Best Practices in Knowledge Management

This guide outlines the best practices and operational workflows for establishing a robust Knowledge Management system. Following these standards ensures that your AI-driven assistant provides accurate, relevant, and secure information to your employees.

---

1. Content Strategy & Selection

To maintain a high-quality knowledge base, focus on high-value content while eliminating "noise" that can confuse the assistant.

Docs: Best practices for structuring 'Knowledge Articles'

Scope & Format

• Target content: Focus on departmental knowledge such as HR and IT. Avoid syncing redundant information or departments that are not needed for your use case.
• Supported formats: PDF, Word (.docx), PowerPoint (.ppt/.pptx), HTML, and spreadsheets (.xlsx/.csv). Spreadsheets are fully supported and, when well structured, can even be queried for calculations and lookups. Native pages from connected sources (such as SharePoint and Confluence) and Google Docs/Sheets are also supported.
• Not supported today: Audio and video files — such as recordings, webinars, and training videos — and Outlook email files (.msg) cannot be read or answered from, and other non-text media is not indexed for search. If important knowledge exists only in a recording, add a written summary or transcript as the article so the content becomes searchable. Save an email you need as a PDF or Word document.
• Handle with care: Some content is harder to interpret and should be prepared thoughtfully rather than avoided outright:
  • Images that carry meaning (charts, diagrams, screenshots) are supported and searchable, but they work best when they are clear and accompanied by a short caption. Avoid using an image purely to display text — type the text out instead.
  • Spreadsheets are supported, but only well-structured sheets can be used for calculations. See the structuring guide for how to lay these out.
  • Scanned or handwritten documents can be read but far less reliably than digitally created files , especially for special characters (like non-English alphabets) or icons (like checkbox). So, please avoid relying on them for important content.
  • Redundant or duplicate content adds noise and should be removed.
• File size and length: Keep individual files under 200 MB and 500 pages (while these are hard limits, recommendation would be to stay within 100 MB and 100 pages). Files over 200 MB are skipped during sync. Content past page 500 is not read, and no error is shown. Word and PowerPoint files are converted to PDF before they are read, so pages are counted after conversion (each slide is one page). More importantly, favour focused, single-topic articles over large multi-topic files: split a big manual (e.g. a full HR handbook) into per-topic articles so answers retrieve cleanly.

For detailed, article-level guidance on all of the above, see Best practices for structuring 'Knowledge Articles'.

Documentation Standards

Pro tip: Always include external links to official policies or source sites. This allows employees to verify information and provides additional context for complex queries.

Check the Parsed Content Before Publishing

Every article shows both the original file and the Parsed content — the text the assistant actually reads and answers from. The two are not always the same, and it is the parsed version that determines the answers your employees receive.

For any article that carries real weight — a policy, a compliance document, a procedure people will follow — spend a minute comparing the two before publishing:

• Is anything missing, particularly text that sits inside an image or a diagram?
• Is the reading order correct, especially in scans, forms and multi-column pages?
• Are field values still alongside their labels?
• Has any hidden or boilerplate text been picked up from a web source?

A fluent-sounding parse is not necessarily a complete one. Both reading-order problems and summarised images produce text that looks perfectly reasonable on its own — comparing against the original is the only way to see the gap.

---

2. Technical Configuration: Metadata & Permissions

Ensuring the right people see the right information at the right time.

Permissions Management

Security is paramount. You can use base system permissions or platform-converted permissions from metadata.

• Action required: Always ensure the "Fetch permissions" toggle is enabled when adding a new connector to maintain data privacy. Article audiences then follow the access rules in the source system, which stays the one place where access is managed.
• Keep the permission model simple: Permissions are fetched for each document separately. Unique per-file permissions across a large library slow the sync and raise the chance it fails, so set access at site, library or folder level where you can.
• Check after permission changes: A change in the source takes effect in Leena AI after the next sync. After a folder-level change, confirm it reached every file in the folder, including files added later.
• Spot-check restricted content: After each major sync, open a few articles from restricted folders (for example, R&D-only) and check that their audience matches the source.

Docs: SharePoint, ServiceNow, etc.

Audience Creation from Metadata (available for Sharepoint)

In cases where permissions are not managed properly in your source knowledge base (e.g. SharePoint), it is recommended to create an audience using metadata.

The Dynamic Audience Mapping feature introduces an Access Control step within the Connector Setup Wizard, enabling administrators to govern document visibility by mapping user attributes, such as Department or Country, directly to document metadata.

• Metadata tagging: Specific audience identifiers are assigned to documents.
• Dynamic filtering: Search results are filtered at the time of a query based on the user's current profile.
• Real-time updates: Access updates (e.g. an employee changing departments) take effect without requiring a full re-sync of the knowledge base.

Docs: Dynamic Audience from Connector Metadata

---

3. Quality Assurance & Testing

Before going live, use these built-in tools to audit your knowledge base for accuracy and redundancy.

Conflict Analysis

Identify duplicates or stale information by selecting folders (up to 1,000 articles).

1. Run the analysis to find conflicting information.
2. Correct the source document at the connector level (e.g. in SharePoint).
3. Re-check the conflict status to ensure a clean deployment.

Docs: Conflict Analysis

Smart Testing (Automated Audit)

Smart Testing is an automated tool that measures assistant accuracy before launch.

• The process: The system generates question-and-answer pairs from your uploaded documents, tests them against the assistant, and provides an accuracy score (0–100%).
• Evaluation criteria: Content accuracy, completeness, and relevance.
• Limits: Up to 100 articles per run; 1,000 cumulative test cases per day.

Docs: Smart Testing

---

4. UAT to Production Workflow

Use the User Acceptance Testing (UAT) environment for incremental testing of new content within your knowledge base. This applies to all integrations (e.g. SharePoint, ServiceNow) by utilising a secondary account with the same credentials. The KM dashboard allows you to manage up to five such instances for incremental data testing.

Docs: Managing Multiple Instances of a Connector

Details about the different environments in Knowledge Management can be found here: Environments in KM.

Here is an example of such a workflow for a SharePoint integration.

The Workflow

1. Connect: Add a new site (e.g. a specific site rather than the whole tenant) in UAT.
2. Test: Validate the content and assistant responses thoroughly.
3. Delete from UAT: Delete the test site from UAT, then delete the UAT connection.
4. Deploy: Add the same site to Production. You have two options, described below.

Deployment Options

Rolling Out Content

• Roll out in phases: Start with a small, high-value scope, check the answers, then expand. Several smaller syncs are more reliable than one large migration. After the first sync, scheduled syncs only reprocess content that changed.
• One sync at a time: A new sync can't start while another sync for the same connector is running or queued. Wait for it to finish before starting the next one.
• Sandbox (trial) limit: On sandbox bots, a sync is blocked if the selected scope holds more than 5,000 articles. This applies to SharePoint, Google Drive, Confluence, ServiceNow and SFTP.

---

5. Ongoing Maintenance & Governance

Answer quality depends on content that stays current. Treat these as standing habits, not a one-time setup.

• Keep the source system as the single source of truth: Make edits in SharePoint, Confluence or the other source, and let the sync carry them into Leena AI. Avoid keeping copies of the same document in different places.
• Assign a content owner: Give each site, space or library a named owner who keeps its structure tidy and its content current.
• Review content on a schedule: Work with subject-matter experts to retire outdated or duplicate documents. Stale content dilutes answers even when a correct document exists. Run Conflict Analysis first to find duplicates and contradictions.
• Let real questions decide what to add: In Leena AI Analytics, add the Not handled Request Dump table to a custom dashboard to see the questions the assistant couldn't answer. Add or fix content for those topics first, then re-run Smart Testing on the articles you change.

---