Vector Store only stores last page of scanned PDF

[Justus Kilian Wolff (justuswolff-audit)]

While testing thread vector stores I observed that the thread info shows that the file search tool always returned just one result. Since we are using the standard setup for our agents I checked the search index to see if I can query more results. It turns out that the index only holds one document (which is the last page of the scanned pdf)

I'd assume that something with the pdf embedding code is broken on Azures side? Should I open an issue somewhere?

Additional info: the document is scanned, a pdf and contains 19 pages. The one item (document in AI Search terms) that is added to the search index holds exactly the context of the last page. I added the document to the file search on thread level via the AI Foundry playground.

[Justus Kilian Wolff (justuswolff-audit)]

Seems to be an issue with the Azure skillset in the document processing pipeline.