Read pdf's? Currently if microsoft_docs_fetch_article model hits maxtokens

[christianarg]

The microsoft_docs_search sometimes returns pdf's but when trying to fetch with the 'microsoft_docs_fetch_article' the model hits maxtokens (tested with gpt-5.1, 400k context window)

The pdf was not very large it, it's this one:
https://www.microsoft.com/content/dam/microsoft/final/en-us/microsoft-brand/documents/Modern-Work-Plan-Comparison-SMB-1-16-2024.pdf

It's probably complicated but I just ask as a "wish" if it would be possible to add a tool that reads pdf's, as markdown maybe? Or in a format that the model could receive?

For now I had to instruct my agent NOT to use microsoft_docs_fetch_article for pdf's, as a workarround

[pdebruin]

Thanks for reporting this experience. We are investigating how to improve. @partychen

[partychen]

@christianarg We’ve resolved this issue.
Specifically, we made the following change:
Exclude PDF inputs when fetching data – The microsoft_docs_fetch step will no longer attempt to process PDF files.
This prevents the model from trying to ingest large PDF content and avoids hitting the max token limit.

[christianarg]

@partychen great, before closing the issue, it's realistic for you to be able to add a "fetch_pdf" tool that reads a pdf as markdown or that is out of the question?

[Eimasogi]

We can look into “fetch_pdf” (or enable retrieval of pdf as markdown). There are a few considerations here namely token sizes, the corpus size of pdf content across Learn etc.

@christianarg can you expand a bit more on your use case? Are you simply looking to use the pdf content as additional context for the agent (assuming the content page itself doesn't have this data already)?

[christianarg]

@Eimasogi yes essentially use the pdf content as additional context. For example this concrete pdf https://www.microsoft.com/content/dam/microsoft/final/en-us/microsoft-brand/documents/Modern-Work-Plan-Comparison-SMB-1-16-2024.pdf that was actually returned by the "microsoft_docs_search" tool.
Microsoft licensing is kinda hard and the learn mcp tool is a great way for agents to be able to get licensing information. Some documents like this one are pdf.
Ideally if would be nice to be able to "feed" the agent these documents.
Not even sure if implementing this tool is the right call, just a thought/wish. As you say token sizes can be a problem, this particular pdf doesn't seem to be a problem if transformed to markdown but there may be some large pdf that could run out of context many models. Maybe take an opinionated decision on how much context can this hypothetical "fetch_pdf" tool and over a threshold the tool simply returns "pdf too large"?