How to Build a Reference Library and Product Knowledge Base for Copilot
The SharePoint structure and retirement rules that get Microsoft 365 Copilot and Copilot Studio agents to cite the current version
What You'll Achieve
A SharePoint reference library where each document has one owner and one current version, tagged with metadata Copilot can use, plus a retirement path for superseded content and a citation test you rerun every month
Who This Is For
Knowledge managers and Microsoft 365 admins who own a reference library or product knowledge base
Before You Start
- SharePoint Online with admin access to the Content type gallery and the Term Store
- At least one Microsoft 365 Copilot license in the tenant. Copilot Studio agents need it for tenant graph grounding with semantic search and for SharePoint files over 7 MB
- Microsoft Purview sensitivity labels published and enabled for files in SharePoint and OneDrive
- A named owner for each product or reference domain, with authority to retire content
- The permission model and governance cadence from our knowledge hub guide, or an equivalent you already run
- A set of 30 real questions per product, taken from support tickets and field requests
Readiness Checklist
Before you begin implementation, ensure you have these items in place:
- Library scope written: products and document types in, drafts and project files out
- One library or site agreed for each product
- Owner named for each product, with authority to retire content
- Content types and columns agreed: Product, Version, Document Status, Effective Date, Owner, Supersedes
- Term sets for Product and Region created as closed sets, with synonyms
- Sensitivity labels enabled for SharePoint and OneDrive files
- Microsoft 365 Copilot license status in the tenant confirmed
- Largest files and SPFx-based pages identified
- Question set of 30 real questions per product collected
- Reader test accounts ready, with and without access to restricted content
Step-by-Step Guide
Decide What Belongs in the Library and Where It Lives
Define the reference set and give each product one home before you design metadata
- Write a one-page scope that lists the products and the document types the library holds. A reference library records what is true now, so working drafts and project files stay in team sites.
- Give each product one library or site. Copilot Studio SharePoint knowledge searches a registered URL and its subfolders, so one location per product lets you scope agents cleanly.
- Publish reference content as modern SharePoint pages or as DOCX, PPTX and PDF files. Copilot Studio does not use classic ASPX pages, or modern pages built on SPFx components, to generate answers.
- Keep each document to one product and one subject. A spec sheet that covers four product lines makes every citation ambiguous.
- Put the product name and version in the title and the effective date in the first paragraph. Copilot Studio cannot answer questions that reference a file name and cannot filter on column values, so the text itself has to carry these facts.
- Carry over owner and review-date rules from your hub governance. Our knowledge hub guide covers the audit, permission model and review cadence, so this guide does not repeat them.
Build Content Types and Metadata Once, Centrally
Publish shared content types and term sets so every product library tags content the same way
- Create a site content type for each reference document type, for example Product Specification, Release Note, Installation Guide and FAQ. Publish them from the Content type gallery in the SharePoint admin center so every product site receives the same definitions.
- Use these columns at minimum: Product, Version, Document Status, Effective Date, Owner and Supersedes.
- Back Product and Region with managed metadata. A closed term set stops people from entering "Product X" and "PX" as two different values.
- Add synonyms to terms. A term can carry several labels, so the name the field uses and the official name resolve to one tag.
- Know what metadata does for Copilot. When a user attaches a library or folder in Microsoft 365 Copilot on the web, Copilot uses that library's column metadata as relevance signals. Copilot Studio SharePoint knowledge can filter only on Title, Author, Modified by and Modified on.
- Backfill metadata on existing files with autofill columns. Each column runs a prompt against the file, managed metadata term sets of up to 100 terms are supported, billing runs through pay-as-you-go, and Microsoft tells you to check the results because AI output can be wrong.
Set Permissions and Sensitivity Labels at the File Level
Copilot answers with each user's own access, so access design decides what any answer can contain
- Make reference content readable by the people who need it. Copilot and agents only return content the asking user can already open.
- Label files as well as sites. Items do not inherit a container label, so a document in a Confidential site shows no label in Copilot unless the file carries one.
- Check every label that applies encryption. Copilot needs the EXTRACT usage right to return text, and Copilot Studio can show an encrypted file as Ready and still return no answer from it.
- Keep Double Key Encryption off reference content. Copilot and agents cannot access items protected with DKE.
- If certain labeled files must never be summarized, ask your Purview admin for a DLP policy on the Microsoft 365 Copilot location. Copilot then points to those files with a link and does not summarize them.
- For sites that are not ready, turn on Restrict content from Microsoft Copilot in the SharePoint admin center. Users keep their access, and the site drops out of Copilot and organization-wide search until you remove the setting.
Control Versions, Drafts and Retirement
Make exactly one current version of each fact visible to readers
- Turn on major and minor versions with Require content approval for submitted items, and use Set who can see draft items to keep drafts away from readers. Copilot works from what the user can open, so confirm with a reader account that answers come from approved versions.
- Set version history limits that fit your records needs. Admins set limits at organization, site or library level, and items under a retention policy keep their versions regardless of library limits.
- Retire a superseded document the day its replacement publishes. Set Document Status to Retired and record the replacement in Supersedes. Then move the old file out of the product library.
- Archive retired content you must keep. Archived SharePoint data is excluded from the semantic index that grounds Copilot, and Microsoft 365 Archive keeps permissions and metadata if you reactivate a site.
- Apply retention labels with disposition review where retired content must be deleted on a schedule. Reviewers approve deletion in Purview, and each action is recorded in the audit log.
- If a site must stay live and stay out of Copilot, set Allow this site to appear in search results to No. That removes it from Microsoft Search and the semantic index together.
Connect the Library to Copilot and to Agents
Use the library as it stands for Microsoft 365 Copilot, and scope agents tightly for product questions
- Microsoft 365 Copilot needs no connection step. Text-based files in searchable SharePoint sites are indexed, and Copilot Chat shows inline citations plus a Sources list under each answer.
- For a scoped assistant without code, create an agent in SharePoint from the product library. You need a Microsoft 365 Copilot license, or pay-as-you-go billing for SharePoint agents, plus edit rights on the site.
- For a Copilot Studio agent, add each product library as SharePoint knowledge with a clear description and keep authentication on Authenticate with Microsoft.
- If some users lack Microsoft 365 Copilot licenses, turn on Tenant graph grounding with semantic search so their questions also use the semantic index. It requires Authenticate with Microsoft and bills extra Copilot Credits.
- Serve many products from one agent with a URL variable. Insert a variable into the SharePoint knowledge URL, set it from the user's product choice, and each question searches only that product's library.
- Turn off Allow ungrounded responses and Web Search for agents that answer from the library. The agent then withholds answers that carry no citation to your content.
- Leave citation formatting alone in the instructions. Microsoft warns that instructions which alter citations can cause the orchestrator to discard valid answers. Keep titles short too, because Teams shows at most 20 citations per response and shortens titles beyond about 80 characters.
Test Retrieval and Citations Every Month
Score each product on whether answers cite the current, correct document for a normal user
- Run the question set as a reader account in Microsoft 365 Copilot and in each agent. Score two things per answer: the right document, and its current version.
- In Copilot Studio, load the questions into an Evaluation test set and add Keyword match on the version number or effective date. A wrong version then fails even when the wording reads well.
- Add retirement tests. For each document retired this month, ask a question only the old version answered and confirm the answer now comes from the replacement.
- Add label tests: one question per sensitivity label, run as a user with access and a user without it.
- When a citation misses, check the title and first paragraph, then search for duplicate copies, before you change the agent.
- Track the score per product and report it with the content health figures from your hub governance. A falling score usually shows up before user complaints do.
Implementation Worksheet
Use this worksheet to track implementation. Give every task a named owner and a due date before you start.
| Task | Owner | Due Date | Status | Notes |
|---|---|---|---|---|
| Write library scope and assign product owners | Knowledge Manager | Week 1 | Pending | One home per product |
| Build content types and term sets | SharePoint Admin | Week 1 | Pending | Publish from the Content type gallery |
| Configure versioning, approval and draft visibility | Site Owner | Week 2 | Pending | Readers see approved versions only |
| Backfill metadata and retire duplicates | Content Team | Weeks 2-3 | Pending | Check autofill output by hand |
| Test labels and encryption in Copilot and agents | Purview Admin | Week 3 | Pending | One file per label, two accounts |
| Configure agents and run baseline citation test | Power Platform Maker | Week 4 | Pending | Keyword match on version or effective date |
| Start the monthly citation report | Knowledge Manager | Week 5 | Pending | Score per product, retirement tests included |
Common Pitfalls and How to Avoid Them
Copilot cites last year's spec sheet
Two readable versions of the same document is the cause. Retire the old file the day the new one publishes and record the link in Supersedes. Archive the old version or move it where readers cannot open it, then add a retirement test so the next monthly run catches a repeat.
The metadata is complete, yet answers ignore it
Metadata helps Microsoft 365 Copilot when a user scopes a question to a library or folder. Copilot Studio SharePoint knowledge cannot answer questions that filter on column values. Keep columns for search and governance, and repeat the product name, version and effective date in the title and first paragraph.
Labeled files show Ready in the agent but never appear in answers
Encryption is the usual cause. Copilot needs the EXTRACT usage right to return text, and it never returns content protected with DKE. Files added through the Upload files path cannot be indexed when they carry Confidential or Highly Confidential labels. Test one file per label in each agent before you rely on it.
Large manuals return nothing in Copilot Studio
Without a Microsoft 365 Copilot license in the tenant, Copilot Studio generative answers only use SharePoint files under 7 MB. Split long manuals by chapter, which also improves citations, or license the tenant and turn on tenant graph grounding for larger files.
Metrics to Track
Current-Version Citation Rate
Share of test questions where the answer cites the correct document in its current version, run monthly as a reader account
Retired Content Cited
Answers on the monthly run that cite a document with Document Status set to Retired
Metadata Completeness
Share of published documents with Product, Version, Document Status, Effective Date and Owner filled in
Time to Retire
Days between publishing a new version and removing the old one from the product library
Withheld Answer Rate
Share of agent questions where no grounded answer was returned, reviewed weekly to find content gaps
Prompt Pack
Copy and use these proven prompts to get started quickly. Customize them for your specific needs.
What is the current [specification] for [product] version [version]? Include the effective date.What changed between [product] version [x] and version [y]?Which document replaced [retired document title]?Is [feature] supported on [product] in [region]? Name the document you used.Summarize the installation prerequisites for [product] for a new field engineer.List the known issues in the latest release note for [product].Answer only from the [product] library. If the library does not cover the question, say so and name the owning team.When the user does not name a product, ask which product before you search.State the product version your answer applies to at the start of every answer.If two documents disagree, tell the user both exist and give the effective date stated in each.Go-Live Checklist
Complete these items before going live to ensure a successful launch:
- Content types published from the Content type gallery to every product site
- Metadata complete on every published document
- Content approval and draft visibility set on each product library
- Superseded documents retired, archived or moved out of product libraries
- Retention labels with disposition review applied where deletion is required
- Labeled and encrypted files tested in Copilot and each agent
- Agents scoped per product, with ungrounded responses and Web Search off
- Baseline citation score recorded for each product
- Retirement and label tests added to the monthly run
- Monthly citation report owner named
Frequently Asked Questions
How is a reference library different from a knowledge hub?
A knowledge hub organizes policies and procedures for the whole company and focuses on permissions and findability. A reference library holds versioned facts, such as product specifications and release notes, where the wrong version is the main risk. Most organizations need both. Build the hub first with our knowledge hub guide, then add product libraries with the version and retirement rules in this guide.
Do I need to rebuild the library for Copilot Studio agents?
Usually you only need to adjust it. Agents read the same SharePoint content as Microsoft 365 Copilot, with a few extra limits: modern pages only, DOCX, PPTX and PDF for files, and a 7 MB file limit when the tenant has no Microsoft 365 Copilot license. Check your largest manuals and any pages built on SPFx components first.
How does Copilot use SharePoint metadata?
Microsoft 365 Copilot uses a library's column metadata as relevance signals when a user attaches that library or folder on the web. Copilot Studio agents can filter a SharePoint source on Title, Author, Modified by and Modified on, and they cannot answer questions that filter on other columns. Metadata still drives views, search refiners and retention rules, so keep it complete.
Should product documentation move to SharePoint or stay in Confluence or ServiceNow?
You can leave it where it is. Copilot connectors bring external content into Microsoft Graph with its access controls, and Copilot Studio can use Confluence, ServiceNow, Salesforce and Zendesk as knowledge sources that sync every four to six hours. Move content only when you also want SharePoint versioning and retention labels to govern it.
How do we prove which file an answer came from?
Users see inline citations and a Sources list in Copilot Chat. For a record, Microsoft Purview Audit captures Copilot interactions, including references to the Microsoft 365 files accessed and their sensitivity labels. Pair that with your monthly citation test and you have both live evidence and a repeatable check.
Can you build this for us?
Yes. We scope one workflow with you, such as a product knowledge agent for one product line, build it in your environment in two weeks, and you pay $10,000 only after the acceptance criteria are met. The library work in this guide is usually the larger effort, so we agree its scope in the same session.
Rather have us build it?
Describe the workflow and get acceptance criteria and a price in under a minute. The first build is $10,000, two weeks, in your environment. You pay only after every signed criterion passes.