graph_rag_server/src/model/multimodal_embedding.py

from typing import List
from PIL import Image
import base64
import io
from langchain_openai import OpenAIEmbeddings
from dashscope import MultiModalEmbedding
from src.conf.settings import model_settings
from langfuse import observe
from src.utils.file.image_util import image_util

class Embedding:
    """Embedding模型工具"""
    
    def __init__(self, model_name: str = None, api_key: str = None):
        """
        初始化Embedding模型
        
        Args:
            model_name: 模型名称，若为None则使用配置文件中的值
            api_key: API密钥，若为None则使用配置文件中的值
        """
        # 获取模型配置
        self.model_provider = model_settings.model_provider
        self.embedding_model_name = model_name or model_settings.embedding_model_name
        self.multi_embedding_model_name = model_settings.multimodal_embedding_model_name
        self.base_url = model_settings.base_url
        self.api_key = api_key or model_settings.api_key
        self.dashscope_api_key =  model_settings.dashscope_api_key
    
    @observe(name="text_embedding", as_type="embedding")
    def get_text_embedding(self, text: str) -> List[float]:
        """
        获取文本的embedding
        
        Args:
            text: 要获取embedding的文本
            
        Returns:
            List[float]: 文本的embedding向量
        """
        try:
            # 使用langchain_openai初始化OpenAI Embeddings模型
            embeddings = OpenAIEmbeddings(
                model=self.embedding_model_name,
                base_url=self.base_url,
                api_key=self.api_key
            )
            embedding = embeddings.embed_query(text)
            return embedding
        except Exception as e:
            raise Exception(f"文本embedding生成失败: {str(e)}")
    
    @observe(name="texts_embedding", as_type="embedding")
    def get_texts_embedding(self, texts: List[str]) -> List[List[float]]:
        """
        获取多个文本的embedding
        
        Args:
            texts: 要获取embedding的文本列表
            
        Returns:
            List[List[float]]: 文本列表的embedding向量列表
        """
        try:
            # 使用langchain_openai初始化OpenAI Embeddings模型
            embeddings = OpenAIEmbeddings(
                model=self.embedding_model_name,
                base_url=self.base_url,
                api_key=self.api_key
            )
            embeddings = embeddings.embed_documents(texts)
            return embeddings
        except Exception as e:
            raise Exception(f"多个文本embedding生成失败: {str(e)}")
    
    @observe(name="image_embedding", as_type="embedding")
    def get_image_embedding(self, image: Image.Image) -> List[float]:
        """
        获取图像的embedding
        
        Args:
            image: PIL图像对象
            
        Returns:
            List[float]: 图像的embedding向量
        """
        try:
            # 将图像转换为base64
            buffer = io.BytesIO()
            image.save(buffer, format="PNG")
            buffer.seek(0)
            
            # 压缩图片字节流
            compressed_bytes = image_util._compress_image_to_bytes(buffer)
            
            image_base64 = base64.b64encode(compressed_bytes).decode("utf-8")
            # 构建输入项
            item = [
                {
                    "image": f"data:image/png;base64,{image_base64}"
                }
            ]
            response = MultiModalEmbedding.call(
                model=self.multi_embedding_model_name,
                api_key=self.dashscope_api_key,
                input=item
            )
            if response.status_code == 200:
                return response.output["embeddings"][0]["embedding"]
        except Exception as e:
            raise Exception(f"图像embedding生成失败: {str(e)}")

    @observe(name="multimodal_embedding", as_type="embedding")
    def get_multimodal_embedding(self, text: str, image: Image.Image) -> List[float]:
        """
        获取多模态（文本+图像）的embedding
        
        Args:
            text: 文本内容
            image: PIL图像对象
            
        Returns:
            List[float]: 多模态的embedding向量
        """
        try:
            item = []
            if image is not None:
                buffer = io.BytesIO()
                image.save(buffer, format="PNG")
                buffer.seek(0)
            
                # 压缩图片字节流
                compressed_bytes = image_util._compress_image_to_bytes(buffer)
            
                image_base64 = base64.b64encode(compressed_bytes).decode("utf-8")
                item.append({'image': f"data:image/png;base64,{image_base64}"})
            # 判断text部位None或者不为空字符串
            if text is not None and text.strip() != "":
                item.append({'text': text})
        
            response = MultiModalEmbedding.call(
                model=self.multi_embedding_model_name,
                api_key=self.dashscope_api_key,
                input=item
            )
            
            if response.status_code == 200:
                return response.output["embeddings"][0]["embedding"]
            else:
                raise Exception(f"Error: {response.message}")   
        except Exception as e:
            raise Exception(f"多模态embedding生成失败: {str(e)}")

# 全局单例
def get_embedding_model() -> Embedding:
    """
    获取全局单例的Embedding模型
    
    Returns:
        Embedding: 全局单例的Embedding模型
    """
    return Embedding(
        model_name=model_settings.embedding_model_name,
        api_key=model_settings.api_key
    )