RagFlow_Plugs/services/model/qwen_vl.py

from typing import Dict, Any
from PIL import Image
import base64
import io
from langchain.chat_models import init_chat_model
from conf.config import ModelConfig

class QWenVLParser:
    """QWEN VL模型图像解析工具"""
    
    def __init__(self, model_name: str = None):
        """
        初始化QWEN VL模型解析器
        
        Args:
            model_name: 模型名称，若为None则使用配置文件中的值
        """
        # 获取模型配置
        self.model_provider = ModelConfig.get_model_provider()
        self.model_name = model_name or ModelConfig.get_model_name()
        self.base_url = ModelConfig.get_base_url()
        self.api_key = ModelConfig.get_api_key()
        
        # 使用langchain的init_chat_model初始化模型
        self.model = init_chat_model(
            model_provider=self.model_provider,
            model=self.model_name,
            base_url=self.base_url,
            api_key=self.api_key
        )
    
    def image_to_base64(self, image: Image.Image) -> str:
        """
        将PIL图像转换为base64编码字符串
        
        Args:
            image: PIL图像对象
            
        Returns:
            str: base64编码的图像字符串
        """
        buffer = io.BytesIO()
        image.save(buffer, format="PNG")
        return base64.b64encode(buffer.getvalue()).decode("utf-8")
    
    def parse_image(self, image: Image.Image, page_number: int, prompt: str = "请详细描述图像中的内容") -> Dict[str, Any]:
        """
        使用OpenAI模型解析图像内容
        
        Args:
            image: PIL图像对象
            page_number: 页码
            prompt: 提示词
            
        Returns:
            Dict: 包含解析结果的字典，包含:
                - page_number: 页码
                - content: 解析内容
                - model: 使用的模型名称
        """
        try:
            # 将图像转换为base64
            image_base64 = self.image_to_base64(image)
            
            # 构建消息，符合OpenAI API格式
            messages = [
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "text",
                            "text": prompt
                        },
                        {
                            "type": "image_url",
                            "image_url": {
                                "url": f"data:image/png;base64,{image_base64}"
                            }
                        }
                    ]
                }
            ]
            
            # 使用langchain模型调用OpenAI API
            response = self.model.invoke(messages)
            
            # 提取解析结果
            content = response.content
            
            return {
                "page_number": page_number,
                "content": content,
                "model": self.model_name
            }
        except Exception as e:
            raise Exception(f"图像解析失败（页码：{page_number}）: {str(e)}")
    
    def parse_image_bytes(self, image_bytes: io.BytesIO, page_number: int, prompt: str = "请详细描述图像中的内容") -> Dict[str, Any]:
        """
        使用OpenAI模型解析图像字节流
        
        Args:
            image_bytes: 图像字节流
            page_number: 页码
            prompt: 提示词
            
        Returns:
            Dict: 包含解析结果的字典
        """
        # 将字节流转换为PIL图像
        image = Image.open(image_bytes)
        return self.parse_image(image, page_number, prompt)