@posthog/ai 8.9.0 → 8.9.2

package/dist/gemini/index.cjs

@@ -422,7 +422,7 @@ function addDefaults(params) {
   };
 }
 
-var version = "8.9.0";
+var version = "8.9.2";
 
 const DEFAULT_MAX_DEPTH = 3;
 const MAX_STACK_LINES = 20;
@@ -579,15 +579,22 @@ const captureAiGeneration = async (client, options) => {
       };
     }
     httpStatus = httpStatus ?? 200;
-    let costOverrideData = {};
+
+    // A configured price applies only to a count the provider reported, so a call with no
+    // reported usage sends no cost instead of asserting $0. $ai_total_cost_usd sums the sides
+    // that were priced, which makes it the cost of the known side alone when the other side
+    // went unreported: a lower bound on the true total, not an assertion of it.
+    const costOverrideData = {};
     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,
-        $ai_total_cost_usd: inputCostUSD + outputCostUSD
-      };
+      if (usage.inputTokens !== undefined) {
+        costOverrideData.$ai_input_cost_usd = (options.costOverride.inputCost ?? 0) * usage.inputTokens;
+      }
+      if (usage.outputTokens !== undefined) {
+        costOverrideData.$ai_output_cost_usd = (options.costOverride.outputCost ?? 0) * usage.outputTokens;
+      }
+      if (Object.keys(costOverrideData).length > 0) {
+        costOverrideData.$ai_total_cost_usd = (costOverrideData.$ai_input_cost_usd ?? 0) + (costOverrideData.$ai_output_cost_usd ?? 0);
+      }
     }
 
     // The caller's own token counts override the SDK-derived ones further down, via the
@@ -631,7 +638,9 @@ const captureAiGeneration = async (client, options) => {
       $ai_input: safeInput,
       $ai_output_choices: safeOutput,
       $ai_http_status: httpStatus,
-      $ai_input_tokens: usage.inputTokens ?? 0,
+      ...(usage.inputTokens !== undefined ? {
+        $ai_input_tokens: usage.inputTokens
+      } : {}),
       ...(usage.outputTokens !== undefined ? {
         $ai_output_tokens: usage.outputTokens
       } : {}),
@@ -749,10 +758,7 @@ class WrappedModels {
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
         modelParameters: getModelParams(params),
-        usage: {
-          inputTokens: 0,
-          outputTokens: 0
-        },
+        usage: {},
         error
       });
       throw error;
@@ -768,11 +774,10 @@ class WrappedModels {
     let firstTokenTime;
     let stopReason;
     let usage = {
-      inputTokens: 0,
-      outputTokens: 0,
       webSearchCount: 0,
       rawUsage: undefined
     };
+    let errored = false;
     try {
       const stream = await this.client.models.generateContentStream(geminiParams);
       for await (const chunk of stream) {
@@ -855,35 +860,8 @@ class WrappedModels {
         }
         yield chunk;
       }
-      const latency = (Date.now() - startTime) / 1000;
-      const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
-      const availableTools = extractAvailableToolCalls('gemini', geminiParams);
-
-      // Format output similar to formatResponseGemini
-      const output = accumulatedContent.length > 0 ? [{
-        role: 'assistant',
-        content: accumulatedContent
-      }] : [];
-      await captureAiGeneration(this.phClient, {
-        ...posthogParams,
-        model: geminiParams.model,
-        provider: 'gemini',
-        input: this.formatInputForPostHog(geminiParams),
-        output,
-        latency,
-        timeToFirstToken,
-        baseURL: 'https://generativelanguage.googleapis.com',
-        modelParameters: getModelParams(params),
-        httpStatus: 200,
-        usage: {
-          ...usage,
-          webSearchCount: usage.webSearchCount,
-          rawUsage: usage.rawUsage
-        },
-        stopReason,
-        tools: availableTools
-      });
     } catch (error) {
+      errored = true;
       const latency = (Date.now() - startTime) / 1000;
       await captureAiGeneration(this.phClient, {
         ...posthogParams,
@@ -894,13 +872,44 @@ class WrappedModels {
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
         modelParameters: getModelParams(params),
-        usage: {
-          inputTokens: 0,
-          outputTokens: 0
-        },
+        usage,
         error
       });
       throw error;
+    } finally {
+      // A consumer that stops iterating resumes the pending yield as a return,
+      // skipping both the loop tail and the catch. Only a finally runs then, so
+      // the success capture lives here to cover completion and cancellation.
+      if (!errored) {
+        const latency = (Date.now() - startTime) / 1000;
+        const timeToFirstToken = firstTokenTime !== undefined ? (firstTokenTime - startTime) / 1000 : undefined;
+        const availableTools = extractAvailableToolCalls('gemini', geminiParams);
+
+        // Format output similar to formatResponseGemini
+        const output = accumulatedContent.length > 0 ? [{
+          role: 'assistant',
+          content: accumulatedContent
+        }] : [];
+        await captureAiGeneration(this.phClient, {
+          ...posthogParams,
+          model: geminiParams.model,
+          provider: 'gemini',
+          input: this.formatInputForPostHog(geminiParams),
+          output,
+          latency,
+          timeToFirstToken,
+          baseURL: 'https://generativelanguage.googleapis.com',
+          modelParameters: getModelParams(params),
+          httpStatus: 200,
+          usage: {
+            ...usage,
+            webSearchCount: usage.webSearchCount,
+            rawUsage: usage.rawUsage
+          },
+          stopReason,
+          tools: availableTools
+        });
+      }
     }
   }
   async embedContent(params) {
@@ -941,9 +950,7 @@ class WrappedModels {
         latency,
         baseURL: 'https://generativelanguage.googleapis.com',
         modelParameters: getModelParams(params),
-        usage: {
-          inputTokens: 0
-        },
+        usage: {},
         error
       });
       throw error;