Prepare content so the AI assistant can provide accurate, relevant answers. Well-structured content improves retrieval, covers common questions, and helps the assistant handle more complex conversations.
The assistant searches indexed knowledge, including Knowledge Management articles and documents, as well as connected SharePoint sites.
For information on which content is indexed, see Content the AI assistant can use.
Identify key knowledge sources
Give the assistant the policies and guides employees already ask about. You do not need to index every file in your organization.
- Publish HR policies, how-to articles, and FAQs as Knowledge Management articles whenever possible.
- Upload existing PDF, Word, and Excel documents as knowledge documents when you do not need a full article.
- Connect SharePoint sites that contain intranet policies, local procedures, site pages, and useful lists.
- Prepare unsupported sources before adding them. For example, export native Google Docs, Sheets, and Slides to a supported format, or provide a text transcript for video and audio content.
Organize content into clear topics
Make each item easy to find and use.
- Prefer one article or a focused section per topic rather than a long handbook that combines multiple policies.
- Use headings that match the questions employees ask, such as "How do I request parental leave?" rather than generic headings like "Appendix C".
- Assign a category to Knowledge Management articles so related content stays together.
- Make titles, headings, and opening paragraphs specific. These help the assistant identify relevant content.
Keep content up to date
Outdated content can still be retrieved until you retire it.
- Review high-traffic policies regularly, such as quarterly or after a policy change.
- Publish updates in Knowledge Management so the index can refresh.
- Unpublish or delete retired articles and documents to remove them from the index. Removing an article also removes its attachments.
- For SharePoint content, update or delete the content at the source and allow the scheduled sync to process the change. Disconnect a site if it should no longer be used to provide knowledge.
Adapt content for different audiences
The assistant uses information about the employee, such as country and audience, to help identify relevant content.
- Set country or global and audience on Knowledge Management articles. Audiences can include all employees, managers, and HR.
- Create separate articles when a rule genuinely differs by country or role. Do not combine multiple exceptions into a single global document when separate content would be clearer.
- Put the audience or location in the title or first heading when content is intended only for managers or a specific location. SharePoint tagging can also use these signals.
- Keep the tone and level of detail appropriate for the audience. Avoid combining employee self-service and manager-only processes in the same item.
Format content for processing
The assistant can only search content that it can read.
- Use supported formats, including PDF, Word, Excel, CSV, HTML, Markdown, plain text, and common image formats.
- Export Google Docs, Sheets, and Slides to a supported format before adding them.
- For video and audio content, provide a text transcript in a supported format. The media file itself is not searchable.
- Avoid relying on content that exists only inside an iframe.
- Use headings and short sections to structure content. Add alt text or a caption near images that contain important information.
Maintain content over time
Keep the knowledge library accurate as your organization changes.
- Assign an owner to each major topic or SharePoint site.
- After an organizational, policy, or location change, update or retire the related articles and documents.
- Review what employees ask the assistant. Repeated gaps can indicate missing content, an unclear title, or a source that has not been connected or indexed.
- Use conversation feedback to identify incorrect or unhelpful responses. Update the source content rather than relying on corrections made during individual conversations.