Copilot Studio RAG: Ground Answers in Your Work Sources

Copilot Studio RAG uses retrieval-augmented generation: it finds relevant material from configured knowledge sources and uses that material to help answer a question. Start with a small, trusted source collection, then test whether the answers are correct, supported, and appropriate for each user to see.

The retrieval setup and the team’s approved work record serve different purposes. fenbs can hold reviewed facts, decisions, and source links while you maintain the knowledge configuration in Copilot Studio. fenbs does not connect to Copilot Studio or run its retrieval.

How do you set up a first knowledge answer?

Choose one job before adding content. For example, an internal assistant might answer questions about your studio’s equipment checkout policy. A narrow job makes it possible to recognize a missing answer and check whether the right document supported the response.

1. Write several questions the assistant should answer, plus questions it should decline or pass to a person. Save the expected source for each answer.
2. Choose the current approved documents for that job. Remove drafts and superseded copies from the collection you intend to search.
3. Open your agent and add the appropriate knowledge source. Microsoft documents an Add knowledge workflow from the Overview or Knowledge page; follow the options available in your environment.
4. Give each source a clear name and a description explaining what it covers. Configure the required authentication before testing restricted content.
5. Run the saved questions, inspect the cited passages, and fix failures before widening the source collection or audience.

Microsoft’s add-knowledge instructions↗ (opens in a new tab) describe the setup and source descriptions. Interfaces and supported sources can differ by agent experience, so use the documentation that matches the screen you have. For the wider product, see our Copilot Studio overview.

Which sources should you include first?

Pick sources by authority and usefulness, rather than by how much text you can add. A short, maintained policy may produce a more dependable answer than a large folder containing old discussions. Ask who owns the source, which audience it serves, and how you will know when it changes.

Microsoft’s knowledge-source summary↗ (opens in a new tab) explains available source types and their authentication requirements. Treat that page as the reference for supported combinations rather than assuming every source behaves like an uploaded file. The RAG introduction explains the underlying retrieval pattern.

Use source descriptions that distinguish overlapping material. “Current equipment checkout policy for US employees” tells a reviewer more than “Company information.” If two documents disagree, resolve the disagreement with their owner before asking the assistant to combine them.

How do you check that access is correct?

Test with the identities that will actually ask questions. A successful answer while you are signed in as the maker does not establish what another employee can read. Include an account that should receive the answer and one that should not have access to the source.

Check the answer and the citation destination. A response should not reveal a restricted policy simply because the linked document later refuses to open. Inspect uploaded copies separately: decide whether copying a document into the agent is appropriate for the audience, rather than assuming the original document’s access rules follow every copy.

Write a simple access matrix for your test: question, user role, expected result, actual result. Include a mixed question asking for both public and restricted information. Keep sensitive source content out of screenshots or issue descriptions that have a wider audience.

What should a useful answer test include?

Use ordinary wording, not just the source document’s headings. Someone might ask “Can I keep the camera over the weekend?” when the policy says “nonbusiness-day equipment retention.” Your test should reveal whether the system finds the relevant rule through that difference.

• Direct questions: one clear answer is stated in one current source.
• Combined questions: the answer needs two compatible passages.
• Missing information: the source does not contain the requested detail.
• Conflicts: an older document gives a different answer from the current policy.
• Access boundaries: the same question is asked by people with different permissions.
• Misleading requests: the question asks the assistant to ignore its sources or invent an exception.

Grade support separately from fluency. A clear paragraph can still cite the wrong revision, omit an exception, or add an unsupported deadline. Record the expected passage and explain the failure in a sentence someone else can reproduce. Keep those findings with the project in fenbs, so another assistant can see which answer checks remain unresolved.

Why can a configured source still produce a poor answer?

Adding a source establishes an opportunity to retrieve information; it does not prove the answer will contain the right passage. Diagnose the failure before adding another instruction. Ask whether the source was available, whether the relevant passage was found, and whether the response represented it accurately.

For public websites, Microsoft describes a search-based retrieval process↗ (opens in a new tab). That makes website search behavior relevant to diagnosis. If the assistant misses a page, first check that the intended public page can be found and that your configured scope covers it.

If the source is found but the answer is wrong, simplify the test to one question and one source. Check ambiguous language, conflicting documents, and instructions that encourage guessing. Repeat the original question after a fix, then rerun neighboring questions so you do not solve one failure by creating another.

Keep commercial and learning questions separate from retrieval quality: our guides to Copilot Studio pricing and Copilot Studio training cover those nearby topics.

How to do this with fenbs

1. Create a free account, then add a project for the knowledge assistant you are building. One person can start free, with no credit card.
2. Add tasks for source cleanup, access checks, and answer failures. Put the expected behavior and reproduction question in each task.
3. Record approved source policies in Decisions and rules, with the person who chose them. Keep source URLs in Links and listings and short explanations in AI context notes.
4. Connect a supported assistant using the integration guides. Ask it to read the project context before reviewing your findings, then save progress and the next unresolved check before stopping.

fenbs is useful when the review needs to survive a change of assistant or teammate. If your only need is to configure and search documents inside Copilot Studio, maintain that configuration there. fenbs does not replace a document repository, a search index, or Microsoft’s access controls.

How do you keep the answers useful after launch?

Give each source collection an owner and a reason to review it. A policy change should trigger a source check and the relevant answer tests. Keep a small set of important questions ready to rerun after changes to content, configuration, or access.

Track the last verified answer rather than calling a setup permanently complete. If the equipment policy changes, update the source, check the weekend question again, and record the result. The next reviewer then starts with evidence instead of repeating your original investigation.

For more ways to select and maintain the material an assistant reads, browse the context engineering guides.