import json import os import logging import tqdm import aiofiles from openai import OpenAI from app.service.lookbooks.utils.image_utils import base64_encode_image, generate_text_id from app.service.lookbooks.utils.openai_utils import wait_for_job_completion from app.service.lookbooks.config.config import DOCUMENT_COLLECTION # 设置日志 logger = logging.getLogger() # OpenAI 配置 OPENAI_API_KEY = "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" OPENAI_API_BASE = "https://pangkaichen-openai-prox-98.deno.dev/v1" client = OpenAI( api_key=OPENAI_API_KEY, base_url=OPENAI_API_BASE, ) def create_image_batch_requests( image_list, output_path, prompt="""You are an AI assistant specializing in fashion analysis and tagging for an advanced clothing indexing system. Your task is to analyze images of outfits and provide concise, relevant information. Please structure your response as follows: 1. Brief Summary: Start with a one-sentence summary of the overall style and vibe of the outfit. 2. Tags: Provide relevant tags in the following categories. Use multiple tags where appropriate, separated by commas. Season: [e.g., Spring/Summer, Fall/Winter] Style: [e.g., Minimalist, Bohemian, Streetwear, Business Casual] Occasion: [e.g., Office, Casual, Party, Outdoor] Colors: [List main colors used] Materials: [e.g., Cotton, Denim, Leather, Silk] Key Elements: [Any distinctive fashion elements] 3. Item Descriptions: - Briefly describe each main clothing item (top, bottom, outerwear). Number and categorize each item (e.g., 1. Top: ..., 2. Bottom: ...). Keep descriptions concise but include key details about style, color, and distinctive features. - In a single sentence, list all accessories and their overall effect on the outfit. Ensure your tags and descriptions are accurate, relevant to current fashion trends, and useful for indexing and retrieval purposes.""" ): # 预处理 prompt：移除多余的空白和换行符 prompt = ' '.join(prompt.split()) completed_id = [] if os.path.exists(os.path.join(output_path, "image_description_results.jsonl")): with open(os.path.join(output_path, "image_description_results.jsonl"), "r") as f: for line in f: completed_id.append(json.loads(line)["custom_id"]) tasks = [] id2img = {} for idx, image_filename in enumerate(image_list): image_base64 = base64_encode_image(image_filename) if not image_base64: continue current_id = generate_text_id(image_base64) if current_id in completed_id: continue task = { "custom_id": current_id, "method": "POST", "url": "/v1/chat/completions", "body": { "model": "gpt-4o", "temperature": 1.0, "max_tokens": 500, "messages": [ { "role": "system", "content": prompt }, { "role": "user", "content": [ { "type": "image_url", "image_url": { "url": f"data:image/jpeg;base64,{image_base64}", "detail": "low" } } ] } ] } } tasks.append(task) id2img[current_id] = image_filename logger.info(f"In total {len(tasks)} images to process") if tasks: batch_file_name = os.path.join(output_path, "image_batch_requests.jsonl") with open(batch_file_name, 'w', encoding='utf-8') as file: for obj in tasks: file.write(json.dumps(obj, ensure_ascii=False) + '\n') batch_file = client.files.create( file=open(batch_file_name, "rb"), purpose="batch" ) batch_job = client.batches.create( input_file_id=batch_file.id, endpoint="/v1/chat/completions", completion_window="24h" ) if wait_for_job_completion(client, batch_job.id): output_file_id = client.batches.retrieve(batch_job.id).output_file_id file_response = client.files.content(output_file_id) file_content_str = file_response.read().decode('utf-8') with open(os.path.join(output_path, "image_description_results.jsonl"), "w", encoding='utf-8') as f: for line in file_content_str.splitlines(): if line.strip(): try: result = json.loads(line) image_id = result['custom_id'] caption = result['response']['body']['choices'][0]['message']['content'] output = json.dumps({ "custom_id": image_id, "summary": caption, "url": id2img[image_id] }) f.write(output + '\n') except json.JSONDecodeError as error: logger.error(f"Error parsing: {error} -- at line: {line}") else: logger.error("Job failed") return os.path.join(output_path, "image_description_results.jsonl") else: return None async def process_lookbook_task(lookbook_list, tag, year): """后台异步任务，用于处理 lookbook 并保存到向量数据库""" image_list = [] try: for look_book_path in tqdm.tqdm(lookbook_list): lookbook_name = os.path.splitext(os.path.basename(look_book_path))[0] output_dir = os.path.join("fashion_documents/lookbook/images", lookbook_name) os.makedirs(output_dir, exist_ok=True) if not os.listdir(output_dir): from unstructured.partition.pdf import partition_pdf partition_pdf( filename=look_book_path, extract_images_in_pdf=True, infer_table_structure=False, chunking_strategy="by_title", max_characters=4000, new_after_n_chars=3800, combine_text_under_n_chars=2000, extract_image_block_output_dir=output_dir, ) else: current_images = os.listdir(output_dir) image_list.extend([os.path.join(output_dir, x) for x in current_images]) # 1. 处理图片并生成批量请求 image_description_results_file = create_image_batch_requests(image_list, "fashion_documents/lookbook/results") # 2. 保存结果到向量数据库 if image_description_results_file: save_to_vector_db(image_description_results_file, tag, year) except Exception as e: logger.error(f"Error processing lookbooks: {str(e)}") raise e def save_to_vector_db(image_description_results_file, tag, year): """保存图像描述到向量数据库""" image_ids = set() image_summaries = [] image_metadatas = [] try: with open(image_description_results_file, "r", encoding="utf-8") as f: for image_content in f: image_content = json.loads(image_content) # 确保ID不重复 if image_content["custom_id"] not in image_ids: image_ids.add(image_content["custom_id"]) image_summaries.append(image_content["summary"]) image_metadatas.append({ "data_type": "image", "url": image_content["url"].replace("\\", "/"), "source": "mitu", "tag": tag, "year": year, "gender": "female" }) # 将图像的描述和元数据添加到向量数据库中 DOCUMENT_COLLECTION.add_texts(texts=image_summaries, metadatas=image_metadatas, ids=list(image_ids)) logger.info("Successfully saved data to vector database") except Exception as e: logger.error(f"Error saving to vector database: {e}") raise e