Practice Exams:

Microsoft AB-100: Knowledge Sources in Copilot Studio

AI & Machine Learning

Knowledge sources determine what a Copilot Studio agent can retrieve when it needs factual context beyond its instructions. Current Copilot Studio supports several knowledge types, including public websites, uploaded documents, SharePoint, and Dataverse, with different authentication and scale characteristics. Choosing a source is therefore a data-architecture decision rather than a simple content-upload step.

On this page
  1. Use SharePoint for governed work content
  2. Use Dataverse for structured business knowledge
  3. Use uploaded documents deliberately
  4. Use public websites for public evidence
  5. Choose sources by freshness
  6. Preserve permissions before retrieval
  7. Preserve provenance
  8. Review source authority
  9. Keep knowledge lifecycle operational

The most useful knowledge source is the one that is authoritative, permissioned correctly, fresh enough for the task, and structured so retrieval can find the right evidence. More sources do not automatically produce better answers. A design review should ask who owns the content, how permissions reach the user making the request, whether the material is still authoritative, and how changes or deletions will reach the index. These answers determine whether a source can safely ground an agent’s output, independent of how convenient it is to connect.

Use SharePoint for governed work content

Copilot Studio SharePoint knowledge sources can search a configured site or folder and its subpaths. With supported user authentication, retrieval is performed on behalf of the signed-in user and respects their SharePoint permissions; this behavior should be verified for the actual channel and agent settings. Microsoft 365 data protection should be assessed before expanding grounding because Copilot makes authorized content easier to discover, including material that an organization may have overshared.

Use Dataverse for structured business knowledge

Dataverse can provide table-based knowledge tied to the same governed environment as Power Platform applications. This is useful when the answer depends on structured records rather than document prose. Power Platform integration becomes stronger when transactional data and conversational knowledge share a clear system of record.

Use uploaded documents deliberately

Uploaded documents can be useful for bounded, curated knowledge sets. They also create a maintenance responsibility: someone needs to know which version is current, when the document should be replaced, and whether the upload still reflects the authoritative source. Do not create a permanent knowledge archive from documents that really belong in a governed repository.

Source onboarding should include a quality review. Check whether documents are current, searchable, consistently titled, permissioned correctly, and free of obvious duplicates before adding them to an agent. Poor source quality is easier to fix before users depend on the answers.

Use public websites for public evidence

Copilot Studio can constrain public-web knowledge to specified websites. This is useful for product documentation, public policy, or other information that should be grounded outside the tenant. Public content should remain distinguishable from internal enterprise evidence so users understand the source of the answer.

Choose sources by freshness

A static uploaded file may be appropriate for a stable handbook but weak for rapidly changing inventory or service status. Knowledge grounding should match source freshness to the business decision the agent is supporting. For high-churn data, a live connector or Dataverse lookup can be stronger than relying on a periodically refreshed document.

Freshness monitoring should identify failed refreshes or stale content. A source can remain technically connected while no longer reflecting current policy. For high-value agents, add a simple operational check such as last update time, owner review date, or source health indicator.

Preserve permissions before retrieval

Security should be enforced at the source or retrieval layer before content enters the prompt. Identity and data should remain aligned across SharePoint, Dataverse, connectors, and agent authentication. The model should not receive sensitive content and then be asked through instructions not to reveal it.

Permission testing should include users with different roles. An administrator or maker can accidentally conclude that grounding works correctly because their own account can access everything. Test with representative end users to confirm that retrieval is trimmed to the right content for each role.

Knowledge design should also minimize unnecessary scope. An agent built for one department rarely needs an entire tenant’s content. Smaller source boundaries can improve relevance, simplify security review, and make failures easier to diagnose. For content with very strict access or rapidly changing state, use a live tool or connector instead of copying the data into a knowledge source. Retrieval architecture should choose between indexed knowledge and live business queries based on authority, latency, freshness, and permission requirements.

Preserve provenance

Store and expose enough source context for users and operators to know where an answer came from. Source title, location, version, owner, and effective date can matter as much as semantic relevance. Enterprise grounding is trustworthy when users can inspect the underlying evidence.

Documents should retain recognizable provenance after ingestion. Users and support teams should be able to trace a response back to the original source, not merely an internal document identifier. Provenance is essential when someone challenges an answer or needs the latest approved version.

Review source authority

Large organizations often have several documents that answer the same question differently. Decide which source is authoritative and retire or demote stale duplicates. Retrieval cannot reliably infer business authority if the source metadata and governance do not express it.

Knowledge-source selection should also consider scale. A handful of curated documents can work well for a narrow internal agent, while a large enterprise corpus needs stronger information architecture, ownership, permission hygiene, and lifecycle. The platform can retrieve from many sources, but governance determines whether those sources remain trustworthy over time.

Source descriptions matter because makers and operators need to understand why each repository is attached. Record the business purpose, owner, intended audience, freshness expectation, and sensitivity. That metadata turns “SharePoint site added” into an understandable design decision.

When sources overlap, define precedence. A formal policy site may need to outrank a project wiki, even if both contain similar words. Use governance, content cleanup, metadata, and source design to reduce contradictions before asking the model to reconcile them.

When several sources answer the same topic, compare their update process and governance. The technically easiest source to connect is not always the right source to trust. A well-managed SharePoint library may be better than an uploaded PDF that nobody owns, and a live Dataverse record may be better than a nightly export.

Keep knowledge lifecycle operational

Knowledge sources need owners, refresh rules, permission review, and retirement just like other production assets. Purview governance can reinforce retention, sensitivity, audit, and information-protection expectations around enterprise content. For current Copilot Studio programs, the durable approach is to choose each source for authority, freshness, permissions, and retrieval quality; keep provenance visible; and remove obsolete knowledge before it becomes a confident wrong answer.

Maintain knowledge quality and scope

As the agent evolves, remove unused knowledge sources. More content can reduce precision, introduce stale evidence, and increase governance burden. A healthy knowledge layer is curated: sources are added for a reason, measured for usefulness, and retired when they stop supporting the agent’s job.

Measure knowledge performance through unanswered questions, low-confidence cases, citation quality, and repeated user corrections. Those signals can reveal missing content or weak source boundaries even when the agent has many documents attached. When a knowledge source is retired, remove its connection and verify that the agent no longer cites or relies on cached versions. Content lifecycle should include deletion as well as addition.

Identify each source of record

Knowledge architecture should also document whether a source is primary, supplementary, or public reference material. That distinction helps reviewers understand how much authority an answer derived from that source should carry. Knowledge-source boundaries should be explained to users where they affect expectations. An HR policy agent may answer only from approved HR repositories, while a broader research agent may use public sites as well. Clear scope reduces false assumptions about what the agent “knows.”

Use source-specific evaluation cases. Ask questions that should be answered by SharePoint, Dataverse, uploaded content, and public websites separately so failures can be traced to the responsible source type rather than one generic knowledge score.

Knowledge maintenance should be part of the agent owner’s calendar. Review source health, permissions, stale content, and unanswered questions on a recurring cadence so the knowledge layer remains aligned with the business process. Source cleanup should be measurable. Track duplicated documents, stale versions, missing owners, failed refreshes, and unanswered high-value questions. These indicators help content teams improve the corpus systematically rather than reacting only when a user reports a bad answer.

Knowledge owners should also review search language. Users may ask with acronyms, product nicknames, customer terminology, or outdated names that do not appear in official documents. Adding useful synonyms, clearer titles, or better structure to the source can improve retrieval more reliably than adding prompt instructions that guess what the user meant.

Knowledge-source changes should pass through lightweight change control for important agents. Adding a new repository can change answer behavior as much as editing instructions because the retriever now has different evidence to rank. Record the change, run representative queries, and keep the previous source configuration until quality is understood. Keep that source-change history available to support and governance teams.

Related Posts

• Mastering the Basics: Your Guide to Microsoft Azure AI

• Microsoft AI-300: How to Roll Back a Model Without Rolling Back the App

• Microsoft AI-103: Managed Identities for AI Apps

• Microsoft AI-901: Computer Vision, Language, and Generative AI

• Microsoft AI-103: Identity Boundaries for Foundry Agents

• Microsoft AI-103: Controlling the Cost of Azure AI Apps

• Microsoft AI-103: Reducing Azure AI Response Latency

• Microsoft AI-103: Reproducible Azure ML Pipelines

• Microsoft AB-100: Copilot Adoption Without AI Sprawl

• Microsoft AB-100: GitHub Copilot Repository Instructions