feat: add GPT-4o-mini as default PPR model (#69)

* feat: add GPT-4o-mini model * feat: add generation token limit * feat: set GPT-4o-mini as default PPR scraper model * docs: update CHANGELOG.md
apify-projects · Jul 30, 2024 · f0bb799 · f0bb799
1 parent b34526c
commit f0bb799
Show file tree

Hide file tree

Showing 7 changed files with 44 additions and 5 deletions.
diff --git a/actors/extended-gpt-scraper/.actor/input_schema.json b/actors/extended-gpt-scraper/.actor/input_schema.json
@@ -39,7 +39,8 @@
                 "gpt-4-32k",
                 "text-davinci-003",
                 "gpt-4-turbo",
-                "gpt-4o"
+                "gpt-4o",
+                "gpt-4o-mini"
             ],
             "enumTitles": [
                 "GPT-3.5 Turbo",
@@ -48,7 +49,8 @@
                 "GPT-4 32k",
                 "GTP-3 (davinci)",
                 "GPT-4 Turbo (Preview)",
-                "GPT-4o"
+                "GPT-4o",
+                "gpt-4o-mini"
             ]
         },
         "includeUrlGlobs": {

diff --git a/code/src/entrypoint.ts b/code/src/entrypoint.ts
@@ -46,7 +46,7 @@ function adjustPprScraperInput(input: Input) {
     return {
         ...input,
         skipGptGlobs: [],
-        model: 'gpt-3.5-turbo',
+        model: 'DEFAULT_PPR_SCRAPER',
         openaiApiKey: getOpenAiApiKeyEnvOrFail(),
         actorType: ACTORS.PPR_PRICING,
     };

diff --git a/code/src/models/models.ts b/code/src/models/models.ts
@@ -84,4 +84,27 @@ const OPEN_AI_MODELS: { [modelKey: string]: ModelConfig } = {
             output: 0.015,
         },
     },
+    'gpt-4o-mini': {
+        modelName: 'gpt-4o-mini',
+        maxTokens: 128_000,
+        interface: 'chat',
+        cost: {
+            input: 0.00015,
+            output: 0.0006,
+        },
+    },
+    /**
+     * Default scraper model that we use for the PPR scraper.
+     * - We limit the max tokens here so the price is capped.
+     */
+    DEFAULT_PPR_SCRAPER: {
+        modelName: 'gpt-4o-mini',
+        limitGenerationTokens: true,
+        maxTokens: 10_000,
+        interface: 'chat',
+        cost: {
+            input: 0.00015,
+            output: 0.0006,
+        },
+    },
 } as const;
diff --git a/code/src/models/openai.ts b/code/src/models/openai.ts
@@ -54,13 +54,16 @@ const wrapInOpenaiError = (error: LangchainError): OpenaiAPIError => {
 
 export class OpenAIModelHandler extends GeneralModelHandler<OpenAIModelSettings> {
     async processInstructions(options: ProcessInstructionsOptions<OpenAIModelSettings>) {
-        const { instructions, content, schema, schemaDescription, modelSettings } = options;
+        const { instructions, content, schema, schemaDescription, modelSettings, remainingTokens } = options;
+        const { modelName, limitGenerationTokens, maxTokens } = this.modelConfig;
+
         log.debug(`Calling Openai API with model ${this.modelConfig.modelName}`);
 
         const handleLLMEndCallback = this.handleLLMEndCallbackHandler();
         const modelOptions = {
             ...modelSettings,
-            modelName: this.modelConfig.modelName,
+            modelName,
+            maxTokens: limitGenerationTokens ? remainingTokens : maxTokens,
             callbacks: [{ handleLLMEnd: handleLLMEndCallback.handleLLMEnd }],
         };
 

diff --git a/code/src/routes/crawl-route.ts b/code/src/routes/crawl-route.ts
@@ -152,6 +152,7 @@ export const crawlRoute = async (context: PlaywrightCrawlingContext) => {
     } else {
         log.info(`Processing page ${url} with GPT instruction...`, { contentLength: pageContent.length });
     }
+    const remainingTokens = getNumberOfTextTokens(pageContent) + instructionTokenLength;
 
     try {
         const answerResult = await model.processInstructionsWithRetry({
@@ -160,6 +161,7 @@ export const crawlRoute = async (context: PlaywrightCrawlingContext) => {
             schema,
             schemaDescription,
             modelSettings,
+            remainingTokens,
             apifyClient: Actor.apifyClient,
         });
         answer = answerResult.answer;

diff --git a/code/src/types/model.ts b/code/src/types/model.ts
@@ -7,6 +7,8 @@ export interface ModelConfig {
     maxOutputTokens?: number;
     interface: 'chat' | 'text';
     cost?: Cost;
+    /** Only used for cases where we need to limit the token limit earlier than the actual maximum (e.g. our PPR Actor) */
+    limitGenerationTokens?: true;
 }
 
 export interface Usage {
@@ -29,6 +31,7 @@ export interface ProcessInstructionsOptions<ModelSettings extends object> {
     modelSettings: ModelSettings;
     schema?: AnySchema;
     schemaDescription: string;
+    remainingTokens: number;
 }
 
 export interface ProcessedInstructions {

diff --git a/shared/CHANGELOG.md b/shared/CHANGELOG.md
@@ -1,5 +1,11 @@
 This changelog tracks updates to both GTP Scraper and Extended GPT Scraper actors.
 
+# 2024-07-30
+*Features*
+- Added support for GPT-4o-mini model. (Extended GPT scraper)
+- Set this model as the default one for the the *Pay Per Result* scraper with a set token limit.
+  - With this, the maximum token limit for the *Pay Per Result* scraper was increased by 150%.
+
 # 2024-05-20
 *Features*
 - Added support for GPT-4o model. (Extended GPT Scraper only)