Agent Builder rejects PDF files that were working previously and remain searchable in Copilot Chat

Question

Agent Builder rejects PDF files that were working previously and remain searchable in Copilot Chat

Patrick Hogan 0

Hello,

I am experiencing what appears to be a discrepancy between Microsoft 365 Copilot Chat and Agent Builder.

I have multiple PDF documents that were successfully attached to and used by my agents as recently as yesterday. The agents were functioning correctly, and these same types of PDF files were being accepted without issue.

Today, those same agents are showing errors on the PDF knowledge sources. When I remove and re-upload the files, they are rejected and display a red error icon. The tooltip indicates that the file may not be usable due to a sensitivity label.

The file can still be successfully found and referenced through Copilot Chat search, and Copilot Chat is able to read and use the document. However, Agent Builder rejects the file when I attempt to add it as a knowledge source.

Troubleshooting already performed:

Removed and re-uploaded the PDFs multiple times.
Verified the issue is occurring today despite the same files working yesterday.
Confirmed the PDF is not password protected.
Confirmed Adobe Acrobat shows "Security Method: No Security."
Checked Windows file properties and did not find any obvious security restrictions.
Created a brand-new test PDF. The test PDF is accepted by Agent Builder, while the MD&A PDFs are rejected.
Confirmed the PDFs remain searchable and usable through Copilot Chat.
Observed the issue on another user's screen before seeing the same behavior within my own agents.

My questions are:

What conditions would allow a PDF to be searchable and usable in Copilot Chat but prevent it from being used as an Agent Builder knowledge source?
Are Agent Builder knowledge sources subject to additional sensitivity label, Purview, compliance, or indexing restrictions that do not apply to Copilot Chat search?
Is there a way to determine the specific reason Agent Builder is rejecting the PDF?
Has there been any recent change or service issue related to Agent Builder PDF ingestion, indexing, or sensitivity label enforcement?
Is there a diagnostic log or error detail available beyond the red error icon that would help identify the root cause?

The most confusing aspect is that these files and agents were functioning normally yesterday, and the PDFs remain accessible through Copilot Chat today. The issue appears to be isolated to Agent Builder's ability to ingest or use the files as knowledge sources.

Any guidance would be greatly appreciated.

Ashlesha Pawar (Tata Consultancy Services Limited) 395 Reputation points Microsoft External Staff

2026-09-04T06:02:53.6233333+00:00
Hello Patrick,

Copilot Chat search and Agent Builder knowledge-source ingestion use different processing paths. A file can remain discoverable through the Microsoft 365 search index while Agent Builder performs a new validation before accepting it.

Possible causes include a Purview sensitivity label, encryption without VIEW and EXTRACT rights, a Copilot DLP rule, Double Key Encryption, a digital signature, unsupported PDF properties, or a temporary ingestion failure. Acrobat’s “No Security” does not exclude Purview metadata or tenant policies.

Yes. Agent Builder validates knowledge sources separately. Microsoft Purview can prevent Copilot from processing labeled content even when the item remains discoverable or appears in citations. Signed and password-protected PDFs also have documented limitations.

Please compare an affected PDF with the working test PDF:

Check the sensitivity label in SharePoint/OneDrive or Purview Content Explorer.

Confirm it is not digitally signed.

Verify encryption and the user’s VIEW/EXTRACT rights.

Check for DKE and Copilot DLP policies.

Review Purview Audit and Activity Explorer for label, DLP, or file-processing events.

I could not find a current public known-issue entry documenting an Agent Builder PDF-ingestion regression. However, previously accepted files failing suddenly for multiple users while a clean control PDF succeeds warrants controlled reproduction and escalation.

In Agent Builder, use Help > Get support and collect the agent’s /debug output where available. Include the Agent ID, tenant ID, affected SharePoint/OneDrive file URL and item ID, UTC timestamp, user UPN, tooltip screenshot, browser/network correlation IDs, and Purview policy findings.

Please also test one affected PDF in a new agent with a second user. If it remains unsigned, unencrypted, permitted by DLP, and readable with the required rights but is consistently rejected, this should be escalated to the Agent Builder product/engineering team as a potential ingestion regression.

References:

Agent Builder

Extensibility known issues

Purview DLP for Microsoft 365 Copilot

Sensitivity labels for SharePoint and OneDrive files

Copilot audit logs

1 answer

Your answer

Answer 1

Yes. Agent Builder knowledge sources have file-specific restrictions that can block a PDF even when the file is still usable elsewhere in Microsoft 365 Copilot.

The conditions documented for embedded files in Agent Builder are:

Sensitivity label with Microsoft Purview Double Key Encryption (DKE) enabled
The file can be embedded but isn’t used as knowledge.
Sensitivity label with user-defined permissions enabled
The file uploads, but agent creation fails without an error message.
Sensitivity label where extract rights are disabled for the user
The file uploads, but agent creation fails without an error message.
Sensitivity label from another tenant with encryption enabled
The file can be embedded but isn’t used as knowledge.
Password-protected files
The builder sees an error message next to the uploaded file.

Based on the symptoms described:

the PDF is not password protected,
a newly created test PDF is accepted,
the rejected files show a tooltip mentioning sensitivity labels,

The most likely cause is that those specific PDFs have a sensitivity-label or encryption condition that Agent Builder blocks for embedded knowledge, even though the files remain searchable or readable in Copilot Chat.

Also, Agent Builder has separate handling for embedded file knowledge. Uploaded files are stored in tenant-owned SharePoint Embedded containers, and Copilot uses only the text content of those files for grounding. That means Agent Builder ingestion is not just a general “Copilot Chat can read it” path.

For diagnostics, the documented behavior is limited:

some blocked conditions fail without an error message,
password-protected files show an error next to the file,
the Share screen shows the sensitivity label if labels are set on embedded content.

No richer end-user diagnostic log is documented in the provided material beyond the UI behavior and support workflow.

A practical way to isolate the cause is:

Compare a rejected PDF with the new test PDF that uploads successfully.
Focus on whether the rejected files have a sensitivity label, encryption, user-defined permissions, disabled extract rights, or came from another tenant.
If the rejected PDFs have any of those label/encryption conditions, remove or replace those files for Agent Builder knowledge use.
The documented guidance is to avoid uploading files with DKE, remove files with user-defined permissions, remove files with extract rights disabled, and avoid encrypted labeled files from another tenant.
If the issue must be escalated, use Send feedback inside Microsoft 365 Copilot and include:
- that the issue is related to Agent Builder in Microsoft 365 Copilot
- agent ID
- tenant ID
- environment ID
- session ID
- if relevant to the Agent preview pane or Describe tab, enter /debug in the chat box and include that output

About a recent service change or incident: no service-change explanation is documented in the provided material. Only the known embedded-file restrictions and support path are documented.

One more important limitation: Microsoft Purview Information Barriers aren’t supported on embedded files. Any user who can access the agent can see responses grounded in that embedded content.

References: