    const exceptionId = core.uuidv7();
-    client.captureException(error, undefined, {
-      $ai_trace_id: traceId
-    }, exceptionId);
-    enrichedError.__posthog_previously_captured_error = true;
-    properties.exceptionId = exceptionId;
-  }
-  await sendEventToPosthog(properties);
-  return enrichedError;
-};
-const sendEventToPosthog = async ({
-  client,
-  eventType = AIEvent.Generation,
-  distinctId,
-  traceId,
-  model,
-  provider,
-  input,
-  output,
-  latency,
-  timeToFirstToken,
-  baseURL,
-  params,
-  httpStatus = 200,
-  usage = {},
-  error,
-  exceptionId,
-  stopReason,
-  tools,
-  captureImmediate = false
-}) => {
+
+var version = "7.17.1";
+
+/**
+ * Options for `captureAiGeneration`. Mirrors the `$ai_generation` event shape
+ * directly so that any caller — first-party SDK wrappers and external code
+ * alike — produces an identical event.
+ */
+
+/**
+ * Capture an `$ai_generation` (or `$ai_embedding`) event to PostHog.
+ *
+ * This is the canonical primitive that every `@posthog/ai` wrapper
+ * (`withTracing`, `OpenAI`, `Anthropic`, `GoogleGenAI`, …) funnels through, so
+ * external code can use it directly to instrument LLM calls made through
+ * arbitrary clients (Cloudflare Workers AI, custom HTTP, etc.) and get the
+ * same events the SDK wrappers produce.
+ *
+ * When `error` is set, the event is captured as an error. If the error is an
+ * object, it is mutated in place to set `__posthog_previously_captured_error`
+ * so callers can re-throw the original error reference safely.
+ */
+const captureAiGeneration = async (client, options) => {
   if (!client.capture) {
-    return Promise.resolve();
-  }
-  // sanitize input and output for UTF-8 validity
-  const safeInput = sanitizeValues(input);
-  const safeOutput = sanitizeValues(output);
-  const safeError = sanitizeValues(error);
+    return;
+  }
+  const traceId = options.traceId ?? uuid.v4();
+  const eventType = options.eventType ?? AIEvent.Generation;
+  const privacyMode = options.privacyMode ?? false;
+  const usage = options.usage ?? {};
+  const safeInput = sanitizeValues(options.input);
+  const safeOutput = sanitizeValues(options.output);
+  let httpStatus = options.httpStatus;
   let errorData = {};
-  if (error) {
+  if (options.error) {
+    if (httpStatus === undefined) {
+      if (typeof options.error === 'object' && 'status' in options.error && typeof options.error.status === 'number') {
+        httpStatus = options.error.status;
+      } else {
+        httpStatus = 500;
+      }
+    }
+    let exceptionId;
+    if (client.options?.enableExceptionAutocapture) {
+      exceptionId = core.uuidv7();
+      client.captureException(options.error, undefined, {
+        $ai_trace_id: traceId
+      }, exceptionId);
+      if (typeof options.error === 'object') {
+        options.error.__posthog_previously_captured_error = true;
+      }
+    }
     errorData = {
       $ai_is_error: true,
-      $ai_error: safeError,
+      $ai_error: sanitizeValues(JSON.stringify(options.error)),
       $exception_event_id: exceptionId
     };
   }
+  httpStatus = httpStatus ?? 200;
   let costOverrideData = {};
-  if (params.posthogCostOverride) {
-    const inputCostUSD = (params.posthogCostOverride.inputCost ?? 0) * (usage.inputTokens ?? 0);
-    const outputCostUSD = (params.posthogCostOverride.outputCost ?? 0) * (usage.outputTokens ?? 0);
+  if (options.costOverride) {
+    const inputCostUSD = (options.costOverride.inputCost ?? 0) * (usage.inputTokens ?? 0);
+    const outputCostUSD = (options.costOverride.outputCost ?? 0) * (usage.outputTokens ?? 0);
     costOverrideData = {
       $ai_input_cost_usd: inputCostUSD,
       $ai_output_cost_usd: outputCostUSD,
@@ -392,50 +394,48 @@ const sendEventToPosthog = async ({
   const properties = {
     $ai_lib: 'posthog-ai',
     $ai_lib_version: version,
-    $ai_provider: params.posthogProviderOverride ?? provider,
-    $ai_model: params.posthogModelOverride ?? model,
-    $ai_model_parameters: getModelParams(params),
-    $ai_input: withPrivacyMode(client, params.posthogPrivacyMode ?? false, safeInput),
-    $ai_output_choices: withPrivacyMode(client, params.posthogPrivacyMode ?? false, safeOutput),
+    $ai_provider: options.providerOverride ?? options.provider,
+    $ai_model: options.modelOverride ?? options.model,
+    $ai_model_parameters: options.modelParameters ?? {},
+    $ai_input: withPrivacyMode(client, privacyMode, safeInput),
+    $ai_output_choices: withPrivacyMode(client, privacyMode, safeOutput),
     $ai_http_status: httpStatus,
     $ai_input_tokens: usage.inputTokens ?? 0,
     ...(usage.outputTokens !== undefined ? {
       $ai_output_tokens: usage.outputTokens
     } : {}),
     ...additionalTokenValues,
-    $ai_latency: latency,
-    ...(timeToFirstToken !== undefined ? {
-      $ai_time_to_first_token: timeToFirstToken
+    $ai_latency: options.latency ?? 0,
+    ...(options.timeToFirstToken !== undefined ? {
+      $ai_time_to_first_token: options.timeToFirstToken
     } : {}),
     $ai_trace_id: traceId,
-    $ai_base_url: baseURL,
-    ...params.posthogProperties,
-    $ai_tokens_source: getTokensSource(params.posthogProperties),
-    ...(distinctId ? {} : {
+    $ai_base_url: options.baseURL ?? '',
+    ...options.properties,
+    $ai_tokens_source: getTokensSource(options.properties),
+    ...(options.distinctId ? {} : {
       $process_person_profile: false
     }),
-    ...(stopReason ? {
-      $ai_stop_reason: stopReason
+    ...(options.stopReason ? {
+      $ai_stop_reason: options.stopReason
     } : {}),
-    ...(tools ? {
-      $ai_tools: tools
+    ...(options.tools ? {
+      $ai_tools: options.tools
     } : {}),
     ...errorData,
     ...costOverrideData
   };
   const event = {
-    distinctId: distinctId ?? traceId,
+    distinctId: options.distinctId ?? traceId,
     event: eventType,
     properties,
-    groups: params.posthogGroups
+    groups: options.groups
   };
-  if (captureImmediate) {
-    // await capture promise to send single event in serverless environments
+  if (options.captureImmediate) {
     await client.captureImmediate(event);
   } else {
     client.capture(event);
   }
-  return Promise.resolve();
 };
 
 class PostHogGoogleGenAI {
@@ -466,8 +466,7 @@ class WrappedModels {
       const availableTools = extractAvailableToolCalls('gemini', geminiParams);
       const metadata = response.usageMetadata;
       const finishReason = response.candidates?.[0]?.finishReason;
-      await sendEventToPosthog({
-        client: this.phClient,
+      await captureAiGeneration(this.phClient, {
         ...posthogParams,
         model: geminiParams.model,
         provider: 'gemini',
@@ -475,7 +474,7 @@ class WrappedModels {
         output: formatResponseGemini(response),
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
-        params: params,
+        modelParameters: getModelParams(params),
         httpStatus: 200,
         usage: {
           inputTokens: metadata?.promptTokenCount ?? 0,
@@ -491,8 +490,7 @@ class WrappedModels {
       return response;
     } catch (error) {
       const latency = (Date.now() - startTime) / 1000;
-      const enrichedError = await sendEventWithErrorToPosthog({
-        client: this.phClient,
+      await captureAiGeneration(this.phClient, {
         ...posthogParams,
         model: geminiParams.model,
         provider: 'gemini',
@@ -500,14 +498,14 @@ class WrappedModels {
         output: [],
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
-        params: params,
+        modelParameters: getModelParams(params),
         usage: {
           inputTokens: 0,
           outputTokens: 0
         },
-        error: error
+        error
       });
-      throw enrichedError;
+      throw error;
     }
   }
   async *generateContentStream(params) {
@@ -611,8 +609,7 @@ class WrappedModels {
         role: 'assistant',
         content: accumulatedContent
       }] : [];
-      await sendEventToPosthog({
-        client: this.phClient,
+      await captureAiGeneration(this.phClient, {
         ...posthogParams,
         model: geminiParams.model,
         provider: 'gemini',
@@ -621,7 +618,7 @@ class WrappedModels {
         latency,
         timeToFirstToken,
         baseURL: 'https://generativelanguage.googleapis.com',
-        params: params,
+        modelParameters: getModelParams(params),
         httpStatus: 200,
         usage: {
           ...usage,
@@ -633,8 +630,7 @@ class WrappedModels {
       });
     } catch (error) {
       const latency = (Date.now() - startTime) / 1000;
-      const enrichedError = await sendEventWithErrorToPosthog({
-        client: this.phClient,
+      await captureAiGeneration(this.phClient, {
         ...posthogParams,
         model: geminiParams.model,
         provider: 'gemini',
@@ -642,14 +638,14 @@ class WrappedModels {
         output: [],
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
-        params: params,
+        modelParameters: getModelParams(params),
         usage: {
           inputTokens: 0,
           outputTokens: 0
         },
-        error: error
+        error
       });
-      throw enrichedError;
+      throw error;
     }
   }
   async embedContent(params) {
@@ -662,8 +658,7 @@ class WrappedModels {
       const response = await this.client.models.embedContent(geminiParams);
       const latency = (Date.now() - startTime) / 1000;
       const inputTokens = extractEmbeddingTokenCount(response);
-      await sendEventToPosthog({
-        client: this.phClient,
+      await captureAiGeneration(this.phClient, {
         ...posthogParams,
         eventType: AIEvent.Embedding,
         model: geminiParams.model,
@@ -672,7 +667,7 @@ class WrappedModels {
         output: null,
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
-        params: params,
+        modelParameters: getModelParams(params),
         httpStatus: 200,
         usage: {
           inputTokens
@@ -681,8 +676,7 @@ class WrappedModels {
       return response;
     } catch (error) {
       const latency = (Date.now() - startTime) / 1000;
-      const enrichedError = await sendEventWithErrorToPosthog({
-        client: this.phClient,
+      await captureAiGeneration(this.phClient, {
         ...posthogParams,
         eventType: AIEvent.Embedding,
         model: geminiParams.model,
@@ -691,13 +685,13 @@ class WrappedModels {
         output: null,
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
-        params: params,
+        modelParameters: getModelParams(params),
         usage: {
           inputTokens: 0
         },
-        error: error
+        error
       });
-      throw enrichedError;
+      throw error;
     }
   }
   formatPartsAsContentBlocks(parts) {