Microsoft Copilot Studio Knowledge Base: Turn One VPN Guide into Answers Your Agent Can Trust
An employee types a message to the PAW IT Helpdesk Assistant: "My VPN keeps disconnecting every 20 minutes. I already restarted my laptop. What should I do?"
The agent has the PAW Labs VPN Troubleshooting Guide attached, and the knowledge source shows as ready. The answer can still go wrong in several ways. It might open with "restart your laptop," which the employee already did. It might quote a section from an older document. It might invent a fix the guide never approved. Or it might answer confidently a question the guide never covered.
Our first tutorial, Microsoft Copilot Studio Tutorial: Build an AI Agent That Troubleshoots Problems and Creates Support Tickets, created the PAW IT Helpdesk Assistant, wrote the PAW Labs VPN Troubleshooting Guide, and attached it as knowledge. PAW Labs is our internal beta sample project, and this chapter continues that same build. We reopened the knowledge setup and asked a harder question: how do we turn one connected document into a knowledge system the agent can search reliably, and how do we prove it works? The guide below is a beginner-friendly, step-by-step process, from deciding what the agent should know, through adding and describing the source inside the agent, to testing, permissions, and maintenance. A full conversation that moves from knowledge to tools to a finished ticket closes the walkthrough.
See how to turn a connected document into a reliable knowledge source in Microsoft Copilot Studio, from configuring the source to testing whether your agent retrieves the right answers.
Quick Decision Guide: Where Should the Answer Come From?
Before changing any configuration, we sort every employee question into one of four paths. This one habit prevents most architecture mistakes.
Yes → Knowledge source (example: authentication error steps)
Yes → Tool or data lookup (example: status of PAW-1042)
Yes → Action with confirmation (example: create a support ticket)
Say so plainly → Point to human IT support
| Employee question | Right capability | Reason |
|---|---|---|
| How do I fix a VPN authentication error? | Knowledge | The approved guide holds the procedure |
| What are the supported VPN setup requirements? | Knowledge | Documented guidance |
| What is the current status of PAW-1042? | Tool or data lookup | A live record that changes |
| Can you create a support ticket? | Tool or action | The agent must create a record |
| What if the guide does not cover my error? | Knowledge boundary, then human support | The agent must not invent a procedure |
One conversation can touch several paths, as the final walkthrough shows. We covered the tool side in Microsoft Copilot Studio Tools Explained: Give Your AI Agent the Power to Search, Act, and Automate, so this guide stays on the knowledge layer. The same read-versus-act split appears in plain automation too, as in How to Automate Repetitive Tasks with AI, where looking something up and doing something are separate jobs.
The Full Path at a Glance
Here are the steps in order. Each one is explained below with the exact PAW Labs example.
Define the questions
Prepare the guide
Fix the headings
Build the inventory
Choose the source type
Add it to the agent
Name and describe it
Agent or topic level
Remove conflicts
Test retrieval
Check citations
Check permissions
Set the fallback
Maintain it
Score it
Walk a full conversation
What "Knowledge Base" Means in Copilot Studio
Copilot Studio does not have a single feature literally called "Knowledge Base." In practice the term covers a few related ideas, and mixing them up causes confusion in design discussions.
A single file or page, such as the PAW Labs VPN Troubleshooting Guide.
A document, site, or data location connected to the agent, with a name and description.
The set of sources attached to the PAW IT Helpdesk Assistant, or to one of its topics.
Our term for the whole practice: what to include, who owns it, how it is described, tested, and retired.
The goal is not to expose every available document. The goal is to make the right approved VPN information retrievable, and nothing else.
Step 1: Define What the Agent Should Know
Beginners often start by asking which files they can upload. We start by asking which questions the PAW IT Helpdesk Assistant must answer. Write them down first. The list tells you which documents you actually need and gives you the raw material for your tests later.
| Question the agent must answer | Where the answer lives | Needs knowledge? |
|---|---|---|
| My VPN will not connect. What should I check? | Guide: basic connectivity checks | Yes |
| The VPN says authentication failed. | Guide: authentication errors | Yes |
| My account seems locked. | Guide: account lock or access problems | Yes |
| When should I contact IT? | Guide: when to stop troubleshooting | Yes |
| What do I put in a support request? | Guide: what to include in a support request | Yes |
| What is the status of PAW-1042? | A live ticket record | No, use a tool |
Notice that five of the six questions are covered by the one guide. That is a good sign for starting small. The sixth does not belong in knowledge at all.
Step 2: Prepare the PAW Labs VPN Troubleshooting Guide
If you followed the first tutorial, you already have the guide, and you should inspect it rather than rewrite it. If you are starting fresh, create a document with the same six sections so the rest of this guide lines up with your setup.
What a working VPN setup looks like
First things to try when the VPN will not connect
Approved checks when sign-in fails
When the account, not the network, is the issue
The point where IT support takes over
The details a ticket needs
That is a sound structure, and the order follows how a person works through an issue. The first tutorial focused on content, so here we audit the guide as a retrievable asset and add what it is missing. Do not silently swap it for a different document.
| Audit question | What we look for | Fix if missing |
|---|---|---|
| Is the title descriptive? | Names the product, topic, and audience | Keep "PAW Labs VPN Troubleshooting Guide" |
| Are headings specific? | Each heading names a problem or decision | Rename vague headings (Step 3) |
| One problem per section? | No section mixes unrelated topics | Split mixed sections |
| Steps in a followable order? | Numbered, simplest check first | Convert paragraphs to numbered steps |
| Unsupported procedures excluded? | Only approved steps appear | Remove anything IT has not approved |
| Clear stop condition? | States when to contact IT support | Strengthen "When to Stop Troubleshooting" |
| Version and review date? | A header block near the top | Add one (template below) |
| Named owner? | A team accountable for updates | Assign IT Support |
Here is the header block we recommend for the updated version of the guide. It costs five lines and answers the freshness and ownership questions at once:
Audience: PAW Labs employees
Owner: IT Support
Version: 1.1 | Last reviewed: [date]
Status: Current. Replaces all earlier VPN guides.
Step 3: Reorganize the Guide So It Can Be Found
A heading tells both humans and the agent what a section is about. Here is a small before-and-after from the VPN guide:
| Before | After | Why it helps |
|---|---|---|
| General Information | Supported VPN Setup for PAW Labs Employees | The topic is in the heading itself |
| Other Problems | When to Stop Troubleshooting and Contact IT | Matches how an employee would phrase the need |
| Errors | VPN Authentication Failed: Approved Checks | Specific symptom plus an approved procedure |
| Notes | What to Include in a VPN Support Request | Links directly to the ticket workflow |
Beyond headings, we apply a few plain habits: one logical topic per section, the same term for the same thing throughout (always "VPN client," never a mix of names), numbered steps for procedures, explicit warnings, and one current version of the document. We would not trust anyone who quotes a magic section length or chunk size without a source you can verify for your own setup.
Step 4: Build a Knowledge Inventory
Before adding anything else, we list every candidate item. The first row exists today. The other rows are optional extensions we use to teach knowledge organization on the VPN scenario. They are not part of the project yet.
| Knowledge item | Purpose | Authority | Owner | Review | Status |
|---|---|---|---|---|---|
| PAW Labs VPN Troubleshooting Guide | Approved employee troubleshooting steps | Official project guidance | IT Support | Monthly or after a procedure change | Exists |
| PAW Labs VPN Access Policy | Who may use VPN, and under what conditions | Approved policy | IT / Security | When policy changes | Proposed |
| PAW Labs VPN FAQ | Common questions and clarifications | Supporting guidance | IT Support | Monthly | Proposed |
| Archived VPN Guide | Historical reference only | Not current guidance | IT Support | Keep only if needed | Keep out of the agent |
For each item we also record the audience, the version, the last-reviewed date, whether the content belongs in knowledge or should come through a tool, and whether the intended users are allowed to see it. Those last two checks catch the problems that cause the most damage later.
Step 5: Choose the Right Source Type
For the PAW Labs VPN guide, an uploaded file is enough. The other options matter once content grows or changes often. Availability depends on your tenant, licensing, administrator policy, and the Copilot Studio experience, so treat these cards as a planning aid and confirm what your environment offers.
Best starting point
Best use: Small, controlled set such as the VPN guide
Watch out: A changed original does not fix the uploaded copy by itself. You must update it.
Good next step
Best use: Documents maintained centrally by a team
Watch out: Connect a controlled site or folder, not everything. Verify permission behavior for your chosen path.
Easy, but narrow fit
Best use: Approved public documentation
Watch out: Wrong home for confidential VPN procedures. A vendor's public page is not PAW Labs policy.
Needs planning
Best use: Structured organizational information
Watch out: Tables do not behave like documents. Live ticket status is usually a lookup job.
Not needed for PAW Labs
Best use: Large collections with existing search infrastructure
Watch out: Indexing, security, and refresh design need specialist skills.
If you plan to upload files, check the current size and format limits before you build around a specific number. We deliberately avoid quoting limits here because they change.
Step 6: Add the Guide to the PAW IT Helpdesk Assistant
This is the hands-on part. Follow it in order. Because the agent and guide usually already exist from the first tutorial, the rule is inspect first, change only what needs changing. Do not create a second agent, and do not add a second copy of the guide.
Sign in, choose the right environment, and open the existing PAW IT Helpdesk Assistant in the standard experience used in the first tutorial. If you are starting without the first tutorial, create the agent there first, then return here.
Look for the Knowledge page, or the Add knowledge option on the agent's Overview page.
Look for the PAW Labs VPN Troubleshooting Guide in the list and read its status. If it is there, open it and inspect it. Skip to step ⑦.
Select the add option and review the source types your tenant offers. These vary by environment.
The guide is a small, controlled document, so an uploaded file fits. Select the file option and pick the guide from your computer. If your tenant does not offer it, ask your administrator what is allowed.
Check that the right file and the right version were selected. Remember the header block with version 1.1.
Use the copy-ready text from Step 7. This is the single most useful edit you will make.
Wait until the source shows as available. Ready means connected, not correct.
The first tutorial relied on generative orchestration, which lets the agent choose among knowledge, topics, and tools. Confirm it is still enabled for this agent.
Ask one simple VPN question to confirm the agent responds, then move on to the full tests in Step 10.
Heads-up: button names, menu positions, and available options change between releases and differ by tenant. If a label here does not match your screen, look for the closest equivalent rather than assuming something is broken. Also confirm that you are in the standard experience, because other agent-building experiences have different navigation. If the guide's content changed since you uploaded it, edit or replace the source so the agent is not using a stale copy.
Step 7: Write the Source Name and Description
This change takes the least effort and, in our view, gives the most benefit. With generative orchestration, the agent uses the name and description to decide whether a source is relevant to a question. A vague description gives it very little to work with.
Name: VPN doc
Description: VPN document.
No scope, no audience, no limits, no fallback.
Name: PAW Labs VPN Troubleshooting Guide (Current)
The name already states the owner, the topic, and that this is the current version, before the description even starts.
Copy-ready description for the existing guide:
| Description element | Example from the text above | What breaks if it is missing |
|---|---|---|
| Purpose and audience | "Approved PAW Labs IT guidance for employees" | Agent cannot tell this from general advice |
| Scope | Setup, network checks, authentication, account lock | Relevant questions may skip the source |
| When to use | "Questions about resolving VPN connectivity issues" | Source selected inconsistently |
| When not to use | "Not for live ticket status, account details, passwords" | Agent may try to answer live-record questions from a static document |
| Missing-answer rule | "Direct the employee to IT support" | Gaps get filled with guesses |
If you add the optional VPN Access Policy, give it a description with a clearly different purpose, for example: "Approved PAW Labs policy describing who may use the company VPN and the conditions of use. Use for eligibility and access-rule questions. Do not use for step-by-step troubleshooting; use the PAW Labs VPN Troubleshooting Guide for that." Two descriptions that overlap blur the line between the sources.
A description does not guarantee correct retrieval. It improves the odds that orchestration picks the right source, and testing is still the only way to know.
Step 8: Decide Between Agent-Level and Topic-Level Knowledge
Available across the agent's conversations.
Fits the VPN guide, because VPN support is the assistant's main job.
Plus: easy reuse
Minus: more chances of irrelevant retrieval as sources multiply
Used inside a specific topic or generative-answer step.
Fits a specialized guide, such as device enrollment, that only one topic needs.
Plus: narrower, tidier retrieval
Minus: not available elsewhere in the agent
For the PAW IT Helpdesk Assistant we keep the VPN guide at the agent level. The right choice depends on what the agent is responsible for and how its topics are built. Where exactly these options live in the interface can vary, so check the current screens in your tenant.
Step 9: Remove Outdated and Conflicting VPN Documents
Suppose the current guide tells employees to use the company-approved VPN client, while an older document describes a previous client and a retired sign-in procedure. If both are reachable, the agent has two plausible answers and no reliable way to know which is current. Uploading the new guide does not erase the old one from wherever else it is connected.
Do not go looking for a hidden "priority" or "ranking" switch to settle conflicts. The dependable fix is governance: keep one authoritative document, label archived material clearly, remove obsolete content from active retrieval, and state in each source description what the source is and is not for. Teams that want to turn this into a repeatable routine can read how we frame production AI agents, where ownership and review are treated as part of the build.
Step 10: Test Whether the Agent Actually Finds the Right Answer
A source showing as ready only means it is connected. It says nothing about whether answers are correct. We build a retrieval test matrix from realistic employee questions and keep it as a standing regression test. Every row below is a recommended test with an expected behavior. Recording the actual results is your job in your own environment.
| # | Test | Employee question | Expected behavior |
|---|---|---|---|
| 1 | Basic troubleshooting | My VPN will not connect. What should I check? | Retrieves the basic connectivity guidance |
| 2 | Authentication | The VPN says authentication failed. | Retrieves the authentication section |
| 3 | Different wording | I can open websites, but cannot reach internal apps. | Recognizes a VPN-related problem and retrieves relevant guidance |
| 4 | Already attempted | I restarted my laptop. What else can I try? | Does not blindly repeat the restart; asks for missing context or gives the next supported step |
| 5 | Stop condition | I tried the approved steps and still cannot connect. | Explains when to contact IT support |
| 6 | Unknown procedure | What is the new 2027 VPN security rule? | Admits the guide does not confirm it |
| 7 | Outdated content | An old guide says to use a different VPN client. | Does not present obsolete steps as current |
| 8 | Out of scope | Can you help me choose a personal VPN provider? | Respects the PAW Labs assistant's scope |
| 9 | Tool boundary | Is ticket PAW-1042 resolved? | Does not invent a status from the guide; uses a ticket lookup only if one has been configured |
| 10 | Sensitive request | What is another employee's VPN password? | Refuses and never exposes credentials |
For every test, record four things:
- The actual response, word for word.
- Which knowledge source was used, where the activity map shows it.
- Whether the answer matches the relevant section of the guide, and whether a citation or source reference appeared in that scenario.
- Whether the agent wrongly triggered a tool or topic, or correctly admitted a gap.
Two conversations show what we are looking for. These are expected-response examples, not recorded outputs, and exact wording will depend on your guide's text.
Employee: The VPN says authentication failed.
Assistant: According to the PAW Labs VPN Troubleshooting Guide, authentication errors are handled with a specific order of checks. [Steps from the authentication section appear here.] If these do not work, the guide points to an account lock as the next possibility. Would you like help contacting IT support?
Employee: What is the new 2027 VPN security rule?
Assistant: The PAW Labs VPN Troubleshooting Guide does not confirm a 2027 VPN security rule, so I cannot give you an answer I can stand behind. IT support can confirm current policy. If it helps, I can help you send them the question.
Worked Example: The Same VPN Question, Two Setups
This is an illustrative scenario we designed to show how setup choices change expected behavior. It is not a recorded benchmark. We use the original question: "My VPN keeps disconnecting every 20 minutes. I already restarted my laptop."
Source description: "VPN document."
Old VPN guide also active.
Headings: "General Information," "Other Problems."
No stated fallback.
Expected risks: repeats the laptop restart, may mix old-client advice with current steps, may guess when the guide is silent.
Scoped description with limits.
Only the current guide is active.
Specific headings, owner, and version.
Fallback rule stated.
Expected behavior: skips the restart, finds the relevant connectivity checks, and points to IT support if the guide does not cover the pattern.
| Fix step | What we changed | Test question to run afterward |
|---|---|---|
| 1. Remove the old guide from active knowledge | One current VPN source remains | "An old guide says to use a different VPN client." |
| 2. Rewrite the source name and description | Scope, limits, and fallback added | "Is ticket PAW-1042 resolved?" |
| 3. Rename vague headings | Problem-based headings | "I can open websites, but cannot reach internal apps." |
| 4. Add the version and owner header | Freshness is visible | "What is the new 2027 VPN security rule?" |
| 5. Re-run the full test matrix | Compare to expected behavior | All ten tests above |
We change one thing at a time and re-test after each change, so we know which one helped. Changing everything at once is the knowledge version of pasting one giant prompt and hoping for the best.
Relative retrieval risk (illustrative only, not a benchmark)
Step 11: Check Citations, but Do Not Trust Them Blindly
When a source and configuration support it, a citation lets an employee see where an answer came from. That is useful, with three caveats. A citation does not prove the document is current. A correct citation can still point to an outdated or non-authoritative file. And not every source type behaves the same way, so confirm what each of your sources shows in your own tests. For important operational steps, employees should verify against the approved guide. This is the same discipline we describe in AI Hallucinations Explained: Why AI Gives Wrong Answers and How to Verify Them, where a confident answer and a correct answer turn out to be different things.
Step 12: Check Permissions, Because Knowledge Is Also an Access Question
Everything in the knowledge layer is also a security decision. Three points stand out.
- The access you have as the maker is not the access your employees have. Test with the identity of the intended user.
- Different source types handle permissions differently. SharePoint behavior depends on the integration path and authentication setup, and uploaded files have their own characteristics. Never assume that adding a document automatically creates a secure, per-employee permission model. Verify it for each source.
- If an employee asks for an internal security procedure they are not authorized to see, the agent must not reveal or summarize the restricted content.
This stoplight shows how we sort content before it ever reaches the VPN guide:
Approved troubleshooting steps, supported setup, when to contact IT, what to put in a ticket.
Policy details, internal contact paths, anything only some employees should see.
Passwords, authentication codes, private keys, other employees' account details.
For a wider look at how agents get misled or over-exposed, see AI Security Threats, which covers the attack patterns that make these access checks worth the effort.
Step 13: Set a Controlled Fallback for Missing Answers
A good agent says "I do not have that" in a controlled way. Here is the flow we designed for the PAW IT Helpdesk Assistant:
Knowledge retrieval and ticket creation are separate responsibilities in the same agent. Knowledge reads, and the ticket workflow acts. This hand-off follows the pattern described in Agentic AI Explained: How AI Agents Plan, Use Tools, and Work Autonomously, where the agent decides which capability a request actually needs.
Copy-ready fallback wording: "The approved PAW Labs VPN Troubleshooting Guide does not cover this situation, so I cannot give you a reliable answer. Please contact IT support. If you would like, I can collect the details and create a support ticket for you."
Step 14: Maintain the VPN Guide Over Time
A guide that was correct last spring can still produce a convincing answer today. Knowledge needs a lifecycle, not a one-time upload.
| PAW Labs maintenance checklist | Done? |
|---|---|
| IT Support is named as owner of the VPN guide | ☐ |
| Guide reviewed after every approved procedure change | ☐ |
| Version and last-reviewed date updated | ☐ |
| Obsolete material removed from active knowledge | ☐ |
| Uploaded copy refreshed after the original changes | ☐ |
| Full retrieval test matrix re-run after each update | ☐ |
| Unanswered and wrongly answered questions reviewed | ☐ |
| A record of the approved source is kept | ☐ |
Do not assume automatic synchronization. Whether and how quickly a change reaches the agent depends on the source type, so confirm this for each source you connect.
Step 15: Score the Source Before You Publish
This is our own practical framework, not an official Microsoft scoring system. We score each source on seven criteria before it goes live.
| Criterion | Key question | Pass | Needs improvement | Fail |
|---|---|---|---|---|
| Relevance | Does it answer real questions? | Matches test questions | Partly | Off-topic |
| Authority | Is it approved? | Named approver | Unclear approval | Unofficial |
| Freshness | Is it current? | Dated and recent | Date missing | Known outdated |
| Structure | Can the right section be found? | Specific headings | Some vague headings | One wall of text |
| Security | Can intended users access it safely? | Tested with user identity | Not yet tested | Over-exposed |
| Testability | Can answers be checked against it? | Test matrix exists | Ad hoc checks | No testing |
| Ownership | Is someone responsible? | Named owner and schedule | Owner, no schedule | No owner |
Scoring guide: any "Fail" means do not publish until it is fixed. Three or more "Needs improvement" ratings mean you should improve the source before adding it. Seven passes means it is ready for a pilot group, and monitoring continues after that.
Step 16: Walk Through One Conversation, From Knowledge to Tools to Done
This final step ties everything together. The conversation shows the knowledge layer, a ticket action, and a live lookup working together on one VPN problem. It is an illustrative expected flow we designed, not a recorded transcript, and exact wording will depend on your guide and configuration. Each exchange shows what the agent is doing behind the scenes.
"My VPN keeps disconnecting every 20 minutes. I already restarted my laptop. What should I do?"
Skips the restart, gives the next supported checks from the guide's basic connectivity section, and asks the employee to report back.
"I tried those. It still drops every 20 minutes."
Explains that the guide says to contact IT after these checks, and offers to create a support ticket.
"Yes, please create one."
Summarizes the details and asks: "Should I create this ticket?"
"Yes, confirm."
Confirms the ticket was created and shares the reference, in our sample environment PAW-1042.
"What is the status of PAW-1042 right now?"
Reports the status returned by the lookup. Without a lookup, it says it cannot check live ticket status and suggests contacting IT support.
"Thanks, that's all."
The employee has tried the approved steps, has a ticket reference, and knows where live status comes from. Nothing was invented along the way.
| Moment | Capability used | Reads or writes? |
|---|---|---|
| Disconnecting every 20 minutes | Knowledge: connectivity guidance | Reads |
| Still failing after the checks | Knowledge: stop condition | Reads |
| Create a ticket | Topic and agent flow, with confirmation | Writes a record |
| Status of PAW-1042 | Lookup tool, if configured | Reads a live record |
The knowledge layer handled the first two moments, and it never touched a record. That separation is why the steps above spend so much effort on keeping the guide accurate and findable.
Common Beginner Mistakes
More documents means more conflicts. Start from the employee questions, not the files.
Two copies of the VPN guide will drift apart.
"VPN document" gives orchestration nothing to work with.
The retired client procedure keeps resurfacing.
PAW-1042's status belongs to a lookup, not a document.
Ready means connected, not correct.
Test as the employee, not as the maker.
The 2027 rule test exists for a reason.
A citation shows where, not whether it is current.
A guide with no owner and no review date quietly goes stale.
Limitations and Reality Check
Better knowledge design improves reliability. It does not make the agent perfect.
- Retrieval can still select irrelevant content, and the agent can still misread a correct passage.
- Conflicting or poor-quality documents produce inconsistent or poor answers.
- Sources differ in authentication, indexing, refresh, and citation behavior.
- Administrator and environment policies can limit which sources you can use.
- Large enterprise search setups may need specialist skills.
- Knowledge does not replace tools for live records or actions.
- Grounding in knowledge reduces hallucination risk. It does not eliminate it.
Check current Microsoft licensing and environment requirements before you scale, and plan time for testing and upkeep. The work does not end at publication.
Where the PAW Labs Project Stands Now
The agent now reads from a described, owned, tested source, and everything that can change a record stays in the ticket workflow. This is the resulting flow:
| Layer | Role in PAW Labs | Read or write? |
|---|---|---|
| Knowledge (VPN guide, optional policy and FAQ) | Approved answers | Read only |
| Source names and descriptions | Help orchestration pick the right source | Configuration |
| Retrieval test matrix | Proves behavior after every change | Process |
| Existing ticket-creation workflow | Creates the support ticket after confirmation | Writes a record |
Frequently Asked Questions
What is a knowledge source in Microsoft Copilot Studio?
It is information the agent can use to ground its answers, such as an uploaded file, a website, SharePoint content, or another supported data source. Which types you see depends on your environment.
Is a knowledge base the same as a document?
No. A document is one file. A knowledge source is a connected item with a name and description. "Knowledge base" is our term for the whole design: what is included, who owns it, and how it is tested.
Can I add knowledge to an existing Copilot Studio agent?
Yes. Open the agent, go to its Knowledge page or the Add knowledge option, and inspect what is already there before adding anything. Labels vary by tenant and release.
Should I upload all company documents?
No. Start from the questions the agent must answer and connect only the authoritative sources those questions need. Extra documents add maintenance work and create conflicts.
What is the difference between agent-level and topic-level knowledge?
Agent-level knowledge is available across the agent's conversations. Topic-level knowledge is scoped to a specific topic or generative-answer step, which can reduce irrelevant retrieval.
Can Copilot Studio use SharePoint as a knowledge source?
Yes, depending on your environment. Permission behavior depends on the integration path and authentication setup, so verify it with the identity of your intended users.
Can Copilot Studio use uploaded PDF and Word files?
Uploaded files are a supported approach. Check the current supported formats, size limits, and prerequisites for your environment before you rely on a particular file.
Why is my agent not finding the right information?
Common causes are a vague source description, unclear headings, conflicting or outdated documents, a question outside the document's scope, or a source the user cannot access. Work through those in order and re-run your test matrix after each change.
Do knowledge-source descriptions matter?
Yes. With generative orchestration, the name and description help the agent decide whether a source is relevant. They improve the odds of good retrieval but do not guarantee it.
Can a knowledge base eliminate hallucinations?
No. Grounding reduces the risk, but retrieval can still pick the wrong passage and the agent can still misread a correct one. Testing and maintenance stay necessary.
Does a knowledge source retrieve live ticket status?
Not reliably. A status like PAW-1042 is a live record, so it belongs to a tool or data lookup rather than a static document.
How do I prevent outdated documents from being used?
Keep one authoritative current source, label or retire older versions, assign an owner, record a review date, and re-test after every update. Uploading a new file does not automatically settle conflicts with older material.
Conclusion
A useful knowledge base is not measured by how many documents are connected. It is measured by whether the agent can find the right approved information, explain it accurately, respect access boundaries, and admit when the answer is not there. The PAW IT Helpdesk Assistant started this chapter with one connected guide and a "ready" status. It ends with a described source, a clearer document, a defined owner, a test matrix, and a fallback that hands off to the ticket workflow built in the first tutorial. Knowledge supplies the approved answers, and tools remain responsible for live data and actions. If you want to keep building along with us, the rest of our Copilot Studio series continues from here.
Comments
Post a Comment