Knowledge Base Best Practices

Your AI Agent is only as good as the knowledge you give it. This article covers how the different knowledge types work together and the most common mistakes to avoid.

How the Knowledge Types Work Together

Your knowledge base has four types of content, and each plays a different role:

Type	Role	Always Active?	Max Length
Guidelines	Behavioral rules the agent follows in every conversation	Yes	500 chars
Verified Answers	Exact responses to specific customer questions	Only when a question matches	5,000 chars
Facts	Authoritative statements the agent references when relevant	Only when relevant	2,000 chars
Knowledge Sources	Documents the agent reads and draws from	Only when relevant	Full documents

When a customer sends a message, the AI considers all of these together. Guidelines shape its overall behavior. If the question matches a Verified Answer, that takes priority. Otherwise, the AI draws from relevant Facts and Knowledge Sources to form its response.

The practical hierarchy:

Guidelines always apply, regardless of the question
A Verified Answer with "Bypass LLM" enabled returns your exact wording, no questions asked
Facts are treated as authoritative when they're relevant, and take priority over conflicting document content
Knowledge Sources provide the broad foundation the AI draws from for everything else

Choosing the Right Type

Scenario	Best Type	Why
"We never discuss competitors"	Guideline	Always-on behavioral rule
"What's your return policy?" needs a precise answer	Verified Answer	You want to control the exact wording
You switched shipping carriers last week	Fact	Quick correction without re-uploading a document
Your full FAQ page	Knowledge Source	Broad document the AI reads for multiple topics
"Always mention our loyalty program to hesitant buyers"	Guideline	Behavioral rule across all conversations
"We don't ship to PO boxes"	Fact	Specific detail, not a full document
"Do you accept PayPal?" needs a one-word answer	Verified Answer	Simple, exact answer you want word-for-word

Monitoring Your Knowledge Base

The Content Health section at the top of the Knowledge Base tab gives you a snapshot of your knowledge base. Expand it to see:

Active Sources and how many are processing or failed
Active Facts and Verified Answers counts
Guidelines (active and pending review)

Use these counts as a quick health check. If you see failed sources, check that the link or file is still accessible. If you see pending review guidelines, review and approve or dismiss them.

Building a Strong Foundation

If you're starting from scratch, build your knowledge base in this order:

Review auto-populated content: Approve pending Shopify Pages, check that your policies were ingested correctly
Add 5-10 Verified Answers for your most common customer questions
Add 3-5 Facts for important details that aren't in your documents
Add 3-5 Guidelines for behavioral rules your agent should always follow
Upload additional documents as needed (product guides, FAQ pages, sizing charts)
Test in the Playground after each batch of additions

You don't need to have everything perfect before going live. Start with the essentials and add more as you see real conversations.

Keeping Content Fresh

Stale content is worse than missing content. If the AI confidently gives an outdated answer, customers lose trust.

What to watch for:

Seasonal information (holiday shipping deadlines, sales, limited-time offers) should be archived or updated when it expires
Facts created during onboarding should be reviewed after you've added full Knowledge Sources covering the same topics
Auto-synced sources (Google Docs, External Pages) re-sync daily, but check that the sync is still working if the source URL or sharing settings change

When policies change:

Update the Shopify policy page (auto-synced sources pick this up automatically)
Update or archive any Facts that referenced the old policy
Update any Verified Answers that quoted specific details from the old policy
Test in the Playground to confirm the AI gives the new answer

Common Mistakes

Duplicating information across types If your return policy exists as a Shopify Policy source AND you've written a Verified Answer AND you've added a Fact about return windows, the AI has three sources to reconcile. Keep one authoritative source per topic. Use Facts for corrections and Verified Answers for questions where exact wording matters.

Too many vague guidelines "Be helpful," "Know our products," and "Give good answers" don't add anything. The AI already does these things based on your persona and tone settings. Focus guidelines on specific behaviors you need enforced.

Adding too much at once without testing After adding or changing knowledge, test in the Playground before relying on it in live conversations. A fact that seems clear to you might interact with existing content in unexpected ways.

Forgetting about temporary content Holiday promotions, limited-time offers, and seasonal shipping changes should be archived when they expire. A fact saying "Order by December 18 for Christmas delivery" is actively harmful in January.

Missing your core policies Make sure your knowledge base covers the topics customers ask about most: shipping, returns/refunds, contact information, and payment methods. If any of these are missing, those are your highest-priority additions.