Add Unstructured Knowledge Sources to a Copilot Studio Agent

Microsoft Copilot Studio can ground an agent in files from OneDrive or SharePoint and in published content from supported knowledge-base systems. Choosing the correct source type matters because synchronization, storage, authentication, citations, and user access differ from a simple file upload.

The procedures below apply to agents using Copilot Studio's standard harness. Available connectors and licensing can vary by tenant and region.

Choose the correct knowledge-source method

Method	Updates	Access behavior	Storage/query behavior
Upload a local file	Static; replace it manually after changes	Anyone who can use the agent can potentially receive answers from the uploaded content	File is stored and indexed in Dataverse
OneDrive files or folders	Synchronizes changes	User credentials are checked against source permissions	Content is processed and indexed in Dataverse and consumes storage
SharePoint files or folders through Upload files	Synchronizes changes	User credentials are checked against source permissions	Content is processed and indexed in Dataverse; supported PDFs can provide page-level citations
Unstructured knowledge base	Uses published collection content	Users sign in with their own credentials after publication	Added at collection level through a supported connector
Real-time Power Platform connector	Queries current source data at runtime	Runtime calls use each user's authentication token	Only metadata is indexed; source data is not copied into Copilot Studio

Do not upload a confidential file merely to avoid configuring authentication. Static upload is not a substitute for source-level permission trimming.

Add OneDrive files or folders

Copilot Studio's OneDrive path supports Word (.doc, .docx), PowerPoint (.ppt, .pptx), PDF (.pdf), and Excel (.xls, .xlsx) files, up to 512 MB per file. Password-protected files and files protected by sensitivity labels cannot be indexed; they may show a ready status without producing answers.

Open the agent in Copilot Studio.
Select Add knowledge from Overview or Knowledge, or from the properties of a generative answers node.
Select OneDrive.
Select Browse items and choose up to five files, folders, or a combination. You can instead enter the URL of an item shared with you.
Select Confirm selection.
Give each source a unique, descriptive name and explain its scope in the description. Generative orchestration uses the description when deciding which source to query.
Select Add to agent.
Wait for the status on the Knowledge page to become Ready.
Ask representative questions in the Test panel. The first query prompts you to connect so Copilot Studio can verify your source permissions.

An individual file appears as one knowledge item. A selected folder appears as one item and includes supported files in its subfolders. Renaming the item in Copilot Studio does not rename the source in OneDrive.

Add SharePoint files or folders

The Upload files > SharePoint path has the same core file formats and 512 MB maximum. MP4 video content is not supported. The picker can select individual files and folders, but it does not currently select an entire document library through this path.

Open Add knowledge and select SharePoint.
Select Browse items, then choose up to five files or folders, or enter a SharePoint item URL you can access.
Confirm the selection, then add a specific name and description.
Select Add to agent and wait for Ready.
Connect in the Test panel and test as users with different permission levels.

Copilot Studio also exposes another SharePoint knowledge option based on the SharePoint connector. Use that route when you need broader SharePoint integration, custom authentication, or advanced query options. Confirm which SharePoint tile you selected before interpreting its limits.

Page-level PDF citations

A PDF added through Upload files > SharePoint can link a citation to the page containing the retrieved passage. The PDF must be approximately 4 KB or larger and meet Microsoft's other indexing requirements. Test the link after publishing; page-level behavior is specific to this ingestion path and should not be assumed for every uploaded file.

Add Confluence, Salesforce, ServiceNow, or Zendesk knowledge

Copilot Studio supports these unstructured knowledge-base sources:

Confluence Cloud
Salesforce
ServiceNow
Zendesk

The source is added at the collection level—such as a set of pages or a knowledge base—not as hand-picked individual articles. Only published content is used; drafts and archived articles are excluded.

Select Add knowledge, then choose the source. If it is not under Featured, check Advanced.
Open Your connections, select Create new connection, and sign in to the external service.
Select Allow, then Next.
Choose the relevant knowledge-base collection from the available sources.
Select Preview and inspect sample records to confirm the correct collection and content.
Select Next, add a clear name and routing description, then select Add to agent.
Wait for Ready, connect in the Test panel, and run permission and retrieval tests.

The maker's setup connection is used to configure the source. It is not reused as a universal identity for published users; each user must authenticate, and Copilot Studio checks that user's source permissions before returning content.

Use real-time Power Platform knowledge connectors

Power Platform real-time knowledge connectors are a preview feature. Microsoft currently lists connectors including Salesforce, ServiceNow, Azure SQL, Azure AI Search, SharePoint, Dataverse, Dynamics 365, Snowflake, Databricks, Zendesk, Confluence Cloud, Oracle Database, SAP OData, and Google Sheets.

Unlike file ingestion, these connectors index table and column metadata, then query the target system at runtime. Data access remains governed by the source and the user's authentication token.

In Add knowledge, select the real-time connector; check Advanced if it is not featured.
Select Add, sign in, and choose the target location.
Select the tables or sources the agent may query.
Add a precise name and description. Add supported synonyms or glossary definitions where available.
Select Add to agent and wait for metadata indexing to finish.
Test with least-privileged accounts before production use.

Because the feature is in preview, do not assume production support, availability, or behavior will remain unchanged. Check your organization's data policies before enabling a new connector.

Test the source before publishing

Ask questions whose answers are present, absent, outdated, and restricted.
Confirm citations open the correct file, article, or PDF page.
Test with a user who has access and another who does not.
Verify that updates synchronize and that deleted or unpublished content stops appearing.
Check source descriptions if the agent retrieves from the wrong knowledge item.
Use supported text alternatives for important diagrams and images; annotated images in PDFs are more retrievable than unexplained graphics.
Measure Dataverse storage consumption for file-based sources.

Common problems

Ready but no answers: check for encryption, a sensitivity label, unsupported content, insufficient permissions, or a weak source description.
Old content appears: confirm that the source is synchronized rather than a static upload and allow time for reindexing.
Excel answers are poor: tabular analysis can be unreliable when a workbook is treated as unstructured text. Use Dataverse, a database connector, or another structured source for analytical questions.
Users see too much: remove static uploads containing restricted material and use an authenticated, permission-aware source.
SharePoint option behaves differently: verify whether you selected Upload files > SharePoint or the full SharePoint connector.

Microsoft updates Copilot Studio limits frequently. Before a large rollout, compare your configuration with the current unstructured-data documentation and current quotas and limits.