Knowledge base
Give AI agents your policies, FAQs, and product information so they answer correctly.
An AI agent without a knowledge base can only answer from what it can look up in Shopify — order status, product details, stock. It can't tell a customer your return window, because that's your policy, not Shopify data.
The knowledge base is where you put everything else.
Source types
| Type | Use for |
|---|---|
| FAQ | A question and its answer. The most precise option. |
| Text | Pasted content — policies, guides, tone instructions. |
| URL | A page Enscale fetches and indexes. |
| File | An uploaded document. |
Processing status
| Status | Meaning |
|---|---|
| Pending | Queued. |
| Processing | Being indexed. |
| Ready | Available to agents. |
| Failed | Couldn't be processed. |
A source only affects answers once it's Ready. If an agent gives a wrong answer right after you add a policy, check the status before assuming the agent ignored it.
Failures are usually an unreachable URL, a page requiring JavaScript to render, or an unsupported file format. Paste the content as Text instead.
What to add first
Start with what customers actually ask, in this order:
- Shipping — how long, where you ship, what it costs.
- Returns and exchanges — the window, the condition, who pays return postage.
- Sizing and fit — if you sell anything worn.
- Payment methods — including COD, if you offer it.
- Order changes — whether an address can be changed after ordering.
- Product care — washing, storage, warranty.
Six good FAQ entries beat a whole website indexed badly.
Writing sources that work
Be specific and quantified. "Returns accepted within 30 days of delivery, unworn with tags attached" is answerable. "We have a flexible returns policy" is not — the agent will either hedge or invent something.
One topic per FAQ entry. An entry covering shipping, returns, and sizing gets retrieved for all three and answers none of them well.
Write how customers ask. "How long does delivery take?" retrieves better than "Fulfilment SLA".
State the exceptions. If sale items aren't returnable, say so in the returns entry. An agent that confidently states the general rule and misses the exception creates a support ticket and an unhappy customer.
Keep it current. A knowledge base describing last season's shipping times is worse than no knowledge base — the agent will state the outdated policy with full confidence.
URL sources
Indexing your policy pages seems efficient, and often is. Two caveats:
- Pages that render content with JavaScript may index as empty.
- A page mixing policy with navigation, promos, and footer text indexes all of it, which dilutes retrieval.
For anything important, paste the specific policy text as a Text source rather than pointing at the page it lives on.
Knowledge base vs. agent instructions
They do different jobs.
| Knowledge base | Agent instructions | |
|---|---|---|
| Holds | Facts about your business | How the agent should behave |
| Example | "Returns accepted within 30 days" | "Be concise. Never promise a refund." |
| Retrieved | When relevant to the question | Always |
Facts go in the knowledge base. Tone, boundaries, and escalation rules go in the agent's instructions. Putting facts in instructions bloats every request; putting behaviour rules in the knowledge base means they only apply when retrieved, which is unpredictable.
Testing
After adding sources, ask the agent the questions you added them for — including awkward phrasings and edge cases.
Worth checking:
- The exception cases, not just the general rule.
- Questions your knowledge base does not cover. The agent should hand over, not guess. See Handover.
- Anything involving money or promises. Those should escalate.
Troubleshooting
| Problem | What to check |
|---|---|
| Agent doesn't use a source | Is it Ready? Does its wording match how customers ask? |
| Agent gives outdated answers | An older source still says the old thing. Remove it. |
| Answers are vague | The source is vague. Add specifics and numbers. |
| URL source failed | Page needs JavaScript, or blocks fetching. Paste as Text. |
| Agent contradicts itself | Two sources disagree. Find and remove the stale one. |
Next steps
- AI agents — configuring agents.
- Handover — when agents escalate.
- AI usage and spend — what agents cost.