107 lines
3.9 KiB
Python
107 lines
3.9 KiB
Python
|
import io
|
|||
|
|
import logging
|
|||
|
|
import re
|
|||
|
|
import base64
|
|||
|
|
|
|||
|
|
from markitdown import MarkItDown
|
|||
|
|
|
|||
|
|
from docreader.models.document import Document
|
|||
|
|
from docreader.parser.base_parser import BaseParser
|
|||
|
|
from docreader.parser.chain_parser import PipelineParser
|
|||
|
|
from docreader.parser.markdown_parser import MarkdownParser
|
|||
|
|
|
|||
|
|
# 尝试导入 VLMClient
|
|||
|
|
try:
|
|||
|
|
from parser.vlm_client import VLMClient
|
|||
|
|
except ImportError:
|
|||
|
|
VLMClient = None
|
|||
|
|
|
|||
|
|
logger = logging.getLogger(__name__)
|
|||
|
|
|
|||
|
|
|
|||
|
|
class StdMarkitdownParser(BaseParser):
|
|||
|
|
"""
|
|||
|
|
Standard MarkItDown Parser Wrapper
|
|||
|
|
|
|||
|
|
This parser uses the markitdown library to convert various document formats
|
|||
|
|
(docx, pptx, pdf, etc.) into text/markdown.
|
|||
|
|
Optionally uses VLM to process images.
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
def __init__(self, *args, vlm_config=None, **kwargs):
|
|||
|
|
# 这里的 super() 会调用 BaseParser 的初始化，确保 self.file_type 被正确赋值
|
|||
|
|
super().__init__(*args, **kwargs)
|
|||
|
|
self.markitdown = MarkItDown()
|
|||
|
|
self.vlm_config = vlm_config
|
|||
|
|
self.vlm_client = None
|
|||
|
|
|
|||
|
|
# 如果有 VLM 配置，初始化 VLM 客户端
|
|||
|
|
if vlm_config and vlm_config.get("enabled") and VLMClient:
|
|||
|
|
try:
|
|||
|
|
self.vlm_client = VLMClient(vlm_config)
|
|||
|
|
logger.info(f"VLM client initialized: provider={vlm_config.get('provider')}, model={vlm_config.get('model')}")
|
|||
|
|
except Exception as e:
|
|||
|
|
logger.warning(f"Failed to initialize VLM client: {e}")
|
|||
|
|
|
|||
|
|
def parse_into_text(self, content: bytes) -> Document:
|
|||
|
|
"""
|
|||
|
|
Parses content using MarkItDown.
|
|||
|
|
Uses self.file_type (inherited from BaseParser) to hint the stream format.
|
|||
|
|
"""
|
|||
|
|
ext = self.file_type
|
|||
|
|
if ext and not ext.startswith('.'):
|
|||
|
|
ext = '.' + ext
|
|||
|
|
|
|||
|
|
# 直接调用 convert，移除 try-catch，让异常由上层 PipelineParser 统一捕获
|
|||
|
|
result = self.markitdown.convert(
|
|||
|
|
io.BytesIO(content),
|
|||
|
|
file_extension=ext,
|
|||
|
|
keep_data_uris=True
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
markdown_content = result.text_content
|
|||
|
|
|
|||
|
|
# 如果有 VLM 客户端，尝试处理图片
|
|||
|
|
if self.vlm_client and markdown_content:
|
|||
|
|
markdown_content = self._process_images_with_vlm(markdown_content)
|
|||
|
|
|
|||
|
|
return Document(content=markdown_content)
|
|||
|
|
|
|||
|
|
def _process_images_with_vlm(self, content: str) -> str:
|
|||
|
|
"""
|
|||
|
|
处理 Markdown 内容中的图片，使用 VLM 分析并替换
|
|||
|
|
"""
|
|||
|
|
# 匹配 data:image 开头的 Base64 图片
|
|||
|
|
pattern = r'!\[([^\]]*)\]\((data:image/([^;]+);base64,([A-Za-z0-9+/=]+))\)'
|
|||
|
|
|
|||
|
|
def replace_image(match):
|
|||
|
|
alt_text = match.group(1)
|
|||
|
|
data_url = match.group(2)
|
|||
|
|
mime_type = match.group(3) or "image/png"
|
|||
|
|
base64_data = match.group(4)
|
|||
|
|
|
|||
|
|
try:
|
|||
|
|
# 解码 Base64 图片
|
|||
|
|
image_bytes = base64.b64decode(base64_data)
|
|||
|
|
|
|||
|
|
# 调用 VLM 分析图片
|
|||
|
|
logger.info(f"Processing image with VLM: {alt_text or 'unnamed'}")
|
|||
|
|
vlm_result = self.vlm_client.analyze_image(image_bytes, mime_type)
|
|||
|
|
|
|||
|
|
if vlm_result.get("success"):
|
|||
|
|
vlm_content = vlm_result.get("content", "")
|
|||
|
|
logger.info(f"VLM processed image successfully, content length: {len(vlm_content)}")
|
|||
|
|
# 替换为 VLM 解析的内容
|
|||
|
|
return f"<!-- Image: {alt_text} -->\n{vlm_content}\n<!-- End Image -->"
|
|||
|
|
else:
|
|||
|
|
logger.warning(f"VLM failed for image: {vlm_result.get('error')}")
|
|||
|
|
return match.group(0) # 保留原图片引用
|
|||
|
|
except Exception as e:
|
|||
|
|
logger.error(f"Error processing image with VLM: {e}")
|
|||
|
|
return match.group(0) # 保留原图片引用
|
|||
|
|
|
|||
|
|
return re.sub(pattern, replace_image, content)
|
|||
|
|
|
|||
|
|
|
|||
|
|
class MarkitdownParser(PipelineParser):
|
|||
|
|
_parser_cls = (StdMarkitdownParser, MarkdownParser)
|