Copilot Studio: The Real Limits of SharePoint Knowledge Sources

A Copilot Studio agent points to a SharePoint library with 400 documents and yet reliably answers only a fraction of questions. The most common suspicion in such projects is that Copilot Studio only knows the first ten documents. This number is not documented anywhere in Microsoft documentation. The real operational limit of knowledge sources lies at a completely different point, and it is significantly narrower than the ten-document legend suggests.

Is there a ten-document limit in Copilot Studio?

No. Microsoft does not document a limit of ten documents per knowledge source for Copilot Studio. The hard limit that determines answer quality in practice is a retrieval limit: For a SharePoint knowledge source, only the three best search results are used to answer each question. Microsoft states verbatim: "When Copilot Studio searches SharePoint, only the top three search results are used to summarize and generate a response." (Source: Microsoft Learn, Generative answers pointing to SharePoint sources don't return results, accessed August 3, 2026).

The number ten likely comes from neighboring values. For declarative agents in Microsoft 365 Copilot, there is a limit of 20 files: Copilot searches the complete content of up to 20 files, beyond that only the 20 most relevant ones (Source: Microsoft Learn, Optimize content retrieval). Anyone mixing such values ends up with a rough estimate that doesn't actually exist.

Why does answer quality drop with a large library?

Answer quality drops because a search engine sits between agent and document, with its hit list cut off hard. Copilot Studio retrieves SharePoint content via Microsoft Graph search and passes only the top results to the language model. If the agent points to a library with hundreds of similarly titled documents, only the ranking of SharePoint search determines which three of them even have a chance at the answer.

This explains the typical failure pattern: The agent confidently answers questions about prominent documents and claims it can't find anything for everything else. This is not an indexing error, but the intended behavior.

Adding to the problem, when Copilot Studio exceeds quantity limits, it silently truncates. Microsoft states this unusually clearly: Copilot Studio indexes up to the maximum number of files and folders, does not process the rest, and does not indicate which items were processed and which were not (Source: Microsoft Learn, Unstructured data as a knowledge source). So there is no error message to pin the problem on.

What limits really apply to SharePoint knowledge sources?

The authoritative values are in the Copilot Studio quotas and limits, as of August 3, 2026:

• Three search results per answer: Only the top three SharePoint results are summarized. This is the actual operational limit.
• 25 SharePoint URLs per agent: With generative orchestration, a maximum of 25 SharePoint site URLs are allowed. In classic mode, it is four URLs per response node.
• 1,000 files, 50 folders, 10 folder levels: This quantity limit applies per uploaded SharePoint knowledge source, with a maximum of 512 MB per file.
• 500 knowledge objects, 5 sources simultaneously: An agent can manage 500 knowledge objects total, but can only use five different source types simultaneously.
• 7 MB without, 200 MB with Copilot license: Without a Microsoft 365 Copilot license in the same tenant, generative answers process only SharePoint files under 7 MB. With a license and enabled Tenant Graph Grounding, up to 200 MB.
• 4 to 6 hours synchronization: Newly uploaded documents are not immediately available to the agent.
• No count and metadata questions: Questions like "How many files are in this folder?" or "List all files from knowledge source X" are explicitly not supported.

What to do instead of simply pointing to the entire library?

The most effective lever is to reduce the amount of accessible documents instead of tweaking agent instructions. Four approaches have proven effective, in this order.

• Curated subset instead of full library: Create a separate library containing only documents that can actually provide answers. Microsoft explicitly recommends for declarative agents to name relevant individual files instead of a folder. The same principle applies to Copilot Studio.
• Cut documents smaller and more thematically focused: Microsoft suggests around 36,000 characters per file as an upper limit for good retrieval quality, roughly 15 to 20 pages. A 120-page manual split into twelve thematic files will be found much more reliably than in one piece.
• Multiple agents instead of one omniscient agent: One agent for HR topics, one for IT policies, each with its own small knowledge source. This bypasses the three-result limit more effectively than any attempt to improve ranking in a mixed library.
• Custom index as a workaround: If that's not enough, the way forward is through a custom retrieval layer, such as Azure AI Search or a vector database. This lets you decide how many passages are retrieved and how they are weighted.

In our Copilot Studio projects, we therefore never start with the question of which library to connect, but with which 20 to 40 documents cover real user questions. This subset goes into a separate library, and only that becomes the knowledge source. More on our page about Copilot Studio agents.

How do you tell if the knowledge source is the problem?

The tell is an agent that fails on precise questions about a specific document but answers general questions adequately. Before making changes to instructions, check these four points:

• Is the knowledge source set to "Ready"? The status jumps briefly to "Ready" after adding and then back to "In progress". Only the second switch to "Ready" counts.
• Is the file too large? Without a Microsoft 365 Copilot license, files over 7 MB are excluded even though Graph Search returns them.
• Does the file have a sensitivity label? Files labeled "Confidential" or "Highly confidential" plus password-protected files are shown as ready but provide no answers.
• Does the asking user have read permissions? Without permissions, there are no results and no error message. To the user it looks like the document doesn't exist.

The last point has a flip side: Copilot exposes accumulated permission errors instead of creating them. We wrote about this in Copilot rollout and the oversharing risk. And if you're still weighing whether automation belongs in Copilot Studio at all, the distinction in Power Automate or Copilot Studio agent flows helps.

Frequently asked questions

Does a Copilot Studio agent really only know ten documents?

No. A ten-document limit is not documented in Microsoft documentation. Up to 1,000 files per knowledge source are possible, but only the three best SharePoint search results are evaluated for the answer itself. This retrieval limit has the practical effect of a much tighter document limit.

How many SharePoint sites can an agent use as a knowledge source?

With generative orchestration, a maximum of 25 SharePoint site URLs per agent is allowed. In classic mode, it is four URLs per generative response node. More URLs don't improve answer quality anyway because only the top three results are still used per question.

Why doesn't the agent find a newly uploaded document?

Synchronization of a SharePoint knowledge source runs every four to six hours. Additionally, the SharePoint search must have indexed the document. A freshly uploaded document is therefore regularly not reliably available until the next day.

What happens if the library contains more than 1,000 files?

Copilot Studio indexes up to the maximum number and does not process the remaining files. There is no error message and no indication of which files were processed. This is why the knowledge source should be deliberately curated instead of pointing to a growing library.