diff --git a/.env b/.env deleted file mode 100644 index 9b14b0a..0000000 --- a/.env +++ /dev/null @@ -1,2 +0,0 @@ -GEMINI_API_KEY=AIzaSyAO4zXFke1bqyrXd9-RGfKJTLerwLSFKww -GOOGLE_APPLICATION_CREDENTIALS="/app/request.json" diff --git a/.gitignore b/.gitignore index 4b7a90f..079b292 100644 --- a/.gitignore +++ b/.gitignore @@ -1,12 +1,6 @@ -.env +/.env .vscode/ -app/core/__pycache__/ -data/db -data/image_data -app/core/data/ +__pycache__/ +data/ .idea/ -*.sqlite3 -*.log -db -*.sqlite -*.png \ No newline at end of file +*.log \ No newline at end of file diff --git a/app/core/config.py b/app/config.py similarity index 80% rename from app/core/config.py rename to app/config.py index 0a5dab9..af3f070 100644 --- a/app/core/config.py +++ b/app/config.py @@ -16,6 +16,8 @@ class Settings(BaseSettings): env_file_encoding='utf-8', extra='ignore' # 忽略环境变量中多余的键 ) + # 调试配饰 + LOCAL: int = Field(default=0, description="是否在本地运行，1表示本地运行，0表示生产环境运行") # Redis 配置 REDIS_HOST: str = Field(default='10.1.1.240', description="Redis服务器地址") @@ -28,10 +30,10 @@ class Settings(BaseSettings): LLM_MODEL_NAME: str = Field(default="gemini-2.5-flash", description="使用的 LLM 模型名称") # 路径配置参数 - DATA_ROOT: str = Field(default="/workspace/lc_stylist_agent/app/core/data", description="数据根目录") - IMAGE_DIR: str = Field(default="/workspace/lc_stylist_agent/app/core/data/image_data", description="图片数据目录") - OUTFIT_OUTPUT_DIR: str = Field(default="/workspace/lc_stylist_agent/app/core/data/outfit_output", description="生成的搭配图片输出目录") - STYLIST_GUIDE_DIR: str = Field(default="/workspace/lc_stylist_agent/app/core/data/stylist_guide", description="风格指南文本目录") + DATA_ROOT: str = Field(default="/workspace/lc_stylist_agent/data", description="数据根目录") + LOCAL_IMAGE_DIR: str = Field(default="/workspace/lc_stylist_agent/Data/image_data", description="图片数据目录") + OUTFIT_OUTPUT_DIR: str = Field(default="/workspace/lc_stylist_agent/data/outfit_output", description="生成的搭配图片输出目录") + STYLIST_GUIDE_DIR: str = Field(default="/workspace/lc_stylist_agent/data/stylist_guide", description="风格指南文本目录") # 向量数据库配置参数 if DEBUG == 1: diff --git a/app/core/__init__.py b/app/core/__init__.py deleted file mode 100644 index e69de29..0000000 diff --git a/app/core/chatbot_agent.py b/app/core/chatbot_agent.py deleted file mode 100644 index 63af21d..0000000 --- a/app/core/chatbot_agent.py +++ /dev/null @@ -1,139 +0,0 @@ -import time -from typing import Dict, List -import asyncio - -from app.core.data_structure import Message, Role -from app.core.llm_interface import AsyncLLMInterface, AsyncGeminiLLM -from app.core.redis_manager import RedisManager -from app.core.system_prompt import BASIC_PROMPT, SUMMARY_PROMPT -from app.core.stylist_agent import AsyncStylistAgent -from app.core.vector_database import VectorDatabase -from app.core.config import settings - - -class ChatbotAgent: - def __init__(self, llm_model: AsyncLLMInterface = None): - self.llm = llm_model if llm_model else AsyncGeminiLLM(model_name=settings.LLM_MODEL_NAME) - self.redis = RedisManager( - host=settings.REDIS_HOST, - port=settings.REDIS_PORT, - db=settings.REDIS_DB, - key_prefix=settings.REDIS_HISTORY_KEY_PREFIX - ) - self.vector_db = VectorDatabase( - vector_db_dir=settings.VECTOR_DB_DIR, - collection_name=settings.COLLECTION_NAME, - embedding_model_name=settings.EMBEDDING_MODEL_NAME - ) - self.stylist_agent_kwages = { - 'local_db': self.vector_db, - 'max_len': 5, - 'outfits_root': settings.OUTFIT_OUTPUT_DIR, - 'image_dir': settings.IMAGE_DIR, - 'stylist_guide_dir': settings.STYLIST_GUIDE_DIR, - 'gemini_model_name': settings.LLM_MODEL_NAME - } - - async def process_query(self, user_id: str, user_message: str) -> str: - """ - 处理用户的最新输入，调用 LLM, 并更新历史记录。 - """ - - # 添加用户消息到历史 - user_msg = Message(role=Role.USER, content=user_message) - chat_history = self.redis.get_history(user_id) - chat_history.append(user_msg) - - # 生成 LLM 回复 - try: - response_text = await self.llm.generate_response(chat_history, system_prompt=BASIC_PROMPT) - except Exception as e: - print(f"LLM 调用失败: {e}") - response_text = "抱歉，系统暂时无法响应，请稍后再试。" - - # 添加助手消息到历史 - if response_text: - assistant_msg = Message(role=Role.ASSISTANT, content=response_text) - else: - assistant_msg = Message(role=Role.ASSISTANT, content="No response generated. Try again later.") - - self.redis.save_message(user_id, user_msg) - self.redis.save_message(user_id, assistant_msg) - - return response_text - - async def get_conversation_summary(self, user_id: str) -> str: - """ - 分析用户的完整会话历史，并打包成一个简洁的需求总结。 - - 这个总结可以直接作为输入 Prompt 传递给 Stylist Agent。` - """ - history_messages = self.redis.get_history(user_id) - input_message = "\n".join([f"{msg.role.value}: {msg.content}" for msg in history_messages]) - - # 临时调用 LLM 或使用本地逻辑生成总结 - summary = await self.llm.generate_response(history=[Message(role=Role.USER, content=input_message)], system_prompt=SUMMARY_PROMPT) - - return summary - - async def recommend_outfit(self, request_summary: str, stylist_name: str, start_outfit: List[Dict[str, str]] = [], num_outfits: int = 1): - """ - 基于用户的对话历史和需求，推荐一套搭配。 - - Args: - request_summary: 用户的request - start_outfit: 可选的初始搭配列表，每个元素包含 'item_id' 和 'category'。 - """ - tasks = [] - for _ in range(num_outfits): - agent = AsyncStylistAgent(**self.stylist_agent_kwages) - task = agent.run_styling_process(request_summary, stylist_name, start_outfit) - tasks.append(task) - print(f"--- Starting {num_outfits} concurrent outfit generation tasks. ---") - - try: - results = await asyncio.gather(*tasks, return_exceptions=True) - - successful_outfits = [] - failed_outfits = [] - for result in results: - if isinstance(result, Exception): - # 任务执行中发生异常 - failed_outfits.append(f"Failed: {result}") - else: - # 任务成功，result 是 run_styling_process 返回的图片路径 - successful_outfits.append(result) - - return { - "successful_outfits": successful_outfits, - "failed_outfits": failed_outfits - } - - except Exception as e: - print(f"An unexpected error occurred during concurrent recommendation: {e}") - return {"error": str(e)} - - -if __name__ == "__main__": - async def run(): - start_time = time.time() - agent = ChatbotAgent() - user_id = "string" - # agent.redis.clear_history(user_id) # 清除历史，便于测试 - # print(await agent.process_query(user_id, "I want a chic outfit for a summer party.")) - # print(await agent.process_query(user_id, "I prefer something floral and light.")) - request_summary = await agent.get_conversation_summary(user_id) - print(f"Conversation Summary:\n{request_summary}") - - recommendation_results = await agent.recommend_outfit(request_summary, stylist_name="crystal", start_outfit=[], num_outfits=4) - - print("\n--- Final Recommendation Results ---") - for i, path in enumerate(recommendation_results.get("successful_outfits", [])): - print(f"✅ Outfit {i + 1} saved to: {path}") - - for error in recommendation_results.get("failed_outfits", []): - print(f"❌ {error}") - print(time.time() - start_time) - - - asyncio.run(run()) diff --git a/app/core/chatbot_agent_stream.py b/app/core/chatbot_agent_stream.py deleted file mode 100644 index d68b8d9..0000000 --- a/app/core/chatbot_agent_stream.py +++ /dev/null @@ -1,163 +0,0 @@ -from google.genai import types -from typing import Dict, List -import asyncio - -from google import genai - -from app.core import system_prompt -from app.core.data_structure import Message, Role -from app.core.llm_interface_stream import AsyncLLMInterface, AsyncGeminiLLM -from app.core.redis_manager import RedisManager -from app.core.system_prompt import BASIC_PROMPT, SUMMARY_PROMPT -from app.core.stylist_agent import AsyncStylistAgent -from app.core.vector_database import VectorDatabase -from app.core.config import settings - - -class ChatbotAgent: - def __init__(self, llm_model: AsyncLLMInterface = None): - self.llm = llm_model if llm_model else AsyncGeminiLLM(model_name=settings.LLM_MODEL_NAME) - self.redis = RedisManager( - host=settings.REDIS_HOST, - port=settings.REDIS_PORT, - db=settings.REDIS_DB, - key_prefix=settings.REDIS_HISTORY_KEY_PREFIX - ) - self.vector_db = VectorDatabase( - vector_db_dir=settings.VECTOR_DB_DIR, - collection_name=settings.COLLECTION_NAME, - embedding_model_name=settings.EMBEDDING_MODEL_NAME - ) - self.stylist_agent_kwages = { - 'local_db': self.vector_db, - 'max_len': 5, - 'outfits_root': settings.OUTFIT_OUTPUT_DIR, - 'image_dir': settings.IMAGE_DIR, - 'stylist_guide_dir': settings.STYLIST_GUIDE_DIR, - 'gemini_model_name': settings.LLM_MODEL_NAME - } - self.gemini_client = genai.Client( - vertexai=True, project='aida-461108', location='us-central1' - ) - - async def process_query(self, user_id: str, user_message: str) -> str: - """ - 处理用户的最新输入，调用 LLM, 并更新历史记录。 - """ - - # 添加用户消息到历史 - user_msg = Message(role=Role.USER, content=user_message) - chat_history = self.redis.get_history(user_id) - chat_history.append(user_msg) - - contents = [] - - for msg in chat_history: - gemini_role = "user" if msg.role == Role.USER else "model" - content = types.Content( - role=gemini_role, - parts=[types.Part.from_text(text=msg.content)] - ) - contents.append(content) - - response_parts = [] - response_stream = await self.gemini_client.aio.models.generate_content_stream( - model='gemini-2.5-flash', - contents=contents, - config=types.GenerateContentConfig( - system_instruction=BASIC_PROMPT, - # temperature=0.3, - ) - ) - async for chunk in response_stream: - # 您可以在这里处理每一个文本块，例如发送给前端 - print(chunk.text, end="", flush=True) - response_parts.append(chunk.text) - # 3. 将所有文本块合并成最终的字符串 - response_text = "".join(response_parts) - - # 添加助手消息到历史 - if response_text: - assistant_msg = Message(role=Role.ASSISTANT, content=response_text) - else: - assistant_msg = Message(role=Role.ASSISTANT, content="No response generated. Try again later.") - - self.redis.save_message(user_id, user_msg) - self.redis.save_message(user_id, assistant_msg) - - return response_text - - async def get_conversation_summary(self, user_id: str) -> str: - """ - 分析用户的完整会话历史，并打包成一个简洁的需求总结。 - - 这个总结可以直接作为输入 Prompt 传递给 Stylist Agent。` - """ - history_messages = self.redis.get_history(user_id) - input_message = "\n".join([f"{msg.role.value}: {msg.content}" for msg in history_messages]) - - # 临时调用 LLM 或使用本地逻辑生成总结 - summary = await self.llm.generate_response(history=[Message(role=Role.USER, content=input_message)], system_prompt=SUMMARY_PROMPT) - - return summary - - async def recommend_outfit(self, request_summary: str, stylist_name: str, start_outfit: List[Dict[str, str]] = [], num_outfits: int = 1): - """ - 基于用户的对话历史和需求，推荐一套搭配。 - - Args: - request_summary: 用户的request - start_outfit: 可选的初始搭配列表，每个元素包含 'item_id' 和 'category'。 - """ - tasks = [] - for _ in range(num_outfits): - agent = AsyncStylistAgent(**self.stylist_agent_kwages) - task = agent.run_styling_process(request_summary, stylist_name, start_outfit) - tasks.append(task) - print(f"--- Starting {num_outfits} concurrent outfit generation tasks. ---") - - try: - results = await asyncio.gather(*tasks, return_exceptions=True) - - successful_outfits = [] - failed_outfits = [] - for result in results: - if isinstance(result, Exception): - # 任务执行中发生异常 - failed_outfits.append(f"Failed: {result}") - else: - # 任务成功，result 是 run_styling_process 返回的图片路径 - successful_outfits.append(result) - - return { - "successful_outfits": successful_outfits, - "failed_outfits": failed_outfits - } - - except Exception as e: - print(f"An unexpected error occurred during concurrent recommendation: {e}") - return {"error": str(e)} - - -if __name__ == "__main__": - async def run(): - # 阶段一：用户对话 - agent = ChatbotAgent() - user_id = "string" - # agent.redis.clear_history(user_id) # 清除历史，便于测试 - # await agent.process_query(user_id, "I want a chic outfit for a summer party.") - # print(await agent.process_query(user_id, "I prefer something floral and light.")) - - # 阶段二：读取聊天记录，生成推荐搭配 - request_summary = await agent.get_conversation_summary(user_id) - print(f"Conversation Summary:\n{request_summary}") - - recommendation_results = await agent.recommend_outfit(request_summary, stylist_name="crystal", start_outfit=[], num_outfits=1) - - print("\n--- Final Recommendation Results ---") - for i, path in enumerate(recommendation_results.get("successful_outfits", [])): - print(f"✅ Outfit {i + 1} saved to: {path}") - - for error in recommendation_results.get("failed_outfits", []): - print(f"❌ {error}") - asyncio.run(run()) diff --git a/app/core/data_structure.py b/app/core/data_structure.py deleted file mode 100644 index 241ec6a..0000000 --- a/app/core/data_structure.py +++ /dev/null @@ -1,16 +0,0 @@ -from typing import List, Dict, Any -from enum import Enum -from pydantic import BaseModel, Field -import datetime - -# 角色枚举，用于区分用户和系统的消息 -class Role(str, Enum): - USER = "user" - ASSISTANT = "assistant" - SYSTEM = "system" - -# 单条消息的数据模型 -class Message(BaseModel): - role: Role = Field(..., description="Role of message sender") - content: str = Field(..., description="Content of the message") - # timestamp: str = Field(default_factory=lambda: datetime.datetime.now().isoformat()) # 记录时间戳 \ No newline at end of file diff --git a/app/core/llm_interface.py b/app/core/llm_interface.py deleted file mode 100644 index 1bd6d92..0000000 --- a/app/core/llm_interface.py +++ /dev/null @@ -1,56 +0,0 @@ -from abc import ABC, abstractmethod -from typing import List - -from google import genai -from google.genai import types - -from app.core.data_structure import Message, Role - -class AsyncLLMInterface(ABC): - @abstractmethod - async def generate_response(self, history: List[Message], system_prompt: str) -> str: - """ - 根据对话历史和系统指令生成回复. - - Args: - history: 包含多条 Message 的列表。 - - Returns: - LLM 生成的回复字符串。 - """ - raise NotImplementedError("Subclasses must implement this method") - -class AsyncGeminiLLM(AsyncLLMInterface): - def __init__(self, model_name: str = "gemini-2.5-flash"): - self.model_name = model_name - try: - self.gemini_client = genai.Client( - vertexai=True, project='aida-461108', location='us-central1' - ) - except Exception as e: - raise type(e)(f"Failed to initialize Gemini Client. Check if GEMINI_API_KEY is set. Original error: {e}") - - async def generate_response(self, history: List[Message], system_prompt: str) -> str: - contents = [] - - for msg in history: - gemini_role = "user" if msg.role == Role.USER else "model" - content = types.Content( - role=gemini_role, - parts=[types.Part.from_text(text=msg.content)] - ) - contents.append(content) - - try: - response = await self.gemini_client.aio.models.generate_content( - model=self.model_name, - contents=contents, - config=types.GenerateContentConfig( - system_instruction=system_prompt, - # temperature=0.3, - ) - ) - return response.text - except Exception as e: - raise type(e)(f"Gemini API call failed: {e}") - diff --git a/app/core/llm_interface_stream.py b/app/core/llm_interface_stream.py deleted file mode 100644 index 6850fc8..0000000 --- a/app/core/llm_interface_stream.py +++ /dev/null @@ -1,61 +0,0 @@ -from abc import ABC, abstractmethod -from typing import List - -from google import genai -from google.genai import types - -from app.core.data_structure import Message, Role - - -class AsyncLLMInterface(ABC): - @abstractmethod - async def generate_response(self, history: List[Message], system_prompt: str) -> str: - """ - 根据对话历史和系统指令生成回复. - - Args: - history: 包含多条 Message 的列表。 - - Returns: - LLM 生成的回复字符串。 - """ - raise NotImplementedError("Subclasses must implement this method") - - -class AsyncGeminiLLM(AsyncLLMInterface): - def __init__(self, model_name: str = "gemini-2.5-flash"): - self.model_name = model_name - try: - self.gemini_client = genai.Client( - vertexai=True, project='aida-461108', location='us-central1' - ) - except Exception as e: - raise type(e)(f"Failed to initialize Gemini Client. Check if GEMINI_API_KEY is set. Original error: {e}") - - async def generate_response(self, history: List[Message], system_prompt: str): - contents = [] - - for msg in history: - gemini_role = "user" if msg.role == Role.USER else "model" - content = types.Content( - role=gemini_role, - parts=[types.Part.from_text(text=msg.content)] - ) - contents.append(content) - - return contents - # response_stream = await self.gemini_client.aio.models.generate_content_stream( - # model=self.model_name, - # contents=contents, - # config=types.GenerateContentConfig( - # system_instruction=system_prompt, - # # temperature=0.3, - # ) - # ) - # - # # 3. 异步迭代流，并 yield 每个块的文本 - # async for chunk in response_stream: - # # 确保 chunk 中有可用的文本 - # if chunk.text: - # print(chunk.text) - # yield chunk.text diff --git a/app/core/redis_manager.py b/app/core/redis_manager.py deleted file mode 100644 index 23cb2c5..0000000 --- a/app/core/redis_manager.py +++ /dev/null @@ -1,68 +0,0 @@ -import logging - -import redis -from typing import List, Optional -from app.core.data_structure import Message, Role - -logger = logging.getLogger(__name__) - - -# 这是一个同步 Redis 客户端，用于演示如何替换内存存储。 -# 在生产环境和异步 Web 框架中，应替换为 aioredis 等异步客户端。 -class RedisManager: - """同步管理器，用于在 Redis 中存储和检索对话历史。""" - - def __init__(self, host: str = 'localhost', port: int = 6379, db: int = 0, key_prefix: str = "chat:history:"): - self.r: Optional[redis.Redis] = None - self.key_prefix = key_prefix - try: - # 尝试连接 Redis - self.r = redis.Redis(host=host, port=port, db=db, decode_responses=True) - self.r.ping() - logger.info("Successfully connected to Redis at {}:{}".format(host, port)) - except Exception as e: - logger.error(f"⚠️ Failed to connect to Redis: {e}. Falling back to No-Op.") - self.r = None # 连接失败时设置为 None，避免后续操作报错 - - def _get_key(self, user_id: str) -> str: - """生成用户历史记录的 Redis 键名。""" - return f"{self.key_prefix}{user_id}" - - def _message_to_json(self, message: Message) -> str: - """将 Message 对象序列化为 JSON 字符串以便存储。""" - return message.model_dump_json() - - def _json_to_message(self, data: str) -> Message: - """将 JSON 字符串反序列化回 Message 对象。""" - try: - return Message.model_validate_json(data) - except Exception as e: - logger.error(f"Error deserializing message data: {data[:50]}... Error: {e}") - return Message(role=Role.ASSISTANT, content="[Deserialization Error]") - - def save_message(self, user_id: str, message: Message): - """将单条消息保存到用户历史记录列表的末尾。""" - if not self.r: - return - - message_json = self._message_to_json(message) - # RPUSH：将元素添加到列表的尾部 - self.r.rpush(self._get_key(user_id), message_json) - - def get_history(self, user_id: str) -> List[Message]: - """检索用户的完整会话历史记录。""" - if not self.r: - return [] - - # LRANGE：获取列表的所有元素 (0 到 -1) - raw_history = self.r.lrange(self._get_key(user_id), 0, -1) - - # 将 JSON 字符串列表转换为 Message 对象列表 - messages = [self._json_to_message(data) for data in raw_history] - return messages - - def clear_history(self, user_id: str): - """删除用户的完整历史记录。""" - if self.r: - self.r.delete(self._get_key(user_id)) - logger.info(f"History cleared for {user_id}") diff --git a/app/core/stylist_agent.py b/app/core/stylist_agent.py deleted file mode 100644 index c02d25d..0000000 --- a/app/core/stylist_agent.py +++ /dev/null @@ -1,302 +0,0 @@ -import json -import os -import random -import uuid -from typing import List, Dict, Any, Optional - -from google import genai - -from app.core.utils import merge_images_to_square - - -class AsyncStylistAgent: - CATEGORY_SET = {'Activewear', 'Watches', 'Shopping Totes', 'Underwear', 'Sunglasses', 'Dresses', 'Outerwear', 'Handbags', 'Backpacks', 'Belts', 'Hats', 'Skirts', 'Swimwear', 'Jewelry', 'Briefcases', 'Socks', 'Neckties', 'Pants', 'Suits', 'Shoes', 'Shirts & Tops', 'Scarves & Shawls'} - - def __init__(self, local_db, max_len: int, outfits_root: str, image_dir: str, stylist_guide_dir: str, gemini_model_name: str): - # self.outfit_items: List[Dict[str, str]] = [] - self.outfit_id = str(uuid.uuid4()) - self.gemini_client = genai.Client( - vertexai=True, project='aida-461108', location='us-central1' - ) - self.local_db = local_db - self.max_len = max_len - self.output_outfit_path = os.path.join(outfits_root, f"{self.outfit_id}.jpg") - self.output_json_path = os.path.join(outfits_root, f"{self.outfit_id}_items.json") - self.image_dir = image_dir - self.stylist_guide_dir = stylist_guide_dir - self.gemini_model_name = gemini_model_name - self.stop_reason = "" - - def _load_style_guide(self, path: str) -> str: - """加载 markdown 风格指南内容。""" - try: - with open(path, 'r', encoding='utf-8') as f: - return f.read() - except Exception as e: - raise FileNotFoundError(f"Failed to load style guide from {path}: {e}") - - def _build_system_prompt(self, request_summary: str = "") -> str: - """Constructs the complete System Prompt.""" - # Insert the style_guide content into the template - template = f""" - You are a professional fashion stylist Agent, specialized in creating complete outfits for the user. - - Your task is to **create a cohesive and complete outfit**, strictly adhering to **BOTH** the user's explicit **Request Summary** and the **Outfit Style Guide**. You must decide the next logical item to add to the outfit based on the currently selected items (if any). - - --- - - ## Request from the User: - - {request_summary} - - ## Core Guidance Document: Outfit Style Guide - - {self.style_guide} - - --- - - ## Your Workflow and Constraints - - 1. **Style Adherence**: You must strictly observe all rules in the Style Guide concerning **color palette, fit, layering principles, pattern restrictions, accessory stacking, and shoe/bag coordination**. - 2. **Step Planning**: The styling sequence must follow a **top-down, inside-out** approach: First major garments (tops/outerwear/bottoms/dresses), then shoes and bags, and finally accessories. - 3. **Structured Output**: Every response must recommend the **next single item**. You must strictly use the **JSON format** for your output, as follows: - - ```json - {{ - "action": "recommend_item", - "category": "YOUR_ITEM_CATEGORY", - "description": "YOUR_DETAILED_DESCRIPTION" - }} - ``` - - * `action`: Must always be `"recommend_item"` until the outfit is complete. - * `category`: Must be the category of the item you are recommending, strictly selected from the following list: {list(self.CATEGORY_SET)}. - * `description`: This must be an **extremely detailed and precise** description of the item. This description is used for **high-accuracy vector search** in the database and must include: - * **Color** (e.g., milk tea, pure white, dark gray) - * **Fit/Silhouette** (e.g., Oversize, loose, slim-fit) - * **Material/Detail** (e.g., 100% cotton, linen, gold clasp, thin stripe, checkered pattern) - * **Role in the Outfit** (e.g., serves as the innermost base layer for layering; acts as the crucial tie accent for the smart casual look) - * **[CRITICAL FOR JEWELRY] If recommending 'Jewelry' (especially Necklaces), the description must specify its distinction (length, thickness, pendant style) from all previously selected necklaces to ensure layered variety.** - - 4. **Termination Condition**: Only when you deem the entire outfit complete and **all mandatory elements stipulated in the Style Guide are met**, you must output the following JSON format to terminate the process: - - ```json - {{ - "action": "stop", - "reason": "OUTFIT_COMPLETE_AND_MEETS_ALL_MINI_GUIDELINES" - }} - ``` - Normally, five or six items are totally enough for an outfit. - - 5. **Context Dependency**: The user's next input (if not `Start`) will contain the **image and description of the selected item**. When recommending the next item, you must consider the coordination between the **already selected items** and the Style Guide. - - **Now, please start building an outfit and output the JSON for the first item.** - """ - return template.strip() - - def _clear_uploaded_files(self): - for f in self.gemini_client.files.list(): - self.gemini_client.files.delete(name=f.name) - - async def _call_gemini(self, user_input: str) -> str: - """ - 实际调用 Gemini API 的函数，接受文本和可选的图片路径列表。 - - Args: - user_input: 发送给模型的主文本内容。 - image_paths: 待发送图片的本地路径列表。 - - Returns: - 模型的响应文本（预期为 JSON 字符串）。 - """ - content_parts = [] - # self._clear_uploaded_files() - # 1. 添加图片内容 - # if self.outfit_items: - # merged_image_path = merge_images_to_square(self.outfit_items, max_len=self.max_len, output_path=self.output_outfit_path) - # try: - # myfile = await self.gemini_client.aio.files.upload(file=merged_image_path) - # content_parts.append(myfile) - # except Exception as e: - # print(f"Error loading image {merged_image_path}: {e}") - - # 2. 添加文本内容 - content_parts.append(user_input) - - # print(f"\n--- Calling Gemini with {len(self.outfit_items) if self.outfit_items else 0} images and query:\n{user_input}") - - try: - # 3. 实际 API 调用 - response = await self.gemini_client.aio.models.generate_content( - model=self.gemini_model_name, - contents=content_parts, - config={ - "system_instruction": self.system_prompt, - # 确保模型返回 JSON 格式 - "response_mime_type": "application/json", - "response_schema": { - "type": "object", - "properties": { - "action": {"type": "string", "enum": ["recommend_item", "stop"]}, - "category": {"type": "string"}, - "description": {"type": "string"}, - "reason": {"type": "string"} - }, - "required": ["action"] - } - } - ) - - # response.text 将包含一个 JSON 字符串 - return response.text - - except Exception as e: - print(f"Gemini API Call failed: {e}") - # 返回一个停止信号以防止循环继续 - return json.dumps({"action": "stop", "reason": f"API_ERROR: {str(e)}"}) - - def _parse_gemini_response(self, response_text: str) -> Optional[Dict[str, Any]]: - """安全解析 Gemini 的 JSON 响应。""" - try: - # 有时 Gemini 可能会在 JSON 外面添加文字,尝试清理 - response_text = response_text.strip().replace('```json', '').replace('```', '') - data = json.loads(response_text) - print(f"The agent response is: {data}") - return data - except json.JSONDecodeError as e: - print(f"Error parsing JSON from Gemini: {e}") - print(f"Raw response: {response_text}") - return None - - def _get_next_item(self, item_description: str, category: str) -> Optional[Dict[str, str]]: - """ - 1. 根据描述生成嵌入。 - 2. 查询本地数据库以找到最佳匹配项。 - 3. 模拟 Agent 审核匹配项(这里简化为总是通过)。 - """ - try: - # 1. 生成查询嵌入 - query_embedding = self.local_db.get_clip_embedding(item_description, is_image=False) - - # 2. 执行查询,并过滤类别 - results = self.local_db.query_local_db(query_embedding, category, n_results=1) - - if not results: - print(f"❌ 数据库中未找到符合 '{category}' 和描述的单品。") - return None - - # 3. 模拟 Agent 审核(实际应用中,你需要将图片发回给 Agent进行审核) - best_meta = results['metadatas'][0][0] # 第一个 batch 的第一个 metadata - return { - "item_id": best_meta['item_id'], # 从 metadata 字典中安全获取 - "category": category, - "gpt_description": item_description, - 'description': best_meta['description'], - # 假设 'item_path' 存储在 metadata 中，或从 'item_id' 推导 - # 这里假设 item_id 就是文件名的一部分 - "image_path": os.path.join(self.image_dir, f"{best_meta['item_id']}.jpg") - } - - except Exception as e: - print(f"An error occurred during item retrieval: {e}") - return None - - def _build_user_input(self) -> str: - """构建发送给 Gemini 的用户输入,包含已选单品信息。""" - if not self.outfit_items: - return "Start" - - # 将已选单品的信息作为上下文发回给 Agent - context = "Selected fashion items:\n" - for ii, item in enumerate(self.outfit_items): - context += f"{ii + 1}. Category: {item['category']}. Description: {item['description']}\n" - context += "\nPlease recommend the next single item based on the selected items, user's request, and style guide." - return context - - async def run_styling_process(self, request_summary, stylist_name, start_outfit=[]): - self.outfit_items = start_outfit if start_outfit else [] - """主流程控制循环。""" - print(f"--- Starting Agent (Outfit ID: {self.outfit_id}) ---") - - self.style_guide = self._load_style_guide(os.path.join(self.stylist_guide_dir, f"{stylist_name}_en.md")) - self.system_prompt = self._build_system_prompt(request_summary) - - while True: - # 1. 准备用户输入(上下文) - user_input = self._build_user_input() - - # 2. 调用 Gemini Agent - gemini_response_text = await self._call_gemini(user_input) - gemini_data = self._parse_gemini_response(gemini_response_text) - - if not gemini_data: - print("🚨 Agent 返回无效响应,终止流程。") - self.stop_reason = "Agent failed to return response" - break - - # 3. 检查终止条件 - if gemini_data.get('action') == 'stop': - print(f"🛑 搭配完成,终止原因: {gemini_data.get('reason')}") - self.stop_reason = "Finish reason: " + gemini_data.get('reason', 'No reason provided') - break - - # 4. 处理推荐单品 - if gemini_data.get('action') == 'recommend_item': - category = gemini_data.get('category') - description = gemini_data.get('description') - - # 4a. 检查类别是否有效 (重要步骤) - if category not in self.CATEGORY_SET: - print(f"❌ Agent 推荐了无效类别: {category}。要求 Agent 重新输出。") - # 在实际应用中,这里需要将错误信息发回给 Agent,要求它更正 - # 这里简化为跳过本次循环 - continue - - # 4b. 在本地 DB 中查询单品 - new_item = self._get_next_item(description, category) - - if new_item: - # 4c. (实际步骤) 将选中的单品图片和描述发回给 Agent 进行最终审核 - # 这里的代码框架省略了图片回传和二次审核的步骤,直接视为通过 - # 实际你需要： new_user_input = f"Check this item: {new_item['description']}, path: {new_item['image_path']}" - # call_gemini_agent(...) -> 如果返回"pass",则添加到outfit_items - - if new_item['item_id'] in [x['item_id'] for x in self.outfit_items]: - print("This item exists. Stop here.") - self.stop_reason = "Finish reason: Duplicate item selected." - break - - if new_item['item_id'] == "ELG383": - if random.random() < 0.70: - self.stop_reason = "Finish reason: ELG383 is seleced repeatly." - break - - self.outfit_items.append(new_item) - print(f"➕ 成功添加单品: {new_item['category']} ({new_item['item_id']}). 当前搭配数量: {len(self.outfit_items)}") - - else: - print("⚠️ 未找到匹配单品,无法继续搭配。终止。") - self.stop_reason = "Finish reason: No matching item found in local database." - break - - if len(self.outfit_items) >= self.max_len: # 设置一个最大循环限制,防止无限循环 - print("🚨 达到最大搭配数量限制,强制终止。") - self.stop_reason = "Finish reason: Reached max outfit length." - break - - # 5. 流程结束后保存结果 - self._save_outfit_results() - return self.output_outfit_path - - def _save_outfit_results(self): - """保存最终的 JSON 列表和图片到指定文件夹。""" - if not self.outfit_items: - raise ValueError("No outfit items to save.") - - # 1. 保存 JSON 文件 - results_list = [{'item_id': item['item_id'], 'category': item['category'], 'description': item['description'], 'gpt_description': item['gpt_description']} for item in self.outfit_items] - results_list.append({'stop_reason': self.stop_reason}) - with open(self.output_json_path, 'w', encoding='utf-8') as f: - json.dump(results_list, f, ensure_ascii=False, indent=4) - - merge_images_to_square(self.outfit_items, max_len=self.max_len, output_path=self.output_outfit_path, add_text=False) \ No newline at end of file diff --git a/app/core/stylist_agent_server.py b/app/core/stylist_agent_server.py deleted file mode 100644 index baaf1c2..0000000 --- a/app/core/stylist_agent_server.py +++ /dev/null @@ -1,435 +0,0 @@ -import asyncio -import io -import json -import logging -import os -import random -import uuid -from typing import List, Dict, Any, Optional - -from google import genai -from google.cloud import storage -from google.oauth2 import service_account - -from app.core.utils_litserve import merge_images_to_square -from app.server.utils.minio_client import minio_client, oss_upload_image -from app.server.utils.request_post import post_request - -logger = logging.getLogger(__name__) - - -class AsyncStylistAgent: - CATEGORY_SET = {'Activewear', 'Watches', 'Shopping Totes', 'Underwear', 'Sunglasses', 'Dresses', 'Outerwear', 'Handbags', 'Backpacks', 'Belts', 'Hats', 'Skirts', 'Swimwear', 'Jewelry', 'Briefcases', 'Socks', 'Neckties', 'Pants', 'Suits', 'Shoes', 'Shirts & Tops', 'Scarves & Shawls'} - - def __init__(self, local_db, max_len: int, gemini_model_name: str, outfit_id=str): - # self.outfit_items: List[Dict[str, str]] = [] - self.outfit_id = outfit_id - self.gemini_client = genai.Client( - vertexai=True, project='aida-461108', location='us-central1' - ) - self.local_db = local_db - self.max_len = max_len - self.gemini_model_name = gemini_model_name - self.stop_reason = "" - - # 存储桶配置 - try: - # TODO 目前写死路径 生产环境切换路径 - self.credentials = service_account.Credentials.from_service_account_file(os.getenv("GOOGLE_APPLICATION_CREDENTIALS")) - except Exception as e: - # 这里的异常处理应根据实际情况调整 - raise RuntimeError(f"Failed to load credentials from file {os.getenv('GOOGLE_APPLICATION_CREDENTIALS')}: {e}") - - self.gcs_client = storage.Client( - project=self.credentials.project_id, - credentials=self.credentials - ) - self.gcs_bucket = "lc_stylist_agent_outfit_items" - self.minio_bucket = "lanecarford" - - def _load_style_guide(self, path: str) -> str: - """加载 markdown 风格指南内容。""" - parts = path.split('/', 1) - if len(parts) != 2: - raise ValueError("MinIO path must be in 'bucket_name/object_name' format.") - - bucket_name, object_name = parts - try: - # 1. 获取对象 - response = minio_client.get_object(bucket_name, object_name) - - # 2. 读取内容 - content_bytes = response.read() - - # 3. 关闭连接 - response.close() - response.release_conn() - - # 4. 解码并返回 - return content_bytes.decode('utf-8') - - except Exception as e: - raise Exception(f"Failed to load style guide from {path}: {e}") - - def _build_system_prompt(self, request_summary: str = "") -> str: - """Constructs the complete System Prompt.""" - # Insert the style_guide content into the template - template = f""" - You are a professional fashion stylist Agent, specialized in creating complete outfits for the user. - - Your task is to **create a cohesive and complete outfit**, strictly adhering to **BOTH** the user's explicit **Request Summary** and the **Outfit Style Guide**. You must decide the next logical item to add to the outfit based on the currently selected items (if any). - - --- - - ## Request from the User: - - {request_summary} - - ## Core Guidance Document: Outfit Style Guide - - {self.style_guide} - - --- - - ## Your Workflow and Constraints - - 1. **Style Adherence**: You must strictly observe all rules in the Style Guide concerning **color palette, fit, layering principles, pattern restrictions, accessory stacking, and shoe/bag coordination**. - 2. **Step Planning**: The styling sequence must follow a **top-down, inside-out** approach: First major garments (tops/outerwear/bottoms/dresses), then shoes and bags, and finally accessories. - 3. **Structured Output**: Every response must recommend the **next single item**. You must strictly use the **JSON format** for your output, as follows: - - ```json - {{ - "action": "recommend_item", - "category": "YOUR_ITEM_CATEGORY", - "description": "YOUR_DETAILED_DESCRIPTION" - }} - ``` - - * `action`: Must always be `"recommend_item"` until the outfit is complete. - * `category`: Must be the category of the item you are recommending, strictly selected from the following list: {list(self.CATEGORY_SET)}. - * `description`: This must be an **extremely detailed and precise** description of the item. This description is used for **high-accuracy vector search** in the database and must include: - * **Color** (e.g., milk tea, pure white, dark gray) - * **Fit/Silhouette** (e.g., Oversize, loose, slim-fit) - * **Material/Detail** (e.g., 100% cotton, linen, gold clasp, thin stripe, checkered pattern) - * **Role in the Outfit** (e.g., serves as the innermost base layer for layering; acts as the crucial tie accent for the smart casual look) - * **[CRITICAL FOR JEWELRY] If recommending 'Jewelry' (especially Necklaces), the description must specify its distinction (length, thickness, pendant style) from all previously selected necklaces to ensure layered variety.** - - 4. **Termination Condition**: Only when you deem the entire outfit complete and **all mandatory elements stipulated in the Style Guide are met**, you must output the following JSON format to terminate the process: - - ```json - {{ - "action": "stop", - "reason": "OUTFIT_COMPLETE_AND_MEETS_ALL_MINI_GUIDELINES" - }} - ``` - Normally, five or six items are totally enough for an outfit. - - 5. **Context Dependency**: The user's next input (if not `Start`) will contain the **image and description of the selected item**. When recommending the next item, you must consider the coordination between the **already selected items** and the Style Guide. - - **Now, please start building an outfit and output the JSON for the first item.** - """ - return template.strip() - - def _clear_uploaded_files(self): - for f in self.gemini_client.files.list(): - self.gemini_client.files.delete(name=f.name) - - async def _call_gemini(self, user_input: str, user_id: str): - """ - 实际调用 Gemini API 的函数，接受文本和可选的图片路径列表。 - - Args: - user_input: 发送给模型的主文本内容。 - image_paths: 待发送图片的本地路径列表。 - - Returns: - 模型的响应文本（预期为 JSON 字符串）。 - """ - minio_path = "" - content_parts = [] - # self._clear_uploaded_files() - # 1. 添加图片内容 - if self.outfit_items: - merged_image = merge_images_to_square(self.outfit_items, max_len=self.max_len, add_text=False) - image_bytes_io = io.BytesIO() - image_format = 'JPEG' - mime_type = 'image/jpeg' - - merged_image.save(image_bytes_io, format=image_format) - image_bytes = image_bytes_io.getvalue() - - file_name = uuid.uuid4() - blob_name = f"lc_stylist_agent_outfit_items/{user_id}/{file_name}.jpg" - gcs_path = self._upload_to_gcs(bucket_name=self.gcs_bucket, blob_name=blob_name, mime_type=mime_type, image_bytes=image_bytes) - responses = oss_upload_image(oss_client=minio_client, bucket=self.minio_bucket, object_name=blob_name, image_bytes=image_bytes) - minio_path = f"{responses.bucket_name}/{responses.object_name}" - content_parts.append(gcs_path) - - # 2. 添加文本内容 - content_parts.append(user_input) - - # print(f"\n--- Calling Gemini with {len(self.outfit_items) if self.outfit_items else 0} images and query:\n{user_input}") - - try: - # 3. 实际 API 调用 - response = await self.gemini_client.aio.models.generate_content( - model=self.gemini_model_name, - contents=content_parts, - config={ - "system_instruction": self.system_prompt, - # 确保模型返回 JSON 格式 - "response_mime_type": "application/json", - "response_schema": { - "type": "object", - "properties": { - "action": {"type": "string", "enum": ["recommend_item", "stop"]}, - "category": {"type": "string"}, - "description": {"type": "string"}, - "reason": {"type": "string"} - }, - "required": ["action"] - } - } - ) - - # response.text 将包含一个 JSON 字符串 - return response.text, minio_path - - except Exception as e: - print(f"Gemini API Call failed: {e}") - # 返回一个停止信号以防止循环继续 - return json.dumps({"action": "stop", "reason": f"API_ERROR: {str(e)}"}) - - def _parse_gemini_response(self, response_text: str) -> Optional[Dict[str, Any]]: - """安全解析 Gemini 的 JSON 响应。""" - try: - # 有时 Gemini 可能会在 JSON 外面添加文字,尝试清理 - response_text = response_text.strip().replace('```json', '').replace('```', '') - data = json.loads(response_text) - # print(f"The agent response is: {data}") - return data - except json.JSONDecodeError as e: - print(f"Error parsing JSON from Gemini: {e}") - print(f"Raw response: {response_text}") - return None - - def _get_next_item(self, item_description: str, category: str) -> Optional[Dict[str, str]]: - """ - 1. 根据描述生成嵌入。 - 2. 查询本地数据库以找到最佳匹配项。 - 3. 模拟 Agent 审核匹配项(这里简化为总是通过)。 - """ - try: - # 1. 生成查询嵌入 - query_embedding = self.local_db.get_clip_embedding(item_description, is_image=False) - - # 2. 执行查询,并过滤类别 - results = self.local_db.query_local_db(query_embedding, category, n_results=1) - - if not results: - print(f"❌ 数据库中未找到符合 '{category}' 和描述的单品。") - return None - - # 3. 模拟 Agent 审核(实际应用中,你需要将图片发回给 Agent进行审核) - best_meta = results['metadatas'][0][0] # 第一个 batch 的第一个 metadata - return { - "item_id": best_meta['item_id'], # 从 metadata 字典中安全获取 - "category": category, - "gpt_description": item_description, - 'description': best_meta['description'], - # 假设 'item_path' 存储在 metadata 中，或从 'item_id' 推导 - # 这里假设 item_id 就是文件名的一部分 - "image_path": os.path.join(f"{best_meta['item_id']}.jpg") - } - - except Exception as e: - print(f"An error occurred during item retrieval: {e}") - return None - - def _build_user_input(self) -> str: - """构建发送给 Gemini 的用户输入,包含已选单品信息。""" - if not self.outfit_items: - return "Start" - - # 将已选单品的信息作为上下文发回给 Agent - context = "Selected fashion items:\n" - for ii, item in enumerate(self.outfit_items): - context += f"{ii + 1}. Category: {item['category']}. Description: {item['description']}\n" - context += "\nPlease recommend the next single item based on the selected items, user's request, and style guide." - return context - - async def run_styling_process(self, request_summary, stylist_path, start_outfit=None, user_id="test", callback_url=""): - if start_outfit is None: - start_outfit = [] - self.outfit_items = start_outfit if start_outfit else [] - """主流程控制循环。""" - print(f"--- Starting Agent (Outfit ID: {self.outfit_id}) ---") - - self.style_guide = self._load_style_guide(stylist_path) - self.system_prompt = self._build_system_prompt(request_summary) - response_data = {"status": "", - "message": "", - "path": "", - "outfit_id": self.outfit_id, - "items": [] - } - logger.info(response_data) - item_id = "" - item_category = "" - while True: - # 1. 准备用户输入(上下文) - user_input = self._build_user_input() - - # 2. 调用 Gemini Agent - gemini_response_text, minio_path = await self._call_gemini(user_input, user_id) - gemini_data = self._parse_gemini_response(gemini_response_text) - response_data['path'] = minio_path - if item_id: - response_data['items'].append({"item_id": item_id, "category": item_category}) - if not gemini_data: - print("🚨 Agent 返回无效响应,终止流程。") - self.stop_reason = "Agent failed to return response" - response_data['status'] = "failed" - response_data['message'] = self.stop_reason - break - - # 3. 检查终止条件 - if gemini_data.get('action') == 'stop': - print(f"🛑 搭配完成,终止原因: {gemini_data.get('reason')}") - self.stop_reason = "Finish reason: " + gemini_data.get('reason', 'No reason provided') - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - - # 4. 处理推荐单品 - if gemini_data.get('action') == 'recommend_item': - category = gemini_data.get('category') - description = gemini_data.get('description') - - # 4a. 检查类别是否有效 (重要步骤) - if category not in self.CATEGORY_SET: - print(f"❌ Agent 推荐了无效类别: {category}。要求 Agent 重新输出。") - # 在实际应用中,这里需要将错误信息发回给 Agent,要求它更正 - # 这里简化为跳过本次循环 - response_data['status'] = "continue" - response_data['message'] = f"❌ Agent 推荐了无效类别: {category}。要求 Agent 重新输出。", - continue - - # 4b. 在本地 DB 中查询单品 - new_item = self._get_next_item(description, category) - item_id = new_item.get('item_id') - item_category = new_item.get('category') - - if new_item: - # 4c. (实际步骤) 将选中的单品图片和描述发回给 Agent 进行最终审核 - # 这里的代码框架省略了图片回传和二次审核的步骤,直接视为通过 - # 实际你需要： new_user_input = f"Check this item: {new_item['description']}, path: {new_item['image_path']}" - # call_gemini_agent(...) -> 如果返回"pass",则添加到outfit_items - - if new_item['item_id'] in [x['item_id'] for x in self.outfit_items]: - print("This item exists. Stop here.") - self.stop_reason = "Finish reason: Duplicate item selected." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - break - - if new_item['item_id'] == "ELG383": - if random.random() < 0.70: - self.stop_reason = "Finish reason: ELG383 is seleced repeatly." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - break - - self.outfit_items.append(new_item) - # print(f"➕ 成功添加单品: {new_item['category']} ({new_item['item_id']}). 当前搭配数量: {len(self.outfit_items)}") - response_data['status'] = "ok" - response_data['message'] = self.stop_reason - else: - print("⚠️ 未找到匹配单品,无法继续搭配。终止。") - self.stop_reason = "Finish reason: No matching item found in local database." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - break - - if len(self.outfit_items) >= self.max_len: # 设置一个最大循环限制,防止无限循环 - logger.info("🚨 达到最大搭配数量限制,强制终止。") - self.stop_reason = "Finish reason: Reached max outfit length." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - - logger.info(f"request data ：{response_data}") - headers = { - 'Accept': "*/*", - 'Accept-Encoding': "gzip, deflate, br", - 'User-Agent': "PostmanRuntime-ApipostRuntime/1.1.0", - 'Connection': "keep-alive", - 'Content-Type': "application/json" - } - url = f'{callback_url}/api/style/callback' - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(response.text) - return response_data - - # def _save_outfit_results(self, user_id): - # """保存最终的 JSON 列表和图片到指定文件夹。""" - # if not self.outfit_items: - # raise ValueError("No outfit items to save.") - # - # # 1. 保存 JSON 文件 - # results_list = [{'item_id': item['item_id'], 'category': item['category'], 'description': item['description'], 'gpt_description': item['gpt_description']} for item in self.outfit_items] - # results_list.append({'stop_reason': self.stop_reason}) - # - # return upload_json_to_minio_sync( - # minio_client=minio_client, - # bucket_name=f"lanecarford", - # object_name=f"lc_stylist_agent_outfit_items/{user_id}/{uuid.uuid4()}.json", - # data=results_list - # ) - - def _upload_to_gcs(self, bucket_name: str, blob_name: str, mime_type, image_bytes) -> str: - """同步方法：将文件上传到 GCS 并返回 GCS URI。""" - bucket = self.gcs_client.bucket(bucket_name) - blob = bucket.blob(blob_name) - blob.upload_from_string( - data=image_bytes, - content_type=mime_type - ) - - gcs_uri = f"gs://{bucket_name}/{blob_name}" - return gcs_uri - - async def recommend_outfit(self, request_summary: str, stylist_name: str, start_outfit: List[Dict[str, str]] = [], num_outfits: int = 1): - """ - 基于用户的对话历史和需求，推荐一套搭配。 - - Args: - request_summary: 用户的request - start_outfit: 可选的初始搭配列表，每个元素包含 'item_id' 和 'category'。 - """ - tasks = [] - for _ in range(num_outfits): - agent = AsyncStylistAgent(**self.stylist_agent_kwages) - task = agent.run_styling_process(request_summary, stylist_name, start_outfit) - tasks.append(task) - print(f"--- Starting {num_outfits} concurrent outfit generation tasks. ---") - - try: - results = await asyncio.gather(*tasks, return_exceptions=True) - - successful_outfits = [] - failed_outfits = [] - for result in results: - if isinstance(result, Exception): - # 任务执行中发生异常 - failed_outfits.append(f"Failed: {result}") - else: - # 任务成功，result 是 run_styling_process 返回的图片路径 - successful_outfits.append(result) - - return { - "successful_outfits": successful_outfits, - "failed_outfits": failed_outfits - } - - except Exception as e: - print(f"An unexpected error occurred during concurrent recommendation: {e}") - return {"error": str(e)} diff --git a/app/core/system_prompt.py b/app/core/system_prompt.py deleted file mode 100644 index 7156586..0000000 --- a/app/core/system_prompt.py +++ /dev/null @@ -1,56 +0,0 @@ -BASIC_PROMPT = """""" -WOMEN_BASIC_PROMPT = """You are a professional, friendly, and insightful AI women's styling assistant. - -Your primary mission is to engage in a multi-turn conversation with the user to fully understand their dressing intent. You must adopt a professional yet approachable tone. - -CONVERSATION GOALS: -1. **Occasion:** Determine the specific event (e.g., romantic dinner, summer wedding, business meeting). -2. **Style:** Pinpoint the desired aesthetic (e.g., classic elegance, edgy, minimalist, bohemian). -3. **Vibe/Details:** Gather any mood or specific constraints (e.g., needs to be comfortable, requires light colors, no bare shoulders). -4. **Item Preference:** Ask the user if they have any specific preferences for an item type or silhouette (e.g., preference for a dress, skirt, tailored pants, or a particular neckline/length). - -GUIDANCE FOR RESPONSE GENERATION: -- After the user's initial request (e.g., "I want a chic outfit for dinner."), immediately reply with a friendly, targeted follow-up question to elicit the most crucial missing information (usually a combination of **Occasion** and **Style**). -- Be concise. Ask only 1 to 2 essential questions per turn. -- You must gather sufficient, clear intent before proceeding to actual clothing recommendations. - -OUTPUT FORMAT INSTRUCTION: -- **DO NOT** use any Markdown formatting whatsoever (e.g., do not use asterisks (*), bold text (**), lists, or code blocks). -- **ONLY** output the plain text response spoken by the AI Assistant. - -Example Follow-up (mimicking a conversational flow): -User: I want a chic outfit for dinner. -Your Response: Hey there! A chic dinner outfit, I love that! To give you the perfect recommendations, tell me: is this a romantic date, business dinner, or celebration with friends? And what's your go-to style vibe: classic elegance or something with more edge?""" -MEN_BASIC_PROMPT = """You are a professional, friendly, and insightful AI men's styling assistant. - -Your primary mission is to engage in a multi-turn conversation with the user to fully understand their dressing intent. You must adopt a professional yet approachable tone. - -CONVERSATION GOALS: -1. **Occasion:** Determine the specific event (e.g., romantic dinner, summer wedding, business meeting). -2. **Style:** Pinpoint the desired aesthetic (e.g., classic elegance, edgy, minimalist, bohemian). -3. **Vibe/Details:** Gather any mood or specific constraints (e.g., needs to be comfortable, requires light colors, no bare shoulders). -4. **Item Preference:** Ask the user if they have any specific preferences for an item type or silhouette (e.g., preference for a dress, skirt, tailored pants, or a particular neckline/length). - -GUIDANCE FOR RESPONSE GENERATION: -- After the user's initial request (e.g., "I want a chic outfit for dinner."), immediately reply with a friendly, targeted follow-up question to elicit the most crucial missing information (usually a combination of **Occasion** and **Style**). -- Be concise. Ask only 1 to 2 essential questions per turn. -- You must gather sufficient, clear intent before proceeding to actual clothing recommendations. - -OUTPUT FORMAT INSTRUCTION: -- **DO NOT** use any Markdown formatting whatsoever (e.g., do not use asterisks (*), bold text (**), lists, or code blocks). -- **ONLY** output the plain text response spoken by the AI Assistant. - -Example Follow-up (mimicking a conversational flow): -User: I want a chic outfit for dinner. -Your Response: Hey there! A chic dinner outfit, I love that! To give you the perfect recommendations, tell me: is this a romantic date, business dinner, or celebration with friends? And what's your go-to style vibe: classic elegance or something with more edge?""" - -SUMMARY_PROMPT = """Analyze the following chat history. Your task is to extract all user intentions, scenarios, style preferences, and constraints expressed during the conversation, and distill them into a concise, structured JSON object. - -**YOUR OUTPUT MUST BE A JSON OBJECT ONLY, WITH NO SURROUNDING TEXT, MARKDOWN, OR EXPLANATION.** - -JSON FIELD REQUIREMENTS: -- **occasion (string):** The specific event and purpose (e.g., "Romantic date dinner", "Summer outdoor wedding", "Casual Friday at office"). -- **style (string):** The overall aesthetic description (e.g., "Classic elegance", "Modern minimalist", "Bohemian vibe", "Edgy and contemporary"). -- **color_preference (string or list):** User's preferred or excluded colors/tones (e.g., "Light colors only", "Avoid deep shades", "['Cream', 'Pale Blue']", "No preference"). -- **clothing_type (string):** User's preference for specific garment types, material, or silhouette (e.g., "Lightweight maxi dress", "Skirt with silk blouse", "Tailored wide-leg pants", "Floral print"). -- **vibe_or_details (string):** Any other details, mood requirements, or specific constraints (e.g., "Needs to be comfortable and breathable", "Accent on accessories", "Must cover shoulders").""" diff --git a/app/core/utils.py b/app/core/utils.py deleted file mode 100644 index a64b003..0000000 --- a/app/core/utils.py +++ /dev/null @@ -1,160 +0,0 @@ -from typing import List, Dict -import shutil - -from PIL import Image, ImageDraw, ImageFont - -# 9个 341x341 左右的单元格 (ALL_9_CELLS) -# 布局顺序: 从上到下，从左到右 (1 -> 9) -ALL_9_CELLS = [ - # Top Row (Y=0, H=341) - (0, 0, 341, 341), # 1. Top-Left (341x341) - (341, 0, 341, 341), # 2. Top-Middle (341x341) - (682, 0, 342, 341), # 3. Top-Right (342x341) - # Middle Row (Y=341, H=341) - (0, 341, 341, 341), # 4. Mid-Left (341x341) - (341, 341, 341, 341), # 5. Center (341x341) - (682, 341, 342, 341), # 6. Mid-Right (342x341) - # Bottom Row (Y=682, H=342) - (0, 682, 341, 342), # 7. Bottom-Left (341x342) - (341, 682, 341, 342), # 8. Bottom-Middle (341x342) - (682, 682, 342, 342) # 9. Bottom-Right (342x342) -] - - -def merge_images_to_square(outfit_items: List[Dict[str, str]], max_len=9, output_path="temp.jpg", add_text=True): - """ - Loads up to 4 images from the given paths, resizes them while maintaining - aspect ratio, and merges them onto a 1024x1024 white background JPG. - - The layout depends on the number of images: - 1: Center the single image on the 1024x1024 canvas. - 2: Place side-by-side, each scaled to fit a 512x1024 half. - 3: Place in top-left (512x512), top-right (512x512), and bottom-left (512x512). - 4: Place in all four 512x512 quadrants. - - Args: - outfit_items: A list of item metadata (max length 9). - - Returns: - The file path of the temporary merged JPG image. - """ - - # Define the final canvas size - CANVAS_SIZE = 1024 - - # 1. Create the final white canvas - # Using 'RGB' mode for JPG output - canvas = Image.new('RGB', (CANVAS_SIZE, CANVAS_SIZE), 'white') - draw = ImageDraw.Draw(canvas) - font = ImageFont.load_default() - - # 2. Define the quadrants/target areas (x, y, w, h) - # The positions are based on a 512x512 quadrant size - quadrants = { - 1: [(0, 0, CANVAS_SIZE, CANVAS_SIZE)], # Single full-size placement - 2: [(0, 0, 512, CANVAS_SIZE), (512, 0, 512, CANVAS_SIZE)], # Left, Right - 3: [(0, 0, 512, 512), (512, 0, 512, 512), (0, 512, 512, 512)], # Top-Left, Top-Right, Bottom-Left - 4: [(0, 0, 512, 512), (512, 0, 512, 512), (0, 512, 512, 512), (512, 512, 512, 512)], # All Four - 5: ALL_9_CELLS[:5], # 布局前5个单元格 (1-5) - 6: ALL_9_CELLS[:6], # 布局前6个单元格 (1-6) - 7: ALL_9_CELLS[:7], # 布局前7个单元格 (1-7) - 8: ALL_9_CELLS[:8], # 布局前8个单元格 (1-8) - 9: ALL_9_CELLS[:9] # 布局全部9个单元格 (1-9) - } - - # 3. Load and Filter Images - valid_images = [] - image_paths = [item['image_path'] for item in outfit_items] - for path in image_paths: - try: - # We use Image.open() and convert to 'RGB' to handle potential transparency (RGBA) - # and ensure compatibility with the final 'RGB' canvas and JPG output. - img = Image.open(path).convert('RGB') - valid_images.append(img) - except Exception as e: - print(f"Error loading image {path}. Skipping: {e}") - - num_images = len(valid_images) - - if num_images == 0: - raise ValueError("No valid images were loaded.") - - if num_images > max_len: - raise ValueError(f"Valid item number {num_images} exceed max limit {max_len}") - - # Get the correct list of target areas based on the number of valid images - target_areas = quadrants.get(num_images, []) - - # 4. Resize and Paste - for i, (img, item) in enumerate(zip(valid_images, outfit_items)): - item_id = item['item_id'] - category = item['category'] - if i >= len(target_areas): - # This should not happen if num_images <= 4 - break - - # Target area dimensions (x_start, y_start, width, height) - x_start, y_start, target_w, target_h = target_areas[i] - - # Calculate new size while maintaining aspect ratio - original_w, original_h = img.size - - # Calculate the ratio needed to fit within the target area - ratio_w = target_w / original_w - ratio_h = target_h / original_h - - # Use the *smaller* of the two ratios to ensure the image fits entirely - resize_ratio = min(ratio_w, ratio_h) - - # Calculate the new dimensions - new_w = int(original_w * resize_ratio) - new_h = int(original_h * resize_ratio) - - # Resize the image. Image.Resampling.LANCZOS provides high-quality scaling. - # Pillow documentation recommends ANTIALIAS or BICUBIC for downscaling, - # but LANCZOS is a good general high-quality filter. - # Note: In Pillow versions > 9.0.0, Image.LANCZOS is now Image.Resampling.LANCZOS - resized_img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) - - # Calculate the paste position to center the resized image within its target area - # Center X: (Target Width - New Width) / 2 + X Start - paste_x = (target_w - new_w) // 2 + x_start - # Center Y: (Target Height - New Height) / 2 + Y Start - # paste_y = (target_h - new_h) // 2 + y_start - - TEXT_RESERVE_HEIGHT = 30 - paste_y = (target_h - new_h - TEXT_RESERVE_HEIGHT) // 2 + y_start - paste_y = max(paste_y, y_start) - - # Paste the resized image onto the canvas - canvas.paste(resized_img, (paste_x, paste_y)) - - full_text = f"ID: {item_id}, Category: {category}" - try: - # 推荐使用：计算文本的实际尺寸 (width, height) - bbox = draw.textbbox((0, 0), full_text, font=font) - text_w = bbox[2] - bbox[0] - text_h = bbox[3] - bbox[1] - except AttributeError: - # 兼容旧版本 Pillow - text_w, text_h = draw.textsize(full_text, font=font) - - # 计算 X 轴起始位置：使其在目标区域 (target_w) 中居中 - text_x_center = x_start + target_w // 2 - text_x_start = text_x_center - text_w // 2 - - # 计算 Y 轴起始位置：将其放在目标区域的底部 - # (目标区域的起始Y + 目标区域的高度 - 文本行的高度) - text_y_start = y_start + target_h - text_h - 5 # 减去 5 像素作为边距 - - # 3. 绘制合并后的文本 - if add_text: - draw.text((text_x_start, text_y_start), - full_text, - fill='black', - font=font) - - # Save as a high-quality JPG (quality=90 is a good balance) - # canvas.save(output_path, 'JPEG', quality=90) - - return canvas diff --git a/app/core/vector_database.py b/app/core/vector_database.py deleted file mode 100644 index 4707ebf..0000000 --- a/app/core/vector_database.py +++ /dev/null @@ -1,62 +0,0 @@ -import os -from typing import List, Dict, Any, Optional - -import torch -import chromadb -from transformers import CLIPProcessor, CLIPModel -from PIL import Image - - -class VectorDatabase(): - def __init__(self, vector_db_dir: str, collection_name: str, embedding_model_name: str): - self.client = chromadb.PersistentClient(path=vector_db_dir) - - self.collection = self.client.get_or_create_collection(name=collection_name) - - self.device = "cuda" if torch.cuda.is_available() else "cpu" - - self.model = CLIPModel.from_pretrained(embedding_model_name).to(self.device) - self.processor = CLIPProcessor.from_pretrained(embedding_model_name) - - def get_clip_embedding(self, data: str | Image.Image, is_image: bool) -> List[float]: - """生成图像或文本的 CLIP 嵌入，并进行 L2 归一化。""" - - if is_image: - inputs = self.processor(images=data, return_tensors="pt").to(self.device) - with torch.no_grad(): - features = self.model.get_image_features(**inputs) - else: - # 强制截断，解决序列长度问题 - inputs = self.processor( - text=[data], - return_tensors="pt", - padding=True, - truncation=True - ).to(self.device) - with torch.no_grad(): - features = self.model.get_text_features(**inputs) - - # L2 归一化 - features = features / features.norm(p=2, dim=-1, keepdim=True) - - return features.cpu().numpy().flatten().tolist() - - def query_local_db(self, embedding: List[float], category: str, n_results: int = 3) -> List[Dict[str, Any]]: - """ - 基于嵌入向量在本地数据库中查询相似单品。 - 实际应执行 ChromaDB 查询,并根据 category 进行过滤(metadatas)。 - """ - # 实际应执行向量查询 - # 为了演示流程,返回一个模拟结果 - results = self.collection.query( - query_embeddings=[embedding], - n_results=n_results, - where={ - "$and": [ - {"category": category}, - {"modality": "image"}, - ] - }, - include=['documents', 'metadatas', 'distances'] - ) - return results diff --git a/app/main.py b/app/main.py index 64ecfa9..e5cb807 100644 --- a/app/main.py +++ b/app/main.py @@ -1,7 +1,7 @@ import logging.config import os import litserve as ls -from app.core.config import DEBUG, settings +from app.config import DEBUG, settings from app.server.ChatbotAgent.agent_server import LCAgent from app.server.ChatbotAgent.chatbot_server import LCChatBot from app.server.ReFace.server import ReFace diff --git a/app/server/ChatbotAgent/agent_server.py b/app/server/ChatbotAgent/agent_server.py index db0b73f..6ca7bfb 100644 --- a/app/server/ChatbotAgent/agent_server.py +++ b/app/server/ChatbotAgent/agent_server.py @@ -1,10 +1,13 @@ import asyncio import logging import uuid +from enum import Enum +from typing import List +from pydantic import Field import litserve as ls from pydantic import BaseModel -from app.core.config import settings +from app.config import settings from app.server.ChatbotAgent.core.data_structure import Message, Role from app.server.ChatbotAgent.core.llm_interface import AsyncGeminiLLM from app.server.ChatbotAgent.core.redis_manager import RedisManager @@ -15,11 +18,40 @@ from app.server.ChatbotAgent.core.vector_database import VectorDatabase logger = logging.getLogger(__name__) +class OccasionEnum(str, Enum): + CASUAL = "Casual" + FORMAL = "Formal" + ACTIVEWEAR = "Activewear" + RESORT = "Resort" + EVENING = "Evening" + OUTDOOR = "Outdoor" + BUSINESS_WORKWEAR = "Business / workwear" + COCKTAIL_SEMI_FORMAL = "Cocktail / Semi-Formal" + BLACK_TIE_WHITE_TIE = "Black Tie / White Tie" + BRIDAL_WEDDING = "Bridal / Wedding" + FESTIVAL_CONCERT = "Festival / Concert" + PARTY_CLUBBING = "Party / Clubbing" + TRAVEL_TRANSIT = "Travel / Transit" + ATHLEISURE = "Athleisure" + BEACH_SWIM = "Beach / Swim" + SKI_SNOW_MOUNTAIN = "Ski / Snow / Mountain" + GARDEN_PARTY_DAYTIME = "Garden Party / Daytime Event" + +class StylistResponse(BaseModel): + occasions: List[OccasionEnum] = Field( + description="A list of **applicable** occasions that are most strongly implied or explicitly requested by the user's conversation history. These occasions are used later in item retrieval for filtering and must strictly match the predefined OccasionEnum list." + ) + summary: str = Field( + description="A detailed summary of the user's styling requirements, preferences, constraints, and specific item requests." + ) + + class AgentRequestModel(BaseModel): user_id: str session_id: str num_outfits: int stylist_path: str + batch_source: str callback_url: str gender: str max_len: int = 9 @@ -41,7 +73,6 @@ class LCAgent(ls.LitAPI): ) self.stylist_agent_kwages = { 'local_db': self.vector_db, - 'max_len': 9, 'gemini_model_name': settings.LLM_MODEL_NAME } @@ -73,40 +104,68 @@ class LCAgent(ls.LitAPI): async def background_run(self, request: AgentRequestModel, outfit_ids): # 1. 根据用户ID查询对话历史，总结对话内容 - request_summary = await self.get_conversation_summary(request.session_id) + request_summary, occasions = await self.get_conversation_summary(request.session_id) logger.info(f"request_summary: {request_summary}") # 2.根据对话总结推荐搭配 - recommendation_results = await self.recommend_outfit(request_summary=request_summary, - stylist_name=request.stylist_path, - start_outfit=[], - num_outfits=request.num_outfits, - user_id=request.user_id, - gender=request.gender, - callback_url=request.callback_url, - max_len=request.max_len, - outfit_ids=outfit_ids) + recommendation_results = await self.recommend_outfit( + request_summary=request_summary, + occasions=occasions, + stylist_name=request.stylist_path, + batch_source=request.batch_source, + start_outfit=[], + num_outfits=request.num_outfits, + user_id=request.user_id, + gender=request.gender, + callback_url=request.callback_url, + max_len=request.max_len, + outfit_ids=outfit_ids + ) logger.info("--- Final Recommendation Results ---") for i, path in enumerate(recommendation_results.get("successful_outfits", [])): logger.info(f"✅ Outfit {i + 1} saved to: {path}") for failed in recommendation_results.get("failed_outfits", []): logger.error(f"❌ {failed}") - async def get_conversation_summary(self, session_id: str) -> str: + async def get_conversation_summary(self, session_id: str) -> dict: """ - 分析用户的完整会话历史，并打包成一个简洁的需求总结。 - - 这个总结可以直接作为输入 Prompt 传递给 Stylist Agent。` + 分析用户的完整会话历史，并返回结构化的需求数据。 + + Returns: + occasions: List[str], # 用于 Vector DB 筛选 + summary: str # 用于 recommendation 的输入 """ history_messages = self.redis.get_history(session_id) - input_message = "\n".join([f"{msg.role.value}: {msg.content}" for msg in history_messages]) - # 临时调用 LLM 或使用本地逻辑生成总结 - summary = await self.llm.generate_response(history=[Message(role=Role.USER, content=input_message)], - system_prompt=SUMMARY_PROMPT) - return summary + if not history_messages: + # 处理无历史记录的情况 + return {"occasions": [], "summary": "User has no history provided."} - async def recommend_outfit(self, request_summary: str, stylist_name: str, start_outfit=None, num_outfits: int = 1, - user_id: str = "test", gender: str = "male", callback_url: str = None, max_len: int = 9, outfit_ids=None): + input_message = "\n".join([f"{msg.role.value}: {msg.content}" for msg in history_messages]) + json_schema = StylistResponse.model_json_schema() + + raw_response = await self.llm.generate_response( + history=[Message(role=Role.USER, content=input_message)], + system_prompt=SUMMARY_PROMPT, + json_schema=json_schema + ) + + try: + # 验证并解析 JSON + parsed_result = StylistResponse.model_validate_json(raw_response) + + print(f"Occasions: {[occ.value for occ in parsed_result.occasions]}") + print(f"Summary: {parsed_result.summary}") # 这是一个 string + + except Exception as e: + logger.error(f"Schema validation failed: {e}") + + return str(parsed_result.summary), [occ.value for occ in parsed_result.occasions] + + async def recommend_outfit( + self, request_summary: str, occasions: List[str], batch_source: str, stylist_name: str, start_outfit=[], + num_outfits: int = 1, user_id: str = "test", gender: str = "male", + callback_url: str = None, max_len: int = 9, outfit_ids=None + ): """ 基于用户的对话历史和需求，推荐一套搭配。 @@ -116,8 +175,6 @@ class LCAgent(ls.LitAPI): """ if outfit_ids is None: outfit_ids = [] - if start_outfit is None: - start_outfit = [] tasks = [] task_map = {} @@ -128,7 +185,9 @@ class LCAgent(ls.LitAPI): agent = AsyncStylistAgent(**stylist_agent_kwages) task = agent.run_styling_process( request_summary=request_summary, - stylist_path=stylist_name, + occasions=occasions, + batch_source=batch_source, + stylist_name=stylist_name, start_outfit=start_outfit, user_id=user_id, callback_url=callback_url, @@ -167,7 +226,9 @@ class LCAgent(ls.LitAPI): agent = AsyncStylistAgent(**stylist_agent_kwages) new_task = agent.run_styling_process( request_summary=request_summary, - stylist_path=stylist_name, + occasions=occasions, + batch_source=batch_source, + stylist_name=stylist_name, start_outfit=start_outfit, user_id=user_id, callback_url=callback_url @@ -209,3 +270,48 @@ class LCAgent(ls.LitAPI): "failed_outfits": failed_outfits, "error": "" } + + +if __name__ == "__main__": + async def test(): + # 1. 准备测试实例 + agent_api = LCAgent() + agent_api.setup(device='cpu') + + # 2. 准备请求数据 + import json + stylist_agent_kwages = agent_api.stylist_agent_kwages.copy() + with open("./data/2025_q4/request_test.json", "r") as f: + request_data = json.load(f) + + tasks = [] + for test_content in request_data[:30]: + occasions = test_content['occasions'] + request_summary = test_content['request_summary'] + stylist_agent_kwages['max_len'] = 5 + for stylist_name in ["edi", "vera"]: + stylist_agent_kwages['outfit_id'] = test_content['test_case_id'] + "_" + "_".join(occasions) + f"_{stylist_name}" + agent = AsyncStylistAgent(**stylist_agent_kwages) + task = agent.run_styling_process( + request_summary=request_summary, + occasions=occasions, + batch_source="2025_q4", + stylist_name=stylist_name, + start_outfit=[], + user_id=test_content['test_case_id'], + callback_url="http://mock-callback.com/result", + gender="female", + ) + tasks.append(task) + + results = await asyncio.gather(*tasks, return_exceptions=True) + for result in results: + if isinstance(result, Exception): + print(f"❌ 任务失败: {type(result).__name__} - {str(result)}") + continue + + try: + # 使用 asyncio.run() 来执行顶层异步函数 + asyncio.run(test()) + except Exception as e: + logger.error(f"Test failed due to an unexpected error: {e}") \ No newline at end of file diff --git a/app/server/ChatbotAgent/chatbot_server.py b/app/server/ChatbotAgent/chatbot_server.py index 872006e..87766b1 100644 --- a/app/server/ChatbotAgent/chatbot_server.py +++ b/app/server/ChatbotAgent/chatbot_server.py @@ -4,13 +4,13 @@ import litserve as ls from typing import AsyncGenerator from google import genai from pydantic import BaseModel -from app.core.config import settings +from app.config import settings from google.genai import types from app.server.ChatbotAgent.core.data_structure import Message, Role from app.server.ChatbotAgent.core.llm_interface import AsyncGeminiLLM from app.server.ChatbotAgent.core.redis_manager import RedisManager -from app.server.ChatbotAgent.core.system_prompt import MEN_BASIC_PROMPT, WOMEN_BASIC_PROMPT +from app.server.ChatbotAgent.core.system_prompt import BASIC_PROMPT from app.server.ChatbotAgent.core.vector_database import VectorDatabase logger = logging.getLogger(__name__) @@ -25,26 +25,12 @@ class PredictRequest(BaseModel): class LCChatBot(ls.LitAPI): def setup(self, device): - # self.llm = AsyncGeminiLLM(model_name=settings.LLM_MODEL_NAME) self.redis = RedisManager( host=settings.REDIS_HOST, port=settings.REDIS_PORT, db=settings.REDIS_DB, key_prefix=settings.REDIS_HISTORY_KEY_PREFIX ) - # self.vector_db = VectorDatabase( - # vector_db_dir=settings.VECTOR_DB_DIR, - # collection_name=settings.COLLECTION_NAME, - # embedding_model_name=settings.EMBEDDING_MODEL_NAME - # ) - # self.stylist_agent_kwages = { - # 'local_db': self.vector_db, - # 'max_len': 5, - # 'outfits_root': settings.OUTFIT_OUTPUT_DIR, - # 'image_dir': settings.IMAGE_DIR, - # 'stylist_guide_dir': settings.STYLIST_GUIDE_DIR, - # 'gemini_model_name': settings.LLM_MODEL_NAME - # } self.gemini_client = genai.Client( vertexai=True, project='aida-461108', location='us-central1' ) @@ -62,9 +48,9 @@ class LCChatBot(ls.LitAPI): chat_history = self.redis.get_history(session_id) chat_history.append(user_msg) if request.gender == 'male': - BASIC_PROMPT = MEN_BASIC_PROMPT + prompt = BASIC_PROMPT.format(gender='men') else: - BASIC_PROMPT = WOMEN_BASIC_PROMPT + prompt = BASIC_PROMPT.format(gender='women') contents = [] @@ -80,7 +66,7 @@ class LCChatBot(ls.LitAPI): model='gemini-2.5-flash', contents=contents, config=types.GenerateContentConfig( - system_instruction=BASIC_PROMPT, + system_instruction=prompt, # temperature=0.3, ) ) @@ -108,3 +94,45 @@ class LCChatBot(ls.LitAPI): # The for-loop must have async keyword here since output is an AsyncGenerator async for out in output: yield {"output": out} + + +if __name__ == "__main__": + import asyncio + async def run_simple_test(): + """ + 一个简单的异步测试用例，用于测试 LCChatBot 的流式输出。 + """ + print("\n" + "=" * 50) + print("--- 🔬 开始 LCChatBot 简单流式测试 ---") + + # 1. 初始化 LitAPI 和其依赖 + chatbot_api = LCChatBot() + chatbot_api.setup(device="cpu") + print("✅ Setup complete. Mock services initialized.") + + # 2. 构造请求数据 + request_data = PredictRequest( + user_id="simple_user", + session_id="simple_session", + user_message="I want an outfit. I am going to a evening party with friends. Suggest something stylish yet comfortable.", + gender="female" + ) + chatbot_api.redis.clear_history(request_data.session_id) + + print(f"-> 正在发送查询: {request_data.user_message}") + + # 3. 调用 predict 方法并处理流 + response_generator = chatbot_api.predict(request_data) + + print("\n<- 接收流式响应:") + + # 4. 异步迭代生成器，实时打印输出 + async for chunk in response_generator: + print(chunk, end="", flush=True) + + print("\n" + "=" * 50) + # 启动异步事件循环 + try: + asyncio.run(run_simple_test()) + except Exception as e: + print(f"\n发生致命错误: {e}") \ No newline at end of file diff --git a/app/server/ChatbotAgent/core/llm_interface.py b/app/server/ChatbotAgent/core/llm_interface.py index ca071ce..1a41c67 100644 --- a/app/server/ChatbotAgent/core/llm_interface.py +++ b/app/server/ChatbotAgent/core/llm_interface.py @@ -32,7 +32,7 @@ class AsyncGeminiLLM(AsyncLLMInterface): except Exception as e: raise type(e)(f"Failed to initialize Gemini Client. Check if GEMINI_API_KEY is set. Original error: {e}") - async def generate_response(self, history: List[Message], system_prompt: str) -> str: + async def generate_response(self, history: List[Message], system_prompt: str, json_schema=None) -> str: contents = [] for msg in history: @@ -44,14 +44,27 @@ class AsyncGeminiLLM(AsyncLLMInterface): contents.append(content) try: - response = await self.gemini_client.aio.models.generate_content( - model=self.model_name, - contents=contents, - config=types.GenerateContentConfig( - system_instruction=system_prompt, - # temperature=0.3, + if json_schema: + response = await self.gemini_client.aio.models.generate_content( + model=self.model_name, + contents=contents, + config=types.GenerateContentConfig( + system_instruction=system_prompt, + response_mime_type="application/json", + response_schema=json_schema + ) ) - ) - return response.text + return response.text + else: + response = await self.gemini_client.aio.models.generate_content( + model=self.model_name, + contents=contents, + config=types.GenerateContentConfig( + system_instruction=system_prompt, + # temperature=0.3, + ) + ) + return response.text except Exception as e: raise type(e)(f"Gemini API call failed: {e}") + \ No newline at end of file diff --git a/app/server/ChatbotAgent/core/stylist_agent_server.py b/app/server/ChatbotAgent/core/stylist_agent_server.py index 88901ca..d4ab40f 100644 --- a/app/server/ChatbotAgent/core/stylist_agent_server.py +++ b/app/server/ChatbotAgent/core/stylist_agent_server.py @@ -6,32 +6,22 @@ import os import random import uuid from typing import List, Dict, Any, Optional +from copy import deepcopy from google import genai from google.cloud import storage from google.oauth2 import service_account -from app.core.utils_litserve import merge_images_to_square +from app.server.utils.img_operation import merge_images_to_square from app.server.utils.minio_client import minio_client, oss_upload_image from app.server.utils.request_post import post_request +from app.config import settings +from app.taxonomy import CLOTHING_CATEGORY, ACCESSORY_CATEGORY logger = logging.getLogger(__name__) class AsyncStylistAgent: - CATEGORY_SET = { - 'Activewear', 'Dresses', 'Outerwear', 'Pants', 'Shirts & Tops', 'Skirts', 'Suits', 'Shoes', - # 取消推荐配饰 - # 'Swimwear', 'Underwear', - # , 'Watches', 'Shopping Totes', 'Sunglasses', 'Handbags', 'Backpacks', 'Belts', 'Hats', 'Jewelry', 'Briefcases', 'Socks', 'Neckties', 'Scarves & Shawls' - } - CATEGORY_SET_ALL = { - 'Activewear', 'Dresses', 'Outerwear', 'Pants', 'Shirts & Tops', 'Skirts', 'Suits', 'Swimwear', 'Underwear', - 'Watches', 'Shopping Totes', 'Sunglasses', 'Handbags', 'Backpacks', 'Belts', 'Hats', 'Jewelry', - 'Briefcases', 'Neckties', 'Shoes', 'Scarves & Shawls', - # 'Socks', - } - def __init__(self, local_db, max_len: int, gemini_model_name: str, outfit_id=str): # self.outfit_items: List[Dict[str, str]] = [] self.outfit_id = outfit_id @@ -42,6 +32,56 @@ class AsyncStylistAgent: self.max_len = max_len self.gemini_model_name = gemini_model_name self.stop_reason = "" + self.headers = { + 'Accept': "*/*", + 'Accept-Encoding': "gzip, deflate, br", + 'User-Agent': "PostmanRuntime-ApipostRuntime/1.1.0", + 'Connection': "keep-alive", + 'Content-Type': "application/json" + } + self.main_clothing_schema = { + "type": "object", + "properties": { + "action": {"type": "string", "enum": ["recommend_item", "stop"]}, + "category": { + "type": "string", + "description": "The category of the single clothing item being recommended in this step (e.g., 'outerwear', 'bottoms'). Only present if action is 'recommend_item'.", + "enum": CLOTHING_CATEGORY + }, + "description": { + "type": "string", + "description": "an **extremely detailed and precise** description of the item. This description is used for **high-accuracy vector search** in the database. It should include Color, Fit/Silhouette, Material/Detail, Role in the Outfit." + }, + "reason": {"type": "string", "description": "The reason for the current action. Required if action is 'stop' (to summarize the final outfit)."} + }, + "required": ["action"] + } + self.accessory_schema = { + "type": "object", + "properties": { + "reason": { + "type": "string", + "description": "The justification for completing the recommendation and the summary of the final outfit." + }, + "recommended_accessories": { + "type": "array", + "description": "A list of accessories recommended to complete the outfit.", + "items": { + "type": "object", + "properties": { + "category": { + "type": "string", + "description": "The category of the accessory (e.g., jewelry, watches, bags).", + "enum": ACCESSORY_CATEGORY + }, + "description": {"type": "string", "description": "The detailed description for this accessory item."} + }, + "required": ["category", "description"] + } + } + }, + "required": ["recommended_accessories", "reason"] + } # 存储桶配置 try: @@ -57,67 +97,42 @@ class AsyncStylistAgent: self.gcs_bucket = "lc_stylist_agent_outfit_items" self.minio_bucket = "lanecarford" - def _load_style_guide(self, path: str): + def _load_style_guide(self, stylist_name: str): """加载 markdown 风格指南内容。""" - parts = path.split('/', 1) - if len(parts) != 2: - raise ValueError("MinIO path must be in 'bucket_name/object_name' format.") - - bucket_name, object_name = parts + guide_path = os.path.join(settings.STYLIST_GUIDE_DIR, f"{stylist_name}_en.md") + acc_guide_path = os.path.join(settings.STYLIST_GUIDE_DIR, f"{stylist_name}_acc.md") try: - # 获取对象 读取内容 - response = minio_client.get_object(bucket_name, object_name) - content_bytes = response.read() - - json_response = minio_client.get_object(bucket_name, object_name.replace('.md', '.json')) - json_data = json_response.data - - # 关闭连接 - response.close() - json_response.close() - response.release_conn() - json_response.release_conn() - - # 4. 解析 JSON 字符串 - json_string = json_data.decode('utf-8') - json_content = json.loads(json_string) - - return content_bytes.decode('utf-8'), json_content - + with open(guide_path, 'r', encoding='utf-8') as file: + stylist_guide = file.read() + with open(acc_guide_path, 'r', encoding='utf-8') as file: + accessories_guide = file.read() + return stylist_guide, accessories_guide except Exception as e: - raise Exception(f"Failed to load style guide from {path}: {e}") + raise Exception(f"Failed to load style guide from {guide_path}, {acc_guide_path}: {e}") - def _build_system_prompt(self, request_summary: str = "", gender: str = "male") -> str: + def _build_main_clothing_prompt(self, request_summary: str = "", gender: str = "male", stylist_guide: str = "") -> str: """Constructs the complete System Prompt.""" - clothing_gender = "women's clothing" - if gender == "male": - clothing_gender = "men's clothing" - elif gender == "female": - clothing_gender = "women's clothing" + clothing_gender = "men's clothing" if gender == "male" else "women's clothing" # Insert the style_guide content into the template template = template = f""" - You are a professional fashion stylist Agent, specialized in creating complete, tailored outfits exclusively for {clothing_gender}. - + You are a professional fashion stylist Agent, specialized in creating complete, tailored outfits for {clothing_gender}. Only main clothing including 'bags' is needed, excluding accessories like 'jewelry', 'hats', 'belts', etc. + Your task is to **create a cohesive and complete outfit**, strictly adhering to **BOTH** the user's explicit **Request Summary** and the **Outfit Style Guide**. You must decide the next logical item to add to the outfit based on the currently selected items (if any). - + --- - ## Request from the User: - {request_summary} - + ## Core Guidance Document: Outfit Style Guide - - {self.style_guide} - + {stylist_guide} --- ## Your Workflow and Constraints 1. **Style Adherence**: You must strictly observe all rules in the Style Guide concerning **color palette, fit, layering principles, pattern restrictions , shoe coordination**. - 2. **Category Uniqueness Mandate**: Every outfit must follow the **absolute no-repeat rule for clothing categories** — each category from the allowed list ({list(self.CATEGORY_SET)}) can appear **exactly once** in the entire outfit. This rule is non-negotiable, even if the user explicitly requests repeating a category. + 2. **Category Uniqueness Mandate**: Every outfit must follow the **absolute no-repeat rule for clothing categories** — each category from the allowed list can appear **exactly once** in the entire outfit. This rule is non-negotiable, even if the user explicitly requests repeating a category. Furthermore, the categories 'dresses' and 'pants' and 'skirts' are mutually exclusive; they NORMALLY cannot be included in the same outfit. 3. **Step Planning**: The styling sequence must follow a **top-down, inside-out** approach: First major garments (tops/outerwear/bottoms/dresses) then shoes. When selecting the next item, prioritize unused categories from the allowed list to avoid repetition. 4. **Structured Output**: Every response must recommend the **next single item** (from an unused category). You must strictly use the **JSON format** for your output, as follows: @@ -130,7 +145,7 @@ class AsyncStylistAgent: ``` * `action`: Must always be `"recommend_item"` until the outfit is complete. - * `category`: Must be an unused category from the following list: {list(self.CATEGORY_SET)} (strictly no repeats, per the Category Uniqueness Mandate). + * `category`: Must be an unused category from the following list: {CLOTHING_CATEGORY} (strictly no repeats, per the Category Uniqueness Mandate). * `description`: This must be an **extremely detailed and precise** description of the item. This description is used for **high-accuracy vector search** in the database and must include: * **Color** (e.g., milk tea, pure white, dark gray) * **Fit/Silhouette** (e.g., Oversize, loose, slim-fit) @@ -147,112 +162,125 @@ class AsyncStylistAgent: "reason": "OUTFIT_COMPLETE_AND_MEETS_ALL_MINI_GUIDELINES" }} ``` - Normally, five or six items are totally enough for an outfit. + Normally, {self.max_len} items are totally enough for an outfit. 6. **Context Dependency**: The user's next input (if not Start) will contain the **image and description of the selected item**. When recommending the next item: a) First verify the categories of all already selected items to ensure no duplicates; - b) Select an unused category from the allowed list ({list(self.CATEGORY_SET)}) as the priority; + b) Select an unused category from the allowed list as the priority; c) Ensure the recommended item coordinates with the already selected items and complies with all rules in the Style Guide. Now, please start building an outfit (with strictly unique categories for all items) and output the JSON for the first item. """ return template.strip() - - def _clear_uploaded_files(self): - for f in self.gemini_client.files.list(): - self.gemini_client.files.delete(name=f.name) - - async def _call_gemini(self, user_input: str, user_id: str): + + def _build_accessory_prompt(self, request_summary: str, gender: str, accessories_guide: str) -> str: """ - 实际调用 Gemini API 的函数，接受文本和可选的图片路径列表。 + 构建配饰推荐 (Accessories) 的 System Prompt。 + 特点：强调基于现有穿搭 (Context Aware)，批量推荐 (Batch Recommendation)，做最后的点缀。 + """ + clothing_gender = "men's clothing" if gender == "male" else "women's clothing" + + template = f""" + You are an expert Accessories Stylist for {clothing_gender}. + Your task is to select the perfect set of accessories to complete an existing outfit. + + --- + ## CONTEXT + [User Request]: {request_summary} + + [Accessories Style Guide]: + {accessories_guide} + + --- + ## STRICT RULES + 1. **Batch Recommendation**: Do NOT recommend items one by one. You must output the **COMPLETE LIST** of accessories (e.g., jewelry, bag, watch, hat) in a single response using the 'recommended_accessories' list. + 2. **Allowed Categories**: Select only from: {ACCESSORY_CATEGORY}. + 3. **Harmony & Constraints**: + - The accessories must complement the [Current Outfit Base]. + - Strictly follow the [Accessories Style Guide] regarding metals (gold/silver), numbers, and prohibited items. + - If the guide mandates a watch or specific jewelry layering, ensure they are included. + 4. **Quantity**: Typically recommend 2-4 distinct accessory items to complete the look. + + Generate the final accessories list now. + """ + return template.strip() + + async def _call_gemini(self, user_input: str, user_id: str, file_name: str, output_schema: Dict[str, Any], image_bytes: bytes = None, system_prompt: str = "") -> str: + """ + 实际调用 Gemini API 的函数，接受文本和用户的id。 + 会在这个函数中merge图片，然后上传到google cloud，供gemini参考。 Args: user_input: 发送给模型的主文本内容。 - image_paths: 待发送图片的本地路径列表。 + user_id: 用户id。 + file_name: 用于存储图片的文件名。 + image_bytes: 可选的图片字节数据。 Returns: 模型的响应文本（预期为 JSON 字符串）。 """ - minio_path = "" content_parts = [] - # self._clear_uploaded_files() # 1. 添加图片内容 - if self.outfit_items: - merged_image = merge_images_to_square(self.outfit_items, max_len=self.max_len + 1, add_text=False) - image_bytes_io = io.BytesIO() - image_format = 'JPEG' - mime_type = 'image/jpeg' - - merged_image.save(image_bytes_io, format=image_format) - image_bytes = image_bytes_io.getvalue() - - file_name = uuid.uuid4() + if image_bytes: blob_name = f"lc_stylist_agent_outfit_items/{user_id}/{file_name}.jpg" - gcs_path = self._upload_to_gcs(bucket_name=self.gcs_bucket, blob_name=blob_name, mime_type=mime_type, image_bytes=image_bytes) - responses = oss_upload_image(oss_client=minio_client, bucket=self.minio_bucket, object_name=blob_name, image_bytes=image_bytes) - minio_path = f"{responses.bucket_name}/{responses.object_name}" + gcs_path = self._upload_to_gcs(bucket_name=self.gcs_bucket, blob_name=blob_name, mime_type='image/jpeg', image_bytes=image_bytes) content_parts.append(gcs_path) # 2. 添加文本内容 content_parts.append(user_input) - # print(f"\n--- Calling Gemini with {len(self.outfit_items) if self.outfit_items else 0} images and query:\n{user_input}") - try: # 3. 实际 API 调用 response = await self.gemini_client.aio.models.generate_content( model=self.gemini_model_name, contents=content_parts, config={ - "system_instruction": self.system_prompt, + "system_instruction": system_prompt, # 确保模型返回 JSON 格式 "response_mime_type": "application/json", - "response_schema": { - "type": "object", - "properties": { - "action": {"type": "string", "enum": ["recommend_item", "stop"]}, - "category": {"type": "string"}, - "description": {"type": "string"}, - "reason": {"type": "string"} - }, - "required": ["action"] - } + "response_schema": output_schema } ) # response.text 将包含一个 JSON 字符串 - return response.text, minio_path + return response.text except Exception as e: print(f"Gemini API Call failed: {e}") # 返回一个停止信号以防止循环继续 return json.dumps({"action": "stop", "reason": f"API_ERROR: {str(e)}"}) - async def _merge_images(self, user_id: str): + async def _merge_images(self, file_name: str, user_id: str, stylist_name: str): """ - 实际调用 Gemini API 的函数，接受文本和可选的图片路径列表。 + 把所有的item图片组成一张图片并保存到jpg文件 Args: - user_input: 发送给模型的主文本内容。 - image_paths: 待发送图片的本地路径列表。 + user_id: 用户的id + stylist_name: 造型师的name Returns: - 模型的响应文本（预期为 JSON 字符串）。 - """ - minio_path = "" - if self.outfit_items: - merged_image = merge_images_to_square(self.outfit_items, max_len=9, add_text=False) - image_bytes_io = io.BytesIO() - image_format = 'JPEG' + (存储的路径, 内存图片数据) + """ + if not self.outfit_items: + return "", None - merged_image.save(image_bytes_io, format=image_format) - image_bytes = image_bytes_io.getvalue() - - file_name = uuid.uuid4() + merged_image = merge_images_to_square(self.outfit_items, max_len=9, add_text=False) + image_bytes_io = io.BytesIO() + image_format = 'JPEG' + merged_image.save(image_bytes_io, format=image_format) + image_bytes = image_bytes_io.getvalue() + if settings.LOCAL == 1: + local_dir = os.path.join(settings.OUTFIT_OUTPUT_DIR, stylist_name) + os.makedirs(local_dir, exist_ok=True) + local_file_path = os.path.join(local_dir, f"{file_name}.jpg") + + with open(local_file_path, 'wb') as f: + f.write(image_bytes) + return local_file_path, image_bytes + else: blob_name = f"lc_stylist_agent_outfit_items/{user_id}/{file_name}.jpg" responses = oss_upload_image(oss_client=minio_client, bucket=self.minio_bucket, object_name=blob_name, image_bytes=image_bytes) minio_path = f"{responses.bucket_name}/{responses.object_name}" - - return minio_path + return minio_path, image_bytes def _parse_gemini_response(self, response_text: str) -> Optional[Dict[str, Any]]: """安全解析 Gemini 的 JSON 响应。""" @@ -267,7 +295,7 @@ class AsyncStylistAgent: print(f"Raw response: {response_text}") return None - def _get_next_item(self, item_description: str, category: str) -> Optional[Dict[str, str]]: + def _get_next_item(self, item_description: str, category: str, occasions: List[str], batch_source: str = "2025_q4", gender: str = "female") -> Optional[Dict[str, str]]: """ 1. 根据描述生成嵌入。 2. 查询本地数据库以找到最佳匹配项。 @@ -278,92 +306,30 @@ class AsyncStylistAgent: query_embedding = self.local_db.get_clip_embedding(item_description, is_image=False) # 2. 执行查询,并过滤类别 - results = self.local_db.query_local_db(query_embedding, category, n_results=1) + results = self.local_db.get_matched_item(query_embedding, category, occasions=occasions, batch_source=batch_source, gender=gender, n_results=1) if not results: print(f"❌ 数据库中未找到符合 '{category}' 和描述的单品。") return None # 3. 模拟 Agent 审核(实际应用中,你需要将图片发回给 Agent进行审核) - best_meta = results['metadatas'][0][0] # 第一个 batch 的第一个 metadata + best_meta = results[0] # 第一个 batch 的第一个 metadata + item_id = best_meta['item_id'].replace("_img", "") return { - "item_id": best_meta['item_id'], # 从 metadata 字典中安全获取 - "category": category, + "item_id": item_id, # 从 metadata 字典中安全获取 + "category": best_meta['category'], "gpt_description": item_description, 'description': best_meta['description'], # 假设 'item_path' 存储在 metadata 中，或从 'item_id' 推导 # 这里假设 item_id 就是文件名的一部分 - "image_path": os.path.join(f"{best_meta['item_id']}.jpg") + "image_path": os.path.join(f"{item_id}.jpg") } except Exception as e: print(f"An error occurred during item retrieval: {e}") return None - async def _get_random_accessories(self, stylist, item_count): - stylist_item = [] - stylist_item_ids = [] - - # 初始过滤类别 - filter_items = [ - {"item_group_id": {"$ne": "Clothing"}}, - {"item_group_id": {"$ne": "Shoes"}}, - {"category": {"$ne": "Socks"}}, - {"modality": "image"} - ] - random_items = [] - - for i in stylist: - # 1. 根据stylist要求抽取item - query_embedding = self.local_db.get_clip_embedding(i['text'], is_image=False) - stylist_results = self.local_db.query_local_db(query_embedding, i['category'], n_results=10) - stylist_item += random.choices(stylist_results['metadatas'][0], k=i['count']) - stylist_item_ids += [item_id['item_id'] for item_id in stylist_item] - filter_items.append({"category": {"$ne": i["category"]}}) - - accessories_count = 9 - item_count - len(stylist_item) - - if accessories_count > 0: - if accessories_count > 4: - accessories_count = 4 - for i in range(accessories_count): - # 2. 在配饰池中过滤掉已经选中的item ，然后抽两个item - random_poll = self.local_db.load_filtered_ids(filter_items) - logger.info(f"random_poll 数量: {len(random_poll)}") - - item = self.local_db.random_get_accessories(random.choice(random_poll)) - # 如果随机选中了包类 则所有包类别都过滤掉 - if item['metadatas'][0]['category'] in ['Shopping Totes', 'Handbags', 'Backpacks', 'Briefcases']: - filter_items.append({"category": {"$ne": "Shopping Totes"}}) - filter_items.append({"category": {"$ne": "Handbags"}}) - filter_items.append({"category": {"$ne": "Backpacks"}}) - filter_items.append({"category": {"$ne": "Briefcases"}}) - else: - filter_items.append({"category": {"$ne": item['metadatas'][0]['category']}}) - - random_items.append(item['metadatas'][0]) - - all_items = stylist_item + random_items - - else: - all_items = stylist_item - - items_data = [] - - for best_meta in all_items: - items_data.append({ - "item_id": best_meta['item_id'], # 从 metadata 字典中安全获取 - "category": best_meta['category'], - "gpt_description": best_meta['description'], - 'description': best_meta['description'], - # 假设 'item_path' 存储在 metadata 中，或从 'item_id' 推导 - # 这里假设 item_id 就是文件名的一部分 - "image_path": os.path.join(f"{best_meta['item_id']}.jpg") - }) - - return items_data - - def _build_user_input(self) -> str: + def _build_user_input(self, recommend_acc=False) -> str: """构建发送给 Gemini 的用户输入,包含已选单品信息。""" if not self.outfit_items: return "Start" @@ -372,164 +338,145 @@ class AsyncStylistAgent: context = "Selected fashion items:\n" for ii, item in enumerate(self.outfit_items): context += f"{ii + 1}. Category: {item['category']}. Description: {item['description']}\n" - context += "\nPlease recommend the next single item based on the selected items, user's request, and style guide." + if not recommend_acc: + context += "\nPlease recommend the next single item based on the selected items, user's request, and style guide." + else: + context += "\nPlease recommend a complete list of accessories to complement the selected outfit based on the user's request and accessories style guide." return context + + def post_operation(self, response_data: Dict[str, Any], status: str, message: str, callback_url: str): + """处理完成后的回调操作。""" + if settings.LOCAL == 0: + response_data['items'] = deepcopy(self.outfit_items) + response_data['status'] = status + response_data['message'] = message + response = post_request(url=callback_url, data=json.dumps(response_data), headers=self.headers) + logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") - async def run_styling_process(self, request_summary, stylist_path, start_outfit=None, user_id="test", callback_url="", gender: str = "male"): - if start_outfit is None: - start_outfit = [] - self.outfit_items = start_outfit if start_outfit else [] + async def run_styling_process(self, request_summary, occasions, stylist_name, batch_source="2025_q4", start_outfit=[], user_id="test", callback_url="", gender: str = "male"): + self.outfit_items = start_outfit """主流程控制循环。""" print(f"--- Starting Agent (Outfit ID: {self.outfit_id}) ---") - self.style_guide, self.style_accessories_guide = self._load_style_guide(stylist_path) - self.system_prompt = self._build_system_prompt(request_summary, gender) - response_data = {"status": "", - "message": "", - "path": "", - "outfit_id": self.outfit_id, - "items": [] - } - logger.info(response_data) - item_id = "" - item_category = "" - headers = { - 'Accept': "*/*", - 'Accept-Encoding': "gzip, deflate, br", - 'User-Agent': "PostmanRuntime-ApipostRuntime/1.1.0", - 'Connection': "keep-alive", - 'Content-Type': "application/json" + stylist_guide, accessories_guide = self._load_style_guide(stylist_name) + system_prompt = self._build_main_clothing_prompt(request_summary, gender, stylist_guide) + + response_data = { + "status": "", + "message": "", + "path": "", + "outfit_id": self.outfit_id, + "items": [] } + logger.info(response_data) url = f'{callback_url}/api/style/callback' - while True: + file_name = self.outfit_id + + recommend_timestep = 0 + gemini_data = {'action': 'start'} + while recommend_timestep < self.max_len and gemini_data.get('action') != 'stop': + recommend_timestep += 1 # 1. 准备用户输入(上下文) user_input = self._build_user_input() - # 2. 调用 Gemini Agent - gemini_response_text, minio_path = await self._call_gemini(user_input, user_id) + # 2. 把图片组装起来供api调用 + response_data['path'], image_bytes = await self._merge_images(file_name, user_id, stylist_name) + + # 3. 调用 Gemini Agent + gemini_response_text = await self._call_gemini(user_input, user_id, file_name, self.main_clothing_schema, image_bytes, system_prompt) gemini_data = self._parse_gemini_response(gemini_response_text) - response_data['path'] = minio_path - if item_id: - response_data['items'].append({"item_id": item_id, "category": item_category}) if not gemini_data: - # if gemini_data: - print("🚨 Agent 返回无效响应,终止流程。") - self.stop_reason = "Agent failed to return response" - response_data['status'] = "failed" - response_data['message'] = self.stop_reason - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") + print("Agent 返回无效响应,终止流程。") + self.post_operation( + response_data, + status="failed", + message="Agent returned invalid response, terminating process.", + callback_url=url + ) break - # 3. 检查终止条件 - if gemini_data.get('action') == 'stop': - if is_duplicate_by_key(response_data['items'], {"item_id": item_id, "category": item_category}): - print("重复（按item_id判断），不插入") - else: - response_data['path'] = minio_path - response_data['items'].append({"item_id": item_id, "category": item_category}) - response_data['status'] = "ok" - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") - - # 根据stylist要求随机增加配饰 3-4个配饰 - new_item = await self._get_random_accessories(self.style_accessories_guide, len(self.outfit_items)) - for item in new_item: - self.outfit_items.append(item) - response_data['items'].append({"item_id": item.get('item_id'), "category": item.get('category')}) - - response_data['path'] = await self._merge_images(user_id) - - logger.info(f"🛑 搭配完成,终止原因: {gemini_data.get('reason')}") - self.stop_reason = "Finish reason: " + gemini_data.get('reason', 'No reason provided') - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") - break - - # 4. 处理推荐单品 + # 处理推荐单品 if gemini_data.get('action') == 'recommend_item': category = gemini_data.get('category') description = gemini_data.get('description') # 4a. 检查类别是否有效 (重要步骤) - if category not in self.CATEGORY_SET_ALL: - print(f"❌ Agent 推荐了无效类别: {category}。要求 Agent 重新输出。") - # 在实际应用中,这里需要将错误信息发回给 Agent,要求它更正 - # 这里简化为跳过本次循环 - response_data['status'] = "continue" - response_data['message'] = f"❌ Agent 推荐了无效类别: {category}。要求 Agent 重新输出。", - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") + if category not in CLOTHING_CATEGORY: + self.post_operation( + response_data, + status="continue", + message=f"Invalid category recommended by Agent: {category}. Requesting Agent to re-output.", + callback_url=url + ) continue # 4b. 在本地 DB 中查询单品 - new_item = self._get_next_item(description, category) - item_id = new_item.get('item_id') - item_category = new_item.get('category') - - if new_item: - # 4c. (实际步骤) 将选中的单品图片和描述发回给 Agent 进行最终审核 - # 这里的代码框架省略了图片回传和二次审核的步骤,直接视为通过 - # 实际你需要： new_user_input = f"Check this item: {new_item['description']}, path: {new_item['image_path']}" - # call_gemini_agent(...) -> 如果返回"pass",则添加到outfit_items - - if new_item['item_id'] in [x['item_id'] for x in self.outfit_items]: - print("This item exists. Stop here.") - self.stop_reason = "Finish reason: Duplicate item selected." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") - break - - if new_item['item_id'] == "ELG383": - if random.random() < 0.70: - self.stop_reason = "Finish reason: ELG383 is seleced repeatly." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") - break - - self.outfit_items.append(new_item) - # print(f"➕ 成功添加单品: {new_item['category']} ({new_item['item_id']}). 当前搭配数量: {len(self.outfit_items)}") - response_data['status'] = "ok" - response_data['message'] = self.stop_reason - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") + new_item = self._get_next_item(description, category, occasions, batch_source, gender) + if not new_item or new_item['item_id'] in [x['item_id'] for x in self.outfit_items]: + self.post_operation( + response_data, + status="continue", + message=f"No matching item is found or item duplicated. Ask Gemini to re-output.", + callback_url=url + ) + continue else: - print("⚠️ 未找到匹配单品,无法继续搭配。终止。") - self.stop_reason = "Finish reason: No matching item found in local database." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") - break + self.outfit_items.append(new_item) + self.post_operation( + response_data, + status="ok", + message=f"Add new item {new_item['item_id']} in category {new_item['category']} successfully.", + callback_url=url + ) + print(f"Step {recommend_timestep}: {gemini_data}, found item: {new_item}") - if len(self.outfit_items) >= self.max_len: # 设置一个最大循环限制,防止无限循环 - gemini_response_text, response_data['path'] = await self._call_gemini(user_input, user_id) - response_data['items'].append({"item_id": self.outfit_items[-1]['item_id'], "category": self.outfit_items[-1]['category']}) - response_data['status'] = "ok" - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") + # When action is stop or timestep limit reached + logger.info(f"Main clothing stylist process finished: {gemini_data.get('reason')}") + # 根据stylist要求随机增加配饰 3-4个配饰 + response_data['path'], image_bytes = await self._merge_images(file_name, user_id, stylist_name) + accessory_system_prompt = self._build_accessory_prompt(request_summary, gender, accessories_guide) + user_input = self._build_user_input(recommend_acc=True) + gemini_response_text = await self._call_gemini(user_input, user_id, file_name, self.accessory_schema, image_bytes, accessory_system_prompt) + gemini_data = self._parse_gemini_response(gemini_response_text) - # 根据stylist要求随机增加配饰 3-4个配饰 - new_item = await self._get_random_accessories(self.style_accessories_guide, len(self.outfit_items)) - for item in new_item: - self.outfit_items.append(item) - response_data['items'].append({"item_id": item.get('item_id'), "category": item.get('category')}) - response_data['path'] = await self._merge_images(user_id) + recommended_accessories = gemini_data.get('recommended_accessories', []) + reason = gemini_data.get('reason', '') + if not recommended_accessories or not isinstance(recommended_accessories, List): + print("No accessory data from Gemini, terminating process.") + self.post_operation( + response_data, + status="failed", + message="Agent returned invalid response, terminating process.", + callback_url=url + ) + else: + for idx, rec_accessory in enumerate(recommended_accessories): + category = rec_accessory.get('category') + description = rec_accessory.get('description') - logger.info("🚨 达到最大搭配数量限制,强制终止。") - self.stop_reason = "Finish reason: Reached max outfit length." - response_data['status'] = "stop" - response_data['message'] = self.stop_reason - response = post_request(url=url, data=json.dumps(response_data), headers=headers) - logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}") - break + # 4a. 检查类别是否有效 (重要步骤) + if category not in ACCESSORY_CATEGORY: + continue + + # 4b. 在本地 DB 中查询单品 + new_item = self._get_next_item(description, category, occasions, batch_source, gender) + if not new_item or new_item['item_id'] in [x['item_id'] for x in self.outfit_items]: + continue + else: + self.outfit_items.append(new_item) + print(f"Accessory {idx + 1}: {rec_accessory}, found item: {new_item}") + + response_data['path'] = await self._merge_images(file_name, user_id, stylist_name) + self.post_operation( + response_data, + status="stop", + message=reason, + callback_url=url + ) + with open(os.path.join(settings.OUTFIT_OUTPUT_DIR, stylist_name, f'{file_name}.json'), 'w') as f: + json.dump(self.outfit_items, f, indent=2) + return response_data def _upload_to_gcs(self, bucket_name: str, blob_name: str, mime_type, image_bytes) -> str: @@ -543,11 +490,3 @@ class AsyncStylistAgent: gcs_uri = f"gs://{bucket_name}/{blob_name}" return gcs_uri - - -def is_duplicate_by_key(data, target_item): - """基于item_id快速判断重复""" - # 提取所有item_id到集合 - existing_ids = {item['item_id'] for item in data} - # 判断目标item_id是否在集合中 - return target_item['item_id'] in existing_ids diff --git a/app/server/ChatbotAgent/core/system_prompt.py b/app/server/ChatbotAgent/core/system_prompt.py index 39ab0bd..7f742d9 100644 --- a/app/server/ChatbotAgent/core/system_prompt.py +++ b/app/server/ChatbotAgent/core/system_prompt.py @@ -1,27 +1,4 @@ -BASIC_PROMPT = """""" -WOMEN_BASIC_PROMPT = """You are a professional, friendly, and insightful AI women's styling assistant. - -Your primary mission is to engage in a multi-turn conversation with the user to fully understand their dressing intent. You must adopt a professional yet approachable tone. - -CONVERSATION GOALS: -1. **Occasion:** Determine the specific event (e.g., romantic dinner, summer wedding, business meeting). -2. **Style:** Pinpoint the desired aesthetic (e.g., classic elegance, edgy, minimalist, bohemian). -3. **Vibe/Details:** Gather any mood or specific constraints (e.g., needs to be comfortable, requires light colors, no bare shoulders). -4. **Item Preference:** Ask the user if they have any specific preferences for an item type or silhouette (e.g., preference for a dress, skirt, tailored pants, or a particular neckline/length). - -GUIDANCE FOR RESPONSE GENERATION: -- After the user's initial request (e.g., "I want a chic outfit for dinner."), immediately reply with a friendly, targeted follow-up question to elicit the most crucial missing information (usually a combination of **Occasion** and **Style**). -- Be concise. Ask only 1 to 2 essential questions per turn. -- You must gather sufficient, clear intent before proceeding to actual clothing recommendations. - -OUTPUT FORMAT INSTRUCTION: -- **DO NOT** use any Markdown formatting whatsoever (e.g., do not use asterisks (*), bold text (**), lists, or code blocks). -- **ONLY** output the plain text response spoken by the AI Assistant. - -Example Follow-up (mimicking a conversational flow): -User: I want a chic outfit for dinner. -Your Response: Hey there! A chic dinner outfit, I love that! To give you the perfect recommendations, tell me: is this a romantic date, business dinner, or celebration with friends? And what's your go-to style vibe: classic elegance or something with more edge?""" -MEN_BASIC_PROMPT = """You are a professional, friendly, and insightful AI men's styling assistant. +BASIC_PROMPT = """You are a professional, friendly, and insightful AI {gender}'s styling assistant. Your primary mission is to engage in a multi-turn conversation with the user to fully understand their dressing intent. You must adopt a professional yet approachable tone. @@ -44,13 +21,12 @@ Example Follow-up (mimicking a conversational flow): User: I want a chic outfit for dinner. Your Response: Hey there! A chic dinner outfit, I love that! To give you the perfect recommendations, tell me: is this a romantic date, business dinner, or celebration with friends? And what's your go-to style vibe: classic elegance or something with more edge?""" -SUMMARY_PROMPT = """Analyze the following chat history. Your task is to extract all user intentions, scenarios, style preferences, and constraints expressed during the conversation, and distill them into a concise, structured JSON object. +SUMMARY_PROMPT = """ +You are an expert fashion request analyzer. Analyze the conversation history provided by the user. +Your task is to: -**YOUR OUTPUT MUST BE A JSON OBJECT ONLY, WITH NO SURROUNDING TEXT, MARKDOWN, OR EXPLANATION.** +1. Identify the most appropriate occasions from the allowed list based on the user's intent. +2. Write a detailed summary string that captures the user's style preferences, specific item requests, disliked items, body concerns, and color preferences. This summary will be used by a stylist to recommend outfits. -JSON FIELD REQUIREMENTS: -- **occasion (string):** The specific event and purpose (e.g., "Romantic date dinner", "Summer outdoor wedding", "Casual Friday at office"). -- **style (string):** The overall aesthetic description (e.g., "Classic elegance", "Modern minimalist", "Bohemian vibe", "Edgy and contemporary"). -- **color_preference (string or list):** User's preferred or excluded colors/tones (e.g., "Light colors only", "Avoid deep shades", "['Cream', 'Pale Blue']", "No preference"). -- **clothing_type (string):** User's preference for specific garment types, material, or silhouette (e.g., "Lightweight maxi dress", "Skirt with silk blouse", "Tailored wide-leg pants", "Floral print"). -- **vibe_or_details (string):** Any other details, mood requirements, or specific constraints (e.g., "Needs to be comfortable and breathable", "Must cover shoulders").""" \ No newline at end of file +Extract this information accurately from the chat history. +""" \ No newline at end of file diff --git a/app/server/ChatbotAgent/core/utils_litserve.py.py b/app/server/ChatbotAgent/core/utils_litserve.py.py deleted file mode 100644 index fcaa4a4..0000000 --- a/app/server/ChatbotAgent/core/utils_litserve.py.py +++ /dev/null @@ -1,163 +0,0 @@ -import logging -from typing import List, Dict -from PIL import Image, ImageDraw, ImageFont -from app.server.utils.minio_client import oss_get_image, minio_client -from app.server.utils.minio_config import MINIO_LC_DATA_PATH - -logger = logging.getLogger(__name__) -# 9个 341x341 左右的单元格 (ALL_9_CELLS) -# 布局顺序: 从上到下，从左到右 (1 -> 9) -ALL_9_CELLS = [ - # Top Row (Y=0, H=341) - (0, 0, 341, 341), # 1. Top-Left (341x341) - (341, 0, 341, 341), # 2. Top-Middle (341x341) - (682, 0, 342, 341), # 3. Top-Right (342x341) - # Middle Row (Y=341, H=341) - (0, 341, 341, 341), # 4. Mid-Left (341x341) - (341, 341, 341, 341), # 5. Center (341x341) - (682, 341, 342, 341), # 6. Mid-Right (342x341) - # Bottom Row (Y=682, H=342) - (0, 682, 341, 342), # 7. Bottom-Left (341x342) - (341, 682, 341, 342), # 8. Bottom-Middle (341x342) - (682, 682, 342, 342) # 9. Bottom-Right (342x342) -] - - -def merge_images_to_square(outfit_items: List[Dict[str, str]], max_len=9, add_text=True): - """ - Loads up to 4 images from the given paths, resizes them while maintaining - aspect ratio, and merges them onto a 1024x1024 white background JPG. - - The layout depends on the number of images: - 1: Center the single image on the 1024x1024 canvas. - 2: Place side-by-side, each scaled to fit a 512x1024 half. - 3: Place in top-left (512x512), top-right (512x512), and bottom-left (512x512). - 4: Place in all four 512x512 quadrants. - - Args: - outfit_items: A list of item metadata (max length 9). - - Returns: - The file path of the temporary merged JPG image. - """ - - # Define the final canvas size - CANVAS_SIZE = 1024 - - # 1. Create the final white canvas - # Using 'RGB' mode for JPG output - canvas = Image.new('RGB', (CANVAS_SIZE, CANVAS_SIZE), 'white') - draw = ImageDraw.Draw(canvas) - font = ImageFont.load_default() - - # 2. Define the quadrants/target areas (x, y, w, h) - # The positions are based on a 512x512 quadrant size - quadrants = { - 1: [(0, 0, CANVAS_SIZE, CANVAS_SIZE)], # Single full-size placement - 2: [(0, 0, 512, CANVAS_SIZE), (512, 0, 512, CANVAS_SIZE)], # Left, Right - 3: [(0, 0, 512, 512), (512, 0, 512, 512), (0, 512, 512, 512)], # Top-Left, Top-Right, Bottom-Left - 4: [(0, 0, 512, 512), (512, 0, 512, 512), (0, 512, 512, 512), (512, 512, 512, 512)], # All Four - 5: ALL_9_CELLS[:5], # 布局前5个单元格 (1-5) - 6: ALL_9_CELLS[:6], # 布局前6个单元格 (1-6) - 7: ALL_9_CELLS[:7], # 布局前7个单元格 (1-7) - 8: ALL_9_CELLS[:8], # 布局前8个单元格 (1-8) - 9: ALL_9_CELLS[:9] # 布局全部9个单元格 (1-9) - } - - # 3. Load and Filter Images - valid_images = [] - image_paths = [item['image_path'] for item in outfit_items] - for path in image_paths: - try: - # We use Image.open() and convert to 'RGB' to handle potential transparency (RGBA) - # and ensure compatibility with the final 'RGB' canvas and JPG output. - img = oss_get_image(oss_client=minio_client, path=f"{MINIO_LC_DATA_PATH}/{path}", data_type="PIL").convert('RGB') - # img = Image.open(path).convert('RGB') - valid_images.append(img) - except Exception as e: - logger.error(f"Error loading image {path}. Skipping: {e}") - - num_images = len(valid_images) - - if num_images == 0: - raise ValueError("No valid images were loaded.") - - if num_images > max_len: - raise ValueError(f"Valid item number {num_images} exceed max limit {max_len}") - - # Get the correct list of target areas based on the number of valid images - target_areas = quadrants.get(num_images, []) - - # 4. Resize and Paste - for i, (img, item) in enumerate(zip(valid_images, outfit_items)): - item_id = item['item_id'] - category = item['category'] - if i >= len(target_areas): - # This should not happen if num_images <= 4 - break - - # Target area dimensions (x_start, y_start, width, height) - x_start, y_start, target_w, target_h = target_areas[i] - - # Calculate new size while maintaining aspect ratio - original_w, original_h = img.size - - # Calculate the ratio needed to fit within the target area - ratio_w = target_w / original_w - ratio_h = target_h / original_h - - # Use the *smaller* of the two ratios to ensure the image fits entirely - resize_ratio = min(ratio_w, ratio_h) - - # Calculate the new dimensions - new_w = int(original_w * resize_ratio) - new_h = int(original_h * resize_ratio) - - # Resize the image. Image.Resampling.LANCZOS provides high-quality scaling. - # Pillow documentation recommends ANTIALIAS or BICUBIC for downscaling, - # but LANCZOS is a good general high-quality filter. - # Note: In Pillow versions > 9.0.0, Image.LANCZOS is now Image.Resampling.LANCZOS - resized_img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) - - # Calculate the paste position to center the resized image within its target area - # Center X: (Target Width - New Width) / 2 + X Start - paste_x = (target_w - new_w) // 2 + x_start - # Center Y: (Target Height - New Height) / 2 + Y Start - # paste_y = (target_h - new_h) // 2 + y_start - - TEXT_RESERVE_HEIGHT = 30 - paste_y = (target_h - new_h - TEXT_RESERVE_HEIGHT) // 2 + y_start - paste_y = max(paste_y, y_start) - - # Paste the resized image onto the canvas - canvas.paste(resized_img, (paste_x, paste_y)) - - full_text = f"ID: {item_id}, Category: {category}" - try: - # 推荐使用：计算文本的实际尺寸 (width, height) - bbox = draw.textbbox((0, 0), full_text, font=font) - text_w = bbox[2] - bbox[0] - text_h = bbox[3] - bbox[1] - except AttributeError: - # 兼容旧版本 Pillow - text_w, text_h = draw.textsize(full_text, font=font) - - # 计算 X 轴起始位置：使其在目标区域 (target_w) 中居中 - text_x_center = x_start + target_w // 2 - text_x_start = text_x_center - text_w // 2 - - # 计算 Y 轴起始位置：将其放在目标区域的底部 - # (目标区域的起始Y + 目标区域的高度 - 文本行的高度) - text_y_start = y_start + target_h - text_h - 5 # 减去 5 像素作为边距 - - # 3. 绘制合并后的文本 - if add_text: - draw.text((text_x_start, text_y_start), - full_text, - fill='black', - font=font) - - # Save as a high-quality JPG (quality=90 is a good balance) - # canvas.save(output_path, 'JPEG', quality=90) - - return canvas diff --git a/app/server/ChatbotAgent/core/vector_database.py b/app/server/ChatbotAgent/core/vector_database.py index a7b141b..5ff1c26 100644 --- a/app/server/ChatbotAgent/core/vector_database.py +++ b/app/server/ChatbotAgent/core/vector_database.py @@ -1,18 +1,28 @@ import random import time +import numpy as np import torch import chromadb from PIL import Image from typing import List, Dict, Any from transformers import CLIPProcessor, CLIPModel +from app.taxonomy import CATEGORY, OCCASION + class VectorDatabase(): def __init__(self, vector_db_dir: str, collection_name: str, embedding_model_name: str): self.client = chromadb.PersistentClient(path=vector_db_dir) - self.collection = self.client.get_or_create_collection(name=collection_name) + self.collection = self.client.get_or_create_collection( + name=collection_name, + configuration={ + "hnsw": { + "space": "cosine", + } + } + ) self.device = "cuda" if torch.cuda.is_available() else "cpu" @@ -48,25 +58,87 @@ class VectorDatabase(): return features.cpu().numpy().flatten().tolist() - def query_local_db(self, embedding: List[float], category: str, n_results: int = 3) -> List[Dict[str, Any]]: + def query_local_db(self, embedding: List[float], category: str, occasions: List[str] = [], n_results: int = 3) -> List[Dict[str, Any]]: """ 基于嵌入向量在本地数据库中查询相似单品。 - 实际应执行 ChromaDB 查询,并根据 category 进行过滤(metadatas)。 + 实际应执行 ChromaDB 查询，并根据 category 进行过滤(metadatas)。 """ - # 实际应执行向量查询 - # 为了演示流程,返回一个模拟结果 + for occasion in occasions: + where_clauses = { + "$and": [ + {"category": category}, + {"modality": "image"}, + {"batch_source": '2025_q4'} + ] + } + if occasion not in OCCASION: + continue + else: + where_clauses['$and'].append({occasion: 1}) + + results = self.collection.query( + query_embeddings=[embedding], + n_results=n_results, + where=where_clauses, + include=['metadatas', 'distances'] + ) + return results + + def get_matched_item(self, embedding: List[float], category: str, occasions: List[str] = [], batch_source: str = "2025_q4", gender: str = 'female', n_results: int = 1) -> List[Dict[str, Any]]: results = self.collection.query( query_embeddings=[embedding], - n_results=n_results, + n_results=500, where={ "$and": [ {"category": category}, {"modality": "image"}, + {"gender": gender}, + {"batch_source": batch_source} ] }, - include=['documents', 'metadatas', 'distances'] + include=['metadatas', 'distances'] ) - return results + if not results['ids'][0]: + return [] + + metadatas = results['metadatas'][0] # List[Dict[str, Any]] + final_scores = [] + for idx, metadata in enumerate(metadatas): + dist_img = results['distances'][0][idx] + score_vec = 1 - dist_img # cosine similarity range: [-1, 1] + + score_occ = 0.0 + if occasions: + count = 0 + for occ in occasions: + if occ not in OCCASION: + continue + count += 1 + status_val = metadata.get(occ, -1) + if status_val == 1: + score_occ += 1.0 + elif status_val == 0: + score_occ += 0.0 + else: + score_occ -= 100.0 + + score_occ = score_occ / count if count else 0.0 + + final_score = 0.6 * score_vec + 0.3 * score_occ + final_scores.append(final_score) + + scores_arr = np.array(final_scores) + temperature = 0.5 + scores_arr = scores_arr / temperature + + # Softmax: 将分数转换为概率 + exp_scores = np.exp(scores_arr - np.max(scores_arr)) + probabilities = exp_scores / np.sum(exp_scores) + + # 采样 (或直接取 Top 1) + sampled_index = np.random.choice(a=len(results['ids'][0]), p=probabilities, size=n_results, replace=False) # 不重复采样 + sampled_items = [metadatas[i] for i in sampled_index] + return sampled_items def load_filtered_ids(self, filter_item): # print("\n--- 初始化阶段：加载所有符合条件的 ID ---") diff --git a/app/core/utils_litserve.py b/app/server/utils/img_operation.py similarity index 97% rename from app/core/utils_litserve.py rename to app/server/utils/img_operation.py index 7cf0e07..68a6f21 100644 --- a/app/core/utils_litserve.py +++ b/app/server/utils/img_operation.py @@ -1,8 +1,10 @@ import logging +import os from typing import List, Dict from PIL import Image, ImageDraw, ImageFont from app.server.utils.minio_client import oss_get_image, minio_client from app.server.utils.minio_config import MINIO_LC_DATA_PATH +from app.config import settings logger = logging.getLogger(__name__) # 9个 341x341 左右的单元格 (ALL_9_CELLS) @@ -74,7 +76,11 @@ def merge_images_to_square(outfit_items: List[Dict[str, str]], max_len=9, add_te try: # We use Image.open() and convert to 'RGB' to handle potential transparency (RGBA) # and ensure compatibility with the final 'RGB' canvas and JPG output. - img = oss_get_image(oss_client=minio_client, path=f"{MINIO_LC_DATA_PATH}/{path}", data_type="PIL").convert('RGB') + if settings.LOCAL == 1: + image_file_path = os.path.join(settings.LOCAL_IMAGE_DIR, path) + img = Image.open(image_file_path).convert('RGB') + else: + img = oss_get_image(oss_client=minio_client, path=f"{MINIO_LC_DATA_PATH}/{path}", data_type="PIL").convert('RGB') # img = Image.open(path).convert('RGB') valid_images.append(img) except Exception as e: diff --git a/app/taxonomy.py b/app/taxonomy.py new file mode 100644 index 0000000..d8dd259 --- /dev/null +++ b/app/taxonomy.py @@ -0,0 +1,19 @@ +# 这个文件用来储存所有的category和occasion，这是标准文件。 + +CATEGORY = [ + 'shoes', 'bags', 'dresses', 'tops', 'pants', 'skirts', 'outerwear', 'swimwear', 'suits', + 'watches', 'sunglasses', 'belts', 'hats', 'jewelry', 'neckties', 'scarves & shawls' +] +CLOTHING_CATEGORY = [ + 'shoes', 'bags', 'dresses', 'tops', 'pants', 'skirts', 'outerwear', 'swimwear' +] +ACCESSORY_CATEGORY = [ + 'watches', 'sunglasses', 'belts', 'hats', 'jewelry', 'neckties', 'scarves & shawls' +] +OCCASION = [ + "Casual", "Formal", "Activewear", "Resort", "Evening", "Outdoor", + "Business / workwear", "Cocktail / Semi-Formal", "Black Tie / White Tie", + "Bridal / Wedding", "Festival / Concert", "Party / Clubbing", + "Travel / Transit", "Athleisure", "Beach / Swim", "Ski / Snow / Mountain", + "Garden Party / Daytime Event" +] diff --git a/data/stylist_guide/mini_en.md b/data/stylist_guide/mini_en.md deleted file mode 100644 index 8ef1a54..0000000 --- a/data/stylist_guide/mini_en.md +++ /dev/null @@ -1,43 +0,0 @@ -# Outfit Style Guide - -This guide summarizes the stylist's latest preferences, prohibitions, and styling patterns. - -## I. Core Preferences and Prohibitions - -| Category | Preference | Prohibition | -| :--- | :--- | :--- | -| **Primary Colors** | **Black, White, Gray, Earth Tones** (Solid colors preferred) | Avoid **Yellow** | -| **Accent Colors** | Added via **bags, scarves** (Max 2 accent colors) | Avoid **Bright/Vivid** colors dominating | -| **Patterns** | **Plaid/Checkered, Stripes** (Used as accents) | Strictly avoid **Floral/Small Prints** | -| **Fit/Silhouette** | **Loose, Comfortable, Oversize**; **Clean and Sharp** tailoring | Tight-fitting or complicated silhouettes | -| **Pants Requirement** | Must be **Floor-Length/Puddle Pants** | **Cropped/Ankle-Length, Flare/Bootcut Pants** | -| **Footwear** | **White Sneakers**, **Flats/Single Shoes**, **Loafers** | **Heels, Tall Boots** | -| **Leather Bags** | Must be **Black or White** pure colors; Preference for **Small, Neat Black Top-Handle Bags** | Bulky/Large volume bags | -| **Accessories** | **Multiple Gold Accessories layered** (Necklace, Bracelet/Bangle, Earrings mandatory); **Watch** (Mandatory) | **Large/Bulky Earrings or Vector-style** accessories | - ---- - -## II. Styling Pattern: Layering and Balance - -The stylist's outfits emphasize **comfort** and **layering** (creating depth). The common structure is: **3 Upper Body Items + 1 Lower Body Item**. Primary colors are solid, with subtle patterns and accessories adding polish. A **Black Floor-Length Slip Dress** is a key base layer option. - -### 1. Classic Outfit Examples - -| Style | Structure | Keywords | -| :--- | :--- | :--- | -| **Smart Casual** | Milk Tea Oversize Blazer + White Shirt + **Loose Black Floor-Length Pants** + Tie + White Sneakers | Polished, Layered, Puddle Pants | -| **Short Jacket Look** | **Short Jacket/Blazer** + **Short Skirt/Shorts** + White Socks + **Black Loafers** + Small Black Top-Handle Bag | Short-Jacket Rule, Sharp, Black Accents | -| **Layered Base** | Loose Striped Knitwear + **Black Floor-Length Slip Dress (Inner)** + **Layered Gold Accessories** + Loafers | Dress Base, Longline, Gold Jewelry | -| **Sporty Chic** | Milk Tea Oversize Blazer + Black Sleeveless Activewear + **Black Floor-Length Legging** + White Sneakers + Black Fanny Pack | Mixed Style, Dynamic, Balanced | - -### 2. Outfit Extended Rules Summary - -1. **Color Palette**: Main garments limited to Black, White, Gray, and Earth Tones; accent colors in accessories max 2. Avoid yellow or vivid dominance. -2. **Layering Principle**: Min 2-3 layers on the upper body, with visible subtle patterns (collar/scarf). **Key:** The **Black Floor-Length Slip Dress** is a favored base. -3. **Fit Requirements**: All items favor loose/oversize for comfort; tailoring must be clean, avoiding tight or complex silhouettes. -4. **Pattern Restriction**: Only Plaid/Stripes as accents; strictly no florals. -5. **Pant Length**: All pants **must be floor-length**, prohibiting cropped and flare styles. -6. **Short Outerwear Rule**: Short outerwear must be paired with **short skirts or shorts**, and **Black Loafers** are mandatory for footwear. -7. **Shoe/Bag Coordination**: Footwear is flat and casual (Loafers, White Sneakers preferred). Leather bags are pure black/white, with a preference for a **small, neat black top-handle bag**. -8. **Accessory Stacking**: **Gold necklace, bracelet/bangle, and earrings should ideally all be present and layered**; watch is mandatory; rings are minimal. -9. **Overall Balance**: Simple bottoms when the top is layered/complex; overall style is neutral and polished. \ No newline at end of file diff --git a/new_.json b/new_.json deleted file mode 100644 index 8f9c80a..0000000 --- a/new_.json +++ /dev/null @@ -1,13 +0,0 @@ -{ - "type": "service_account", - "project_id": "aida-461108", - "private_key_id": "e0fa4eb8743342ee0d9af77296ec71101bf09706", - "private_key": "-----BEGIN PRIVATE KEY-----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\nxJjMu9TnfDjmd2gG5GkWVqM=\n-----END PRIVATE KEY-----\n", - "client_email": "aida-gcs-test@aida-461108.iam.gserviceaccount.com", - "client_id": "117280364971550320059", - "auth_uri": "https://accounts.google.com/o/oauth2/auth", - "token_uri": "https://oauth2.googleapis.com/token", - "auth_provider_x509_cert_url": "https://www.googleapis.com/oauth2/v1/certs", - "client_x509_cert_url": "https://www.googleapis.com/robot/v1/metadata/x509/aida-gcs-test%40aida-461108.iam.gserviceaccount.com", - "universe_domain": "googleapis.com" -} diff --git a/stylist_guide/crystal_acc.md b/stylist_guide/crystal_acc.md new file mode 100644 index 0000000..54730a7 --- /dev/null +++ b/stylist_guide/crystal_acc.md @@ -0,0 +1,17 @@ +# Crystal's Accessory Guide: Pure Balance +This guide strictly outlines accessory selection, emphasizing Gold Tones and pure, solid colors to stabilize the outfit's primary focus on bold pattern clashing. + +## I. Color, Pattern, and Material Constraints +Color & Tone: All jewelry must be in Gold Tones. Vector-style accessories are prohibited. + +Accessory Pattern: Bags and shoes must be Solid Color only. Patterned or printed bags and shoes are strictly prohibited. + + +## II. Mandatory & Stacking Requirements +The use of accessories is essential to complete the look, focusing on stacking and specific shapes: + +Mandatory Jewelry: A Necklace is mandatory (minimum one piece). A Watch must be included as part of the wrist stack. + +Earrings: Must be Hoop Earrings. + +Stacking: Encourage stacking of Bracelets and Rings alongside the mandatory watch. \ No newline at end of file diff --git a/data/stylist_guide/crystal_en.md b/stylist_guide/crystal_en.md similarity index 50% rename from data/stylist_guide/crystal_en.md rename to stylist_guide/crystal_en.md index 2842fd2..3ffaa86 100644 --- a/data/stylist_guide/crystal_en.md +++ b/stylist_guide/crystal_en.md @@ -1,25 +1,23 @@ # Outfit Style Guide -This guide outlines the preferred styling logic, brand affinities, patterns, and structure for Crystal's outfits, emphasizing **bold pattern mixing** balanced by **pure accessories**. +This guide outlines the preferred styling logic, brand affinities, patterns, and structure for Crystal's outfits, emphasizing **bold pattern mixing** balanced by . ## I. Core Preferences and Prohibitions -| Category | Preference | Prohibition | +| Category | Preference (✔️) | Prohibition (❌) | | :--- | :--- | :--- | | **Brands/Material** | **Sacai** brand; **Denim** items | - | -| **Patterns** | **Plaid, Stripes, Floral, Leopard Print**; Active pattern clashing (min 2, **with specific restrictions**) | Monochromatic or single-pattern outfits; **Mixing different Animal Prints** (e.g., Leopard + Snake) | +| **Patterns** | **Plaid, Stripes, Floral, Leopard Print**; Active pattern clashing (min 2) | Monochromatic or single-pattern outfits | | **Layering** | Max **2 items** on the upper body; Pattern mixing replaces layering | Excessive layering (叠穿) | | **Fit/Silhouette** | Accepts **Oversize**; Flexible (can mix with slim-fit/leggings) | Tight-fitting or complicated silhouettes | | **Shoe Styles** | **Boots, Platform Shoes, Pointed Low Heels, Mesh Shoes, Ballet Flats, White Sneakers** | Full High Heels (高跟鞋), Tall Boots | -| **Bags/Shoes (Pattern)** | **Solid Color** only for bags and shoes | **Patterned/Printed** Bags or Shoes | -| **Bags (Material)** | Must be **Leather** (or natural elements like Rattan/Wicker) | Non-leather materials (except rattan/wicker) | -| **Accessories** | **Gold Tones**; **Necklace** (Mandatory); Rings, **Hoop Earrings**, Bracelet, **Watch** | Vector-style accessories | +| **Shoes (Pattern)** | **Solid Color** only for shoes | **Patterned/Printed** Shoes | --- -## II. Styling Pattern: Pattern Clash and Pure Balance +## I. Styling Pattern: Pattern Clash and Pure Balance -This stylist's style prioritizes visual impact through **clashing prints** rather than layering. Outfits are usually based on a bold print or denim, balanced by pure, solid-colored accessories. +This stylist's style prioritizes visual impact through **clashing prints** rather than layering. Outfits are usually based on a bold print or denim, balanced by pure. ### 1. Classic Outfit Examples @@ -28,15 +26,14 @@ This stylist's style prioritizes visual impact through **clashing prints** rathe | **Edgy Casual** | Milk Tea Oversize Blazer + Beige Shirt + Olive Green Cargo Pants + Black Pointed Low Heels | Utility, Sharp Contrast, Subtle Prints | | **Elegant Flow** | Milk Tea Oversize Blazer + Beige Shirt + Dark Brown Slim-fit Pants + White Mesh Shoes or Ballet Flats | Mixed Fit, Comfort Contrast | | **Sporty Mix** | Milk Tea Oversize Blazer + Dark Brown or Leopard Print Yoga Set + White Sneakers | Athleisure, Pattern Pop | -| **Heavy Print** | Floral Dress + Denim Pants + Rattan Bag + Wrist Cord + Layered Beaded Necklaces | Pattern Clash (Floral + Denim), Natural Accent | +| **Heavy Print** | Floral Dress + Denim Pants | Pattern Clash (Floral + Denim), Natural Accent | ### 2. Outfit Extended Rules Summary -1. **Pattern Clash**: Actively mix patterns (Plaid/Stripe/Leopard) in one outfit, aiming for at least two non-forbidden prints. **CRITICAL RESTRICTIONS:** A) If a **Floral** print item is chosen, all other clothing items must be solid colors (no mixing floral with other prints). B) **Do not mix different animal prints** (e.g., Leopard and Snakeskin cannot be combined). +1. **Pattern Clash**: Actively mix **at least two** patterns (Plaid/Stripe/Floral/Leopard) in one outfit to maximize visual interest. 2. **Layering Limit**: Do not rely on multi-layering for depth; use print complexity instead. The upper body is limited to **max 2 pieces**. 3. **Fit Flexibility**: Oversize is welcome, but tight-fitting items (e.g., leggings) can be mixed. Prioritize durable fabrics like **Denim** for texture. -4. **Color Base**: No strict color restrictions on garments, but all major accessories (**bags and shoes**) must be **pure, solid colors**. -5. **Shoe/Bag Principle**: Shoes should provide height (low heel, platform, boots). Bags should be **leather and solid-colored**, with rattan/wicker bags accepted as natural accents. -6. **Accessory Requirements**: **Gold** jewelry is preferred for unified tone. **Necklace** is mandatory (min 1). Earrings must be **hoops**. Stack **bracelet/ring/watch**. -7. **Overall Balance**: If clothing patterns are complex, shoes/bags must be **simple and pure** to ground the look. Style leans toward **mixed-casual and energetic**, avoiding blandness. +4. **Color Base**: No strict color restrictions on garments, but shoes must be **pure, solid colors**. +5. **Shoe Principle**: Shoes should provide height (low heel, platform, boots). +6. **Overall Balance**: If clothing patterns are complex, shoes must be **simple and pure** to ground the look. Style leans toward **mixed-casual and energetic**, avoiding blandness. 8. **Scene Adaptability**: Add **cargo pants/low heels** for casual settings; **yoga sets** for sportier looks; use **denim** to balance out heavy prints. \ No newline at end of file diff --git a/stylist_guide/edi_acc.md b/stylist_guide/edi_acc.md new file mode 100644 index 0000000..9316169 --- /dev/null +++ b/stylist_guide/edi_acc.md @@ -0,0 +1,10 @@ +# Accessory Style Guide +This guide strictly outlines accessory selection, emphasizing a flexible approach to metal tones and highlighting the aesthetic of long-worn items. + +# I. Color and Material Constraints +Metal Tones: Both Gold and Silver metals are preferred and should be mixed and matched together. + +Aesthetic Preference: Items that show wear, such as Silver items that have changed color over time, are acceptable as they present a unique personal preference. + +# II. Mandatory & Stacking Requirements +Mandatory Items: No specific jewelry piece is listed as mandatory, but the style encourages mixing both gold and silver jewelry. \ No newline at end of file diff --git a/stylist_guide/edi_en.md b/stylist_guide/edi_en.md new file mode 100644 index 0000000..1f20d1b --- /dev/null +++ b/stylist_guide/edi_en.md @@ -0,0 +1,20 @@ +# Outfit Style Guide +This guide summarizes the preferred styling logic, colors, patterns, and structure for Edi's outfits, emphasizing functionality and a flexible approach to color and print mixing. + +## I. Core Preferences and Prohibitions +Primary Colors: Black is preferred as a mixing base. + +Dominant Colors: Maximum of 2 dominant colors per outfit (excluding Black and White). + +Print Mixing: Prints are mixed, specifically by using the same pattern (e.g., stripes, plaid) but in different sizes (e.g., thin stripe mixed with wide stripe, large check with small check). + +Prohibited Prints: None. All prints are acceptable, depending on the combination. + +Shoes: Sneakers are worn most often for comfort and versatility. + +Bags: Functional bags that can hold "everything" are preferred over mini bags. A specific favorite is the Lemaire croissant leather bag in a big size. + +Color Mixing: All color tones are acceptable for mixing together, including highly varied tones (e.g., rainbow colors), as long as they create a mutual harmony. + +## II. Styling Pattern: Functional and Harmonious +This stylist prioritizes practical, comfortable items (functional bag, sneakers) while embracing complex color and print compositions that lean on Black as a foundational element. The style aims for a unique, harmonious look achieved through flexible mixing. \ No newline at end of file diff --git a/stylist_guide/mini_acc.md b/stylist_guide/mini_acc.md new file mode 100644 index 0000000..d21ae19 --- /dev/null +++ b/stylist_guide/mini_acc.md @@ -0,0 +1,12 @@ +Stylist Accessories Guide +I. Gold Jewelry +Preference: Necklace, Bracelet, and Earrings must all be present and layered. +Prohibition: Avoid Large/Bulky Earrings or Vector-style accessories. + +II. Watch +Preference: Mandatory item. +Prohibition: N/A. + +III. Accent Colors +Preference: Added via accessories (e.g., scarves); Max 2 accent colors in total. +Prohibition: Avoid Bright/Vivid colors dominating the outfit. \ No newline at end of file diff --git a/stylist_guide/mini_en.md b/stylist_guide/mini_en.md new file mode 100644 index 0000000..471bb6e --- /dev/null +++ b/stylist_guide/mini_en.md @@ -0,0 +1,37 @@ +# Outfit Style Guide + +This guide summarizes the preferred styling logic, colors, silhouettes, accessories, and layering patterns favored by the stylist. + +## I. Core Preferences and Prohibitions + +| Category | Preference (✔️) | Prohibition (❌) | +| :--- | :--- | :--- | +| **Primary Colors** | **Black, White, Gray, Earth Tones** (Solid colors preferred) | Avoid **Yellow** | +| **Accent Colors** | | Avoid **Bright/Vivid** colors dominating | +| **Patterns** | **Plaid/Checkered, Stripes** (Used as accents) | Strictly avoid **Floral/Small Prints (碎花)** | +| **Style/Fit** | **Loose, Comfortable, Oversize**; **Clean and Sharp** tailoring | Tight-fitting or complicated silhouettes | +| **Footwear** | **White Sneakers**, **Flats/Single Shoes**, **Loafers** | **Heels, Tall Boots** | + +--- + +## II. Styling Pattern: Layering and Balance + +This stylist's outfits emphasize **comfort** and **layering** (creating depth). The common structure is: **3 Upper Body Items + 1 Lower Body Item**. The pattern mainly uses solid primary colors, mixed with subtle patterns, aiming for a look that is **relaxed and casual yet polished**. + +### 1. Classic Outfit Examples + +| Style | Structure | Keywords | +| :--- | :--- | :--- | +| **Smart Casual** | Milk Tea Oversize Blazer + White Shirt + Loose Black Pants + White Sneakers | Neutral, Polished, Clean | +| **Everyday Casual** | Milk Tea Oversize Blazer + Dark Blue Jeans or Dark Gray Shorts + **Loafers**+ Detailed Vest/Tank Top | Relaxed, Comfortable, Detailed | +| **Sporty Chic** | Milk Tea Oversize Blazer + Black/Blue & White Striped Sleeveless Top + Black Leggings + White Sneakers | Mixed Style, Dynamic, Balanced | +| **Dress Base** | Floral Dress (**Exception**) + White Sneakers | Comfortable, Accents, Exception Handling | + +### 2. Outfit Extended Rules Summary + +1. **Color Palette**: Main garment colors are limited to **Black, White, Gray, and Earth Tones** to ensure a minimalist foundation. +2. **Layering Principle**: A minimum of **2-3 layers** on the upper body, with patterned edges (e.g., shirt collar) visible to add depth without complexity. +3. **Fit Requirements**: All items should lean toward a **loose/oversize** fit to prioritize comfort; tailoring must be **clean and sharp**. +4. **Pattern Restriction**: Only **plaid/checkered or stripes** are acceptable as subtle accents; **strictly no florals**, unless an extremely minimal exception is made. +5. **Shoe/Bag Coordination**: Footwear must be **flat and casual** (white sneakers are preferred). +6. **Overall Balance**: When the upper body is complex (layered), the lower body should remain **simple**; the overall style is **neutral and polished**, avoiding highly feminine heels or boots. \ No newline at end of file diff --git a/stylist_guide/vera_acc.md b/stylist_guide/vera_acc.md new file mode 100644 index 0000000..1fb1e1a --- /dev/null +++ b/stylist_guide/vera_acc.md @@ -0,0 +1,9 @@ +# Accessory Style Guide +This guide outlines accessory selection based on the desired overall aesthetic, emphasizing a balanced approach to metal tones. + +# I. Metal Tone and Aesthetic Constraints +Jewelry Tone: Prefers Gold for a vintage and nostalgic feel. + +Jewelry Tone: Prefers Silver for grungier looks. + +Wear Preference: Wears both Gold and Silver, depending on the outfit's desired aesthetic. diff --git a/stylist_guide/vera_en.md b/stylist_guide/vera_en.md new file mode 100644 index 0000000..afb6a4b --- /dev/null +++ b/stylist_guide/vera_en.md @@ -0,0 +1,24 @@ +# Outfit Style Guide +This guide summarizes the preferred styling logic, colors, patterns, and structure for Vera's outfits, emphasizing harmony in color mixing and the use of statement bags. + +# I. Core Preferences and Prohibitions +Primary Colors: Most often wears Khakis, Black, Creams, and sometimes Burgundies. + +Dominant Colors: Maximum of two dominant colors per outfit, handled with care. + +Print Mixing: Prints are mixed. Acceptable combinations include big polka dots with small polka dots or different floral prints. + +Print Rule: Prints must be mixed only as long as the colors and shapes are harmonized. + +Prohibited Prints: None. Most prints are considered wearable. + +Color Mixing Prohibition: Avoid mixing warms and colds, and brights with muted tones. + +Shoes: Wears loafers most often. + +Bags: Prefers Bigger bags and bags that make a statement. Specific examples include The Row Margaux and Saint Laurent Icare. + +Prohibited Bags: Not a fan of crossbody or micro mini bags. + +# II. Styling Pattern: Harmonious Statement +This stylist prioritizes a core color palette of neutrals and deep tones (Khakis, Black, Creams, Burgundies) and uses print mixing (e.g., different sized polka dots or florals) only when colors and shapes are harmonized. The overall look is anchored by comfortable shoes (loafers) and a large, functional, statement bag. \ No newline at end of file