@crewdle/mist-connector-openai 1.0.23 → 1.0.24

package/dist/models/OpenAIGenerativeAIWorkerConnector.js

@@ -246,11 +246,19 @@ Only output the missing remainder. Do not restart or re-open tags already emitte
     /**
      * OpenAI reports cached prompt tokens inside the input token total; split
      * them out so billing can rate the cached share via the model's tokenRates.
+     * Image models (gpt-image) also detail text vs image input tokens — bucket
+     * the text share so docs whose base input SKU is the image price can rate
+     * it (e.g. gpt-image-2 `text: 0.625`). Unrated buckets bill at base, so
+     * this is a no-op for plain text models. cached can overlap the modality
+     * details, hence the clamp.
      */
     usageBuckets(usage) {
         const total = usage?.input_tokens ?? 0;
-        const cached = usage?.input_tokens_details?.cached_tokens ?? 0;
-        return { input: { base: total - cached, cached }, output: { base: usage?.output_tokens ?? 0 } };
+        const details = usage?.input_tokens_details ?? {};
+        const cached = details.cached_tokens ?? 0;
+        const text = details.text_tokens ?? 0;
+        const audio = details.audio_tokens ?? 0;
+        return { input: { base: Math.max(0, total - cached - text - audio), cached, text, audio }, output: { base: usage?.output_tokens ?? 0 } };
     }
     /**
      * whisper-1 reports no token usage; verbose_json carries the audio duration

package/dist/types/models/OpenAIGenerativeAIWorkerConnector.d.ts

@@ -11,6 +11,11 @@ export declare class OpenAIGenerativeAIWorkerConnector implements IGenerativeAIW
     /**
      * OpenAI reports cached prompt tokens inside the input token total; split
      * them out so billing can rate the cached share via the model's tokenRates.
+     * Image models (gpt-image) also detail text vs image input tokens — bucket
+     * the text share so docs whose base input SKU is the image price can rate
+     * it (e.g. gpt-image-2 `text: 0.625`). Unrated buckets bill at base, so
+     * this is a no-op for plain text models. cached can overlap the modality
+     * details, hence the clamp.
      */
     private usageBuckets;
     /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@crewdle/mist-connector-openai",
-  "version": "1.0.23",
+  "version": "1.0.24",
   "description": "",
   "main": "dist/index.js",
   "types": "dist/types/index.d.ts",