embedding_test.py

import os
import logging
from abc import ABC, abstractmethod
from typing import List, Dict, Any
from sentence_transformers import SentenceTransformer
from pymilvus import MilvusClient, DataType
import time
import gradio as gr

from app.config import MILVUS_DB_URL

# 配置日志
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s %(levelname)s %(message)s'
)
logger = logging.getLogger(__name__)

models = [
    'shibing624/text2vec-base-chinese',
    'BAAI/bge-small-zh',
    'BAAI/bge-base-zh',
    'sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2',
    'all-MiniLM-L6-v2',
    'all-MiniLM-L12-v2',
    'multi-qa-mpnet-base-dot-v1',
    # 'bge-small-en-v1.5', 不兼容
    'all-mpnet-base-v2',
    'jinaai/jina-embeddings-v3',
]


searchers = {}


class BaseEmbeddingModel(ABC):
    @abstractmethod
    def encode(self, text: str) -> List[float]:
        pass

    @property
    @abstractmethod
    def dimension(self) -> int:
        pass

    @property
    @abstractmethod
    def model_name(self) -> str:
        pass

class SentenceTransformerModel(BaseEmbeddingModel):
    def __init__(self, model_name: str):
        self.model = SentenceTransformer(model_name, trust_remote_code=True)
        self._model_name = model_name

    def encode(self, text: str) -> List[float]:
        result = self.model.encode(text).tolist()
        return result

    @property
    def dimension(self) -> int:
        return self.model.get_sentence_embedding_dimension()

    @property
    def model_name(self) -> str:
        return self._model_name

class StickerSearcher:
    def __init__(self, model: BaseEmbeddingModel):
        self.model = model
        self.client = MilvusClient(uri=MILVUS_DB_URL)
        self.collection_name = f'test_{model.model_name.replace("/", "_").replace("-", "_")}'

    def init_collection(self) -> bool:
        try:
            self.client.drop_collection(collection_name=self.collection_name)
            self.client.create_collection(
                collection_name=self.collection_name,
                dimension=self.model.dimension,
                primary_field_name='id',
                auto_id=True
            )
            self.client.create_index(
                collection_name=self.collection_name,
                index_type='IVF_SQ8',
                metric_type='COSINE',
                params={'nlist': 128},
                index_params={}
            )
            self.client.load_collection(self.collection_name)
            logger.info(f'Collection initialized: {self.collection_name}')
            return True
        except Exception as e:
            logger.error(f'Collection init failed: {str(e)}')
            return False

    def store_vector(self, title: str, description: str, tags: List[str], file_path: str):
        vector = self.model.encode(description)
        data = [{
            'vector': vector,
            'title': title,
            'description': description,
            'tags': tags,
            'file_name': file_path
        }]
        self.client.insert(self.collection_name, data)

    def search(self, query: str, limit: int = 5) -> List[Dict[str, Any]]:
        start_time = time.time()
        query_vector = self.model.encode(query)
        encode_time = time.time() - start_time
        start_search_time = time.time()
        results = self.client.search(
            collection_name=self.collection_name,
            data=[query_vector],
            limit=limit,
            output_fields=['title', 'description', 'tags', 'file_name']
        )
        search_time = time.time() - start_search_time
        total_time = encode_time + search_time
        logger.info(f'模型 {self.model.model_name} Encoding耗时: ${encode_time:.4f}，搜索耗时: {search_time:.4f} 秒, 总耗时: {total_time:.4f} 秒')
        return results[0]

def create_gradio_ui():

    async def search_model(model_name: str, query: str):
        try:
            if model_name in searchers:
                return searchers[model_name].search(query)
            logger.error(f'Model not loaded: {model_name}')
            return []
        except Exception as e:
            logger.error(f'Search failed: {model_name} | Error: {str(e)}')
            return []

    async def search_all_models(query):
        if not query:
            return []

        print(f'>>>> Searching From Models {query}')
        results = []
        for model_name in models:
            result = await search_model(model_name, query)
            results.append(result)

        formatted_results = []
        max_results = max(len(r) for r in results)

        for i in range(max_results):
            row = [i + 1]
            for model_results in results:
                if i < len(model_results):
                    result = model_results[i]
                    image_url = f'https://huggingface.co/datasets/Nekoko/StickerSet/resolve/main/{result["entity"]["file_name"]}'
                    row.append(f'![Sticker]({image_url})\n相似度: {result["distance"]:.4f}')
                else:
                    row.append('-')
            formatted_results.append(row)
        return formatted_results

    def init_collections():
        try:
            client = MilvusClient(uri=MILVUS_DB_URL)
            stickers = client.query(
                collection_name='stickers',
                filter='',
                limit=1000,
                output_fields=['title', 'description', 'tags', 'file_name']
            )
            logger.info(f'Stickers loaded: {len(stickers)}')

            def init_model(model_name):
                try:
                    searcher = StickerSearcher(SentenceTransformerModel(model_name))
                    if searcher.init_collection():
                        searchers[model_name] = searcher
                        for sticker in stickers:
                            searcher.store_vector(
                                sticker.get('title'),
                                sticker.get('description'),
                                sticker.get('tags'),
                                sticker.get('file_name')
                                )
                        logger.info(f'Model initialized: {model_name}')
                except Exception as e:
                    logger.error(f'Model init failed: {model_name} | Error: {str(e)}')

            for model_name in models:
                print(f'>>>> 初始化模型 {model_name}')
                start_time = time.time()
                init_model(model_name)
                print(f'>>>> 初始化模型 {model_name} 完成 ✅，耗时 {time.time() - start_time:.4f} 秒')
            print(f'>>>> 初始化所有模型完成 ✅')
            return '初始化成功！'
        except Exception as e:
            logger.error(f'Data init failed: {str(e)}')
            return f'初始化失败: {str(e)}'

    with gr.Blocks(title='Neko Sticker Search 🔍', css='.gradio-container img { width: 200px !important; height: 200px !important; object-fit: contain; }') as demo:
        with gr.Row():
            search_input = gr.Textbox(label='搜索关键词')
            search_button = gr.Button('搜索')


        headers = ['序号'] + [f'🧊{model.split("/")[-1]}' for i, model in enumerate(models)]
        results_table = gr.Dataframe(
            headers=headers,
            datatype=['number'] + ['markdown'] * len(models),
            row_count=5,
            col_count=len(models) + 1
        )

        status_box = gr.Textbox(label='状态', interactive=False)
        refresh_button = gr.Button('刷新数据')

        refresh_button.click(fn=init_collections, outputs=status_box)
        # 由于这里只是简单的搜索操作，可以直接使用同步方式调用
        search_button.click(
            fn=search_all_models, 
            inputs=[search_input], 
            outputs=results_table
)

    return demo

if __name__ == '__main__':

    # 提前加载所有模型
    start_time = time.time()
    for index, model_name in enumerate(models):
        try:
            start_time = time.time()
            searchers[model_name] = StickerSearcher(SentenceTransformerModel(model_name))
            print(f'>>>> 预加载模型 {model_name} 完成 ✅, 耗时 {time.time() - start_time:.4f} 秒')
        except Exception as e:
            logger.error(f'Model preload failed: {model_name} | Error: {str(e)}')

    logger.info(f'>>>> 预加载模型完成 ✅: {models}, 耗时 {time.time() - start_time:.4f} 秒')



    demo = create_gradio_ui()
    demo.launch()