Copilot is not reliably reading my documents

Question

Copilot is not reliably reading my documents

Pacetti Christine 0

Hello,

I use Microsoft Copilot, review records, analyze data, and draft reports. I'm hoping someone can help me troubleshoot a document-access problem that has developed over approximately the past 2–3 weeks.

The problem:

When I upload a document, Copilot can usually read and analyze it correctly the first time. However, after just one or two exchanges, it frequently appears to lose access to that same document, even though we're still working with it in the same conversation.

Previously, I could generally ask several questions about different sections of an uploaded document without having to upload it again.

Now, when I ask follow-up questions, Copilot sometimes responds as though it has reviewed the document, but its answers appear to be based on earlier information, assumptions, or inferences rather than the actual source.

What makes this particularly concerning is that Copilot doesn't consistently tell me when it can no longer access the document. Instead, it may generate a plausible, professionally written response that sounds as though it is supported by the source.

For example, I recently asked Copilot what a team had concluded about a student's response to a reading intervention. Rather than locating the team's documented conclusion, Copilot interpreted the available score patterns and generated its own explanation. When I asked where that conclusion appeared in the document, it could not identify the source and eventually acknowledged that it had inferred the answer.

The other problem is that re-uploading the document doesn't always work.

I've found that:

Copilot initially reads an uploaded document successfully.

After one or two exchanges, it may no longer be able to retrieve information from that document.

Re-uploading the exact same file sometimes does not restore access.

Renaming the file and uploading it again has sometimes restored access, even though the document contents have not changed.

I don't know whether this is related to document retrieval, caching, file handling, or something else.

What I'm hoping someone can help me figure out:

Is this a known issue with Copilot's handling of uploaded documents?

Has anything changed recently that could explain why documents that were previously accessible throughout a conversation are now becoming inaccessible after one or two exchanges?

Why might renaming an otherwise identical file allow Copilot to access it when re-uploading the original filename does not?

Is there a setting, workaround, or troubleshooting procedure that would help Copilot maintain access to documents during a conversation?

Is there a reliable way to determine whether Copilot still has access to an uploaded document before relying on its answer?

Why this matters:

My work involves providing specific data from multiple sources to help inform decision. I need to know whether Copilot is reporting information actually documented in a source or generating an interpretation based on partial information.

I understand that AI can make mistakes, and I routinely verify its work. However, this particular issue is difficult to manage because the responses can sound entirely reasonable even when Copilot is no longer working from the actual document.

I would much rather have Copilot tell me it cannot access a document than provide an unsupported answer that appears to be source-based.

Has anyone else experienced this? Is there something I can do to resolve it?

0 comments

1 answer

Your answer

Answer 1

A similar loss of uploaded-file content has been reported, but Microsoft has not published a named known issue specifically confirming that PDFs expire or become inaccessible after follow-up prompts in Microsoft 365 Copilot Chat.^1^

PDF upload and follow-up questions within the same conversation are supported behavior, so losing access after one or two exchanges is not a documented normal limitation.^2^

Use a source-verification prompt before relying on each answer

For work where the distinction between documented facts and interpretation matters, make the grounding requirement explicit:

Use only the attached PDF. Quote the exact text supporting your answer and identify the page or section. If you cannot access the PDF or cannot find an explicit answer, say “Not found in the source.” Do not infer a conclusion from the data.

For every subsequent question, repeat the short form:

Using only [filename.pdf], quote the exact supporting text and identify its page. Do not infer.

Microsoft specifically recommends instructions such as “Use this file only” and asking Copilot to confirm the exact section. If Copilot cannot find a relevant accessible source, it may otherwise provide a general answer.^3^

Treat any of these as a failed retrieval rather than a source-based answer:

No quotation or source entry is provided.
The quotation cannot be found in the PDF.
Copilot identifies a score pattern but cannot identify the documented conclusion.
It changes its answer when asked for the exact page.
It says it is relying on prior conversation context.

Grounding and citations improve verification, but Microsoft still advises checking the cited source before acting on critical details.^3^

Work around the intermittent attachment retrieval

Since re-uploading the unchanged file has already failed, use a different attachment path:

Save the PDF in OneDrive for Business.
In a new Copilot Chat, select Add and manage sources.
Select Attach cloud files, then choose the OneDrive copy.
Include the filename and source-verification instruction in the first prompt.
If retrieval fails later, start a new conversation and attach the same cloud file rather than repeatedly uploading another local copy.

Copilot Chat supports both local uploads and OneDrive cloud-file attachments. Uploaded files themselves are also stored in the user’s OneDrive for Business.^4^^5^

Renaming the file may cause Copilot to treat it as a new attachment, but Microsoft documentation does not establish caching or filename identity as the cause. Therefore, renaming is an observed workaround, not a reliable fix or diagnosis.

If this mainly affects long or complex PDFs, split one affected PDF into smaller page ranges and test the same questions against the relevant section. Microsoft notes that long documents can receive uneven attention and recommends splitting them into smaller documents for more precise responses.^6^

Check for a tenant incident

Ask a Microsoft 365 administrator to check:

Microsoft 365 admin center → Health → Service health.
Review active Copilot or Microsoft 365 incidents and advisories.
Check Issue history for the last 30 days, covering when the behavior began.

Service Health shows tenant-specific incidents and is more useful than the public status page for this purpose.^7^

Report and escalate with a reproducible case

On the first unsupported answer, select Thumbs down and include:

Date and time, including time zone
Web or Microsoft 365 app
Browser/app version and operating system
PDF size and approximate page count
Whether it was uploaded locally or attached from OneDrive
Prompt number where retrieval stopped
The exact prompt and response
Whether renaming changed the result
A sanitized screenshot showing the missing or incorrect source

Copilot’s feedback pane is designed to collect details about incorrect responses; an administrator can review end-user feedback under Microsoft 365 admin center → Health → Product feedback if organizational feedback is enabled.^8^

If Service Health has no matching advisory and the problem reproduces with multiple PDFs or users, the administrator should open a Microsoft 365 support request and provide the same evidence. Do not attach records containing sensitive student information unless the organization’s approved support and data-handling process permits it.

References