Image export

[pizang]

It would be good to have the option to export the images to media directory and add a proper image link in a markdown code.

When we use RAG with DOCX manuals we currently use pandoc with image export. In some cases it does not make any sense to describe pictures as they may be some schemas or diagrams. In that case we provide a link to LLM and then it is displayed in the final answer.

[CarlZhang12]

this is my tmp method to get docx images outputs, maybe good for you:

from markitdown import MarkItDown,StreamInfo
import mammoth
from pathlib import Path
import tempfile

docx_file = "/path/to/docx"
img_dir = Path('image')
img_dir.mkdir(exist_ok=True)

def convert_image(image):
    with image.open() as image_bytes:
        img_name = Path(image_bytes.name).name
        img_path = img_dir.joinpath(img_name)
        with open(img_path,"wb") as f:
            f.write(image_bytes.read())
    return {
        "src":f"{img_path}"
    }
html_content = mammoth.convert_to_html(docx_file, convert_image=mammoth.images.img_element(convert_image))
md = MarkItDown()

with tempfile.NamedTemporaryFile() as tmp:    
    tmp.write(html_content.value.encode('utf-8'))
    tmp.flush()
    res = md.convert(tmp.name,stream_info=StreamInfo(mimetype='text/html'))
    Path(docx_file.replace('docx','md')).write_text(res.text_content, encoding='utf-8')

[tqye2000]

Many thanks for the example. However, I think there is a minor bug though. It should be:

with tempfile.NamedTemporaryFile() as tmp:    
    tmp.write(html_content.value.encode('utf-8'))
    tmp.flush()
    
res = md.convert(tmp.name,stream_info=StreamInfo(mimetype='text/html'))
Path(docx_file.replace('docx','md')).write_text(res.text_content, encoding='utf-8')

Namely, the tmp file needs to be closed before the md process read it.

[CarlZhang12]

Many thanks for the example. However, I think there is a minor bug though. It should be:

with tempfile.NamedTemporaryFile() as tmp:    
    tmp.write(html_content.value.encode('utf-8'))
    tmp.flush()
    
res = md.convert(tmp.name,stream_info=StreamInfo(mimetype='text/html'))
Path(docx_file.replace('docx','md')).write_text(res.text_content, encoding='utf-8')

Namely, the tmp file needs to be closed before the md process read it.

I think postprocess extraction works better: markitdown with --keep-data-uris to get md output first and then use pattern match string base64_pattern = r'data:image/([a-zA-Z0-9]+);base64,([a-zA-Z0-9+/]+={0,2})' to get base64 img content and replace base64 str with local reference
this would keep original data pipeline unchanged and have better control for local image filename generation.