import csv import openai import os import random import re # consistent use random.seed(1929) # get an OpenAI API key and `export OPENAI_API_KEY=sk---` client = openai.OpenAI(api_key=os.getenv('OPENAI_API_KEY')) model = "o4-mini" # dynamic list length returning "1, 2, or 3"; "1 or 2" def natural_enum(n: int) -> str: nums = [str(i) for i in range(1, n + 1)] if n == 1: return nums[0] return ", ".join(nums[:-1]) + f", or {nums[-1]}" def shuffle_with_index_tracking(images, best_index=0, second_index=1): copy = [i for i in images] random.shuffle(copy) return copy, copy.index(images[0]), copy.index(images[1]) def ask_openai_best_outfit(post, images): image_messages = [ { "type": "input_image", "image_url": url } for url in images ] prompt = ( "You're a fashion-savvy assistant. Here are images of different outfits. " f"{post['title']} {post['selftext']} " f"Respond only with the number of the best outfit: {natural_enum(len(images))}." ) try: response = client.responses.create( model=model, input=[ {"role": "user", "content": [ {"type": "input_text", "text": prompt}, *image_messages ]} ], ) except openai.BadRequestError as e: raise Exception("There is an issue possibly retrieving images for the post. It might have been deleted?") # for this model, output[0] is any reasoning text = response.output[1].content[0].text match = re.search(r"\d+", text) if match is not None: return int(match[0]) raise ValueError(f"Invalid response: {text}") posts = csv.DictReader(open('./dataset.csv', 'r')) score = 0.0 items = 0 for post in posts: items += 1 print(post['title']) shuffled_images, best_after_shuffle, second_after_shuffle = shuffle_with_index_tracking(post["images"].split(",")) # print(f"{len(shuffled_images)} options") # this subtracts 1 so AI returning "1" is the 0th image predicted_index = ask_openai_best_outfit(post, shuffled_images) - 1 if predicted_index == best_after_shuffle: score += 1.0 elif predicted_index == second_after_shuffle: score += float(post["secondChoice"]) / float(post["firstChoiceVotes"]) print(f"Total score: {score}/{items}")