Can immage description be suppressed for pptx files?

[Dynteq]

I just found and installed markitdown, and it works very well.
Even more than I expected. I used the command-line to convert a .pptx so I could to have a discussion with my favorite llm about it. To my surprise it also described the pictures in there which is not mentioned on the first page anywhere.
This was thankfully not a sensitive document but for others I would hate to find out after the fact that they (or the images) were send over the internet to some unnamed AI to be analyzed.

What I did after installation:
markitdown example.pptx -o example.md

Result:
an excellent .md of all the text in the document, including a description of the enclosed image

Question 1:
Can the (online) transcription of images be suppressed? (I didn't find an argument to do so)

Question2:
Does marktidown access the internet for other things? In that case it should be clearly stated as this would make it unsafe for sensitive data that should stay local.

Kind regards,
Martijn

[pauldev-hub]

Hey @Dynteq ,

Issue #1531 — Suppressing image transcription in PPTX

Q1 — Can online image transcription be suppressed?
Yes, completely. Image description via LLM only activates when you explicitly provide llm_client and llm_model during initialization. (GitHub)

Simply run markitdown without those parameters:

CLI — no LLM = no image description, no internet call

markitdown example.pptx -o example.md

Python — same result

from markitdown import MarkItDown
md = MarkItDown() # no llm_client = images skipped entirely
result = md.convert("example.pptx").

Q2 — Does markitdown access the internet for other things?

Audio transcription uses speech_recognition which leverages Google's API — so audio conversion does send data online. (DEV Community) Text extraction from PPTX/DOCX/PDF is fully local. The internet is only used when you explicitly configure LLM integration or convert audio files.

For sensitive documents: Use markitdown with no llm_client, no audio conversion, and you're fully offline.

👍 If this helped you, please mark it as the answer — it helps others in the community who run into the same issue find the solution faster!

[Dynteq]

Hi @pauldev-hub,
Thank you for your answer, that is reassuring. Just to be sure: What should my prompt look like to assure no internet calls?
Because my original (command line) prompt:
markitdown example.pptx -o example.md
resulted in the images being described, I did not specify llm_client or llm_model.

So it looks like from the command line some default llm_client and llm_model are used.
I had a peek in main.py but could not spot any argument to specify or suppress these.
Kind regards,
Martijn

[txhno]

Short answer:

• LLM image description is not on by default. In the PPTX converter it only runs if llm_client and llm_model are configured.
• MarkItDown can still include the image’s embedded alt text from the slide deck, so that may be what you saw.
• Other internet-using paths do exist, for example audio transcription and Azure Document Intelligence if you use those features.

So for sensitive PPTX files, plain local conversion should stay local unless you explicitly enable an online feature.