refactor(runner): replace gpt-tokenizer with tiktoken

Author: danielsogl

package.json

@@ -71,17 +71,17 @@
     "file-type": "^21.0.0",
     "genkit": "^1.19.1",
     "genkitx-anthropic": "0.23.1",
-    "gpt-tokenizer": "^3.0.1",
     "handlebars": "^4.7.8",
     "limiter": "^3.0.0",
     "marked": "^16.1.1",
     "node-fetch": "^3.3.2",
     "p-queue": "^8.1.0",
     "puppeteer": "^24.10.1",
     "sass": "^1.89.2",
-    "stylelint": "^16.21.1",
     "strict-csp": "^1.1.1",
+    "stylelint": "^16.21.1",
     "stylelint-config-recommended-scss": "^16.0.0",
+    "tiktoken": "^1.0.22",
     "tinyglobby": "^0.2.14",
     "tsx": "^4.20.3",
     "typescript": "^5.8.3",

pnpm-lock.yaml

@@ -1,4 +1,3 @@
-import { ChatMessage } from 'gpt-tokenizer/GptEncoding';
 import { GenkitPluginV2 } from 'genkit/plugin';
 import { openAI } from '@genkit-ai/compat-oai/openai';
 import { RateLimiter } from 'limiter';
@@ -7,8 +6,7 @@ import {
   PromptDataForCounting,
   RateLimitConfig,
 } from '../model-provider.js';
-import o3 from 'gpt-tokenizer/model/o3';
-import o4Mini from 'gpt-tokenizer/model/o4-mini';
+import { encoding_for_model } from 'tiktoken';
 
 export class OpenAiModelProvider extends GenkitModelProvider {
   readonly apiKeyVariableName = 'OPENAI_API_KEY';
@@ -30,8 +28,17 @@ export class OpenAiModelProvider extends GenkitModelProvider {
         tokensPerInterval: 30_000 * 0.75, // *0.75 to be more resilient to token count deviations
         interval: 1000 * 60 * 1.5, // Refresh tokens after 1.5 minutes to be on the safe side.
       }),
-      countTokens: async (prompt) =>
-        o3.countTokens(this.genkitPromptToOpenAi(prompt)),
+      countTokens: async (prompt) => {
+        const encoding = encoding_for_model('gpt-4o');
+        try {
+          const messages = this.genkitPromptToOpenAi(prompt);
+          const text = messages.map(m => `${m.role}: ${m.content}`).join('\n');
+          const tokens = encoding.encode(text);
+          return tokens.length;
+        } finally {
+          encoding.free();
+        }
+      },
     },
     // See https://platform.openai.com/docs/models/o4-mini
     'openai/o4-mini': {
@@ -43,8 +50,17 @@ export class OpenAiModelProvider extends GenkitModelProvider {
         tokensPerInterval: 100_000 * 0.75, // *0.75 to be more resilient to token count deviations
         interval: 1000 * 60 * 1.5, // Refresh tokens after 1.5 minutes to be on the safe side.
       }),
-      countTokens: async (prompt) =>
-        o4Mini.countTokens(this.genkitPromptToOpenAi(prompt)),
+      countTokens: async (prompt) => {
+        const encoding = encoding_for_model('gpt-4o-mini');
+        try {
+          const messages = this.genkitPromptToOpenAi(prompt);
+          const text = messages.map(m => `${m.role}: ${m.content}`).join('\n');
+          const tokens = encoding.encode(text);
+          return tokens.length;
+        } finally {
+          encoding.free();
+        }
+      },
     },
     // See: https://platform.openai.com/docs/models/gpt-5
     'openai/gpt-5': {
@@ -56,10 +72,17 @@ export class OpenAiModelProvider extends GenkitModelProvider {
         tokensPerInterval: 30_000 * 0.75, // *0.75 to be more resilient to token count deviations
         interval: 1000 * 60 * 1.5, // Refresh tokens after 1.5 minutes to be on the safe side.
       }),
-      // TODO: at the time of writing, the `gpt-tokenizer` doesn't support gpt-5.
-      // See https://github.com/niieani/gpt-tokenizer/issues/73
-      countTokens: async (prompt) =>
-        o3.countTokens(this.genkitPromptToOpenAi(prompt)),
+      countTokens: async (prompt) => {
+        const encoding = encoding_for_model('gpt-5');
+        try {
+          const messages = this.genkitPromptToOpenAi(prompt);
+          const text = messages.map(m => `${m.role}: ${m.content}`).join('\n');
+          const tokens = encoding.encode(text);
+          return tokens.length;
+        } finally {
+          encoding.free();
+        }
+      },
     },
   };
 
@@ -72,8 +95,8 @@ export class OpenAiModelProvider extends GenkitModelProvider {
     return {};
   }
 
-  private genkitPromptToOpenAi(prompt: PromptDataForCounting): ChatMessage[] {
-    const openAiPrompt: string | ChatMessage[] = [];
+  private genkitPromptToOpenAi(prompt: PromptDataForCounting): Array<{ role: string; content: string }> {
+    const openAiPrompt: Array<{ role: string; content: string }> = [];
     for (const part of prompt.messages) {
       for (const c of part.content) {
         openAiPrompt.push({
@@ -82,6 +105,6 @@ export class OpenAiModelProvider extends GenkitModelProvider {
         });
       }
     }
-    return [...openAiPrompt, { content: prompt.prompt }];
+    return [...openAiPrompt, { role: 'user', content: prompt.prompt }];
   }
 }