Knowledge Best Practices: What to Include and What to Leave Out

Your Agent is only as good as the knowledge you give it. This guide covers what makes knowledge work well, a trick most people miss (telling your Agent what isn't true), and what to keep out. Follow it and your Agent answers more accurately, more often, with fewer "I don't know" moments.

Why Knowledge Quality Decides Everything

Think of your Agent as a seasoned concierge: years of experience, polished, completely at ease with guests. The one catch is that they have no memory of your hotel. They've never worked here, so they know nothing specific about your property: your rooms, your policies, your amenities. What they do have is a book, and the remarkable ability to look anything up in it instantly. For every guest question, they flip to the most relevant pages and answer from what's there, in a heartbeat. Give them a clear, complete, well-organized book and they shine. Give them a thin or messy one and they struggle, no matter how skilled they are.

That's essentially how your Agent works: for each guest question it looks things up in the knowledge you've given it, then answers from what it finds. (This look-it-up-then-answer approach is well established in AI research, where it's called retrieval-augmented generation, or RAG. You don't need the term, just the idea.) Your Agent doesn't browse the web or rely on outside facts, so if something isn't in its knowledge, or it's buried in noise, the answer suffers.

So the quality of your Agent comes down to the quality of your knowledge. Good knowledge in, good answers out. For how the pieces fit together, see Understanding Your Agent's Knowledge: Places and Resources.

What Good Knowledge Looks Like

Keep it accurate and current

Outdated facts are worse than missing ones, because a confident wrong answer erodes trust. Review your knowledge regularly and remove anything that no longer applies. Website and Reviews Resources refresh on their own; Documents and Markdown change only when you update them, so they need a closer eye.

Keep each source focused

One clear topic per source beats a giant page that covers everything. Focused content is easier for your Agent to match to a question, and there's less unrelated text to wade through. When you add a web page, trim the navigation, footers, and banners so only the real content comes through. See Adding a Website Resource.

Have one source of truth

If two sources disagree (an old policy says checkout is noon, a new one says 11 a.m.), your Agent can surface either one. Pick the correct version, update it in one place, and remove the duplicates. Conflicting knowledge is one of the most common reasons an Agent gives an inconsistent answer.

Write it clearly and with structure

Plain language and clear headings help your Agent (and your guests). Short sections with descriptive titles are easier for your Agent to find and quote accurately, like clear chapter headings in a well-organized book. Write the way you'd want the answer to come out.

Cover the questions guests actually ask

Start from real guest questions, not from what's easy to document. Check your Insights for questions the Agent couldn't answer, and fill those gaps first. Coverage of the common questions matters more than volume.

Include "Negative Knowledge": Say What Isn't True

Here's the trick most people miss. It's not enough to document what you do offer. Tell your Agent what you don't.

This is the hardest idea for hotels to accept, and it's worth slowing down on. The natural instinct is to document only what you have: your rooms, your restaurant, your amenities. So if your hotel has no pool, you simply don't mention a pool anywhere. That feels logical, but it leaves a hole. When a guest asks "Do you have a pool?", your Agent has nothing to find, and a blank is not an answer.

The fix is to write the negation down as knowledge. Your concierge can't flip to a page that doesn't exist. If you don't have a pool, that fact, "this hotel does not have a swimming pool," needs to live on a knowledge page, just like your check-in time does. An absence has to be stated, not left out. Treat "what we don't have" as real information your guests need, because they ask about it constantly.

When your knowledge is silent on something, the Agent either says it doesn't know or reaches for a vague guess. State the boundaries plainly and it answers with confidence:

• "We do not have a swimming pool."
• "Pets are not allowed, with the exception of service animals."
• "We do not offer airport shuttle service."
• "Late checkout is not available on weekends."

A simple way to do this: keep a short Markdown Resource called something like "What we don't offer" in your Place, and list the common no's there in plain sentences. It takes five minutes and removes a whole category of weak answers. Add a line whenever a guest asks about something you don't have, so the gap never reappears.

This kind of "negative knowledge" turns hedging, "I'm not sure" replies into clear, correct ones, exactly the moments where guests lose trust fastest.

What to Leave Out

More knowledge isn't always better. Some content adds noise or risk:

• Guest-specific or personal data. Your knowledge is shared across every Conversation, with every guest. Keep individual booking details and personal information out of it. (Personalized answers come from connected guest-data sources, not from shared knowledge.)
• Secrets and credentials. Passwords, API keys, internal-only links, and anything you wouldn't want repeated in a chat.
• Outdated or expired content. Last season's offers, old menus, retired policies. Remove them so they can't resurface.
• Duplicates and contradictions. Multiple versions of the same fact compete with each other. Keep one.
• Filler and marketing fluff. Long promotional copy with little factual content dilutes the useful parts and makes it harder for your Agent to find the real details.

Facts Versus Behavior: Knowledge or Instructions?

A useful rule of thumb:

• Facts go in knowledge (Resources). This includes negative facts like "we have no parking on site."
• Behavior rules go in Instructions. A light touch like "lead with our spa and wellness offer when a guest mentions relaxing" shapes how the Agent acts, not what it knows.

A word of caution on that second point: Custom Instructions are not a script. They run on every single message, so the more you pile in, the less focused the Agent gets. A nudge to highlight a current promotion is fine. Trying to spell out everything you want the Agent to say is not, that belongs in knowledge, where the Agent looks it up only when it's relevant. See How the Agent Works for more on using Custom Instructions sparingly.

Putting each in the right place keeps both clean and predictable.

Keep It Fresh

Knowledge isn't set-and-forget. Build a light habit:

1. Review your Insights for new unanswered questions.
2. Add or correct knowledge to close those gaps.
3. Remove anything that's expired or no longer true.
4. Re-check after big changes (a renovation, a new policy, a seasonal menu).

A Quick Checklist

• Is every fact accurate and current?
• Is each source focused on one topic, with the noise trimmed?
• Is there exactly one version of each fact?
• Have you stated what you don't offer, not just what you do?
• Is personal, secret, or outdated content kept out?
• Are behavior rules in Instructions, not buried in knowledge?

Further Reading

The advice above lines up with what researchers have found makes the look-it-up-then-answer approach (the RAG idea mentioned earlier) work well. If you're curious, these are accessible starting points. You don't need any of this to run a great Agent, your concierge and their book is all the model you need.

• Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv:2005.11401). The original idea: an AI answers better when it first retrieves relevant documents instead of relying on memory.
• Gao et al. (2024), Retrieval-Augmented Generation for Large Language Models: A Survey (arXiv:2312.10997). A broad overview, including how much answer quality depends on the quality of the source content.
• Liu et al. (2024), Lost in the Middle: How Language Models Use Long Contexts (arXiv:2307.03172). Models pay the most attention to the beginning and end of what they're given, which is why concise, focused knowledge beats dumping in everything.
• Barnett et al. (2024), Seven Failure Points When Engineering a Retrieval-Augmented Generation System (arXiv:2401.05856). Common ways these systems go wrong, several of them rooted in missing or messy source content.
• Anthropic (2024), Introducing Contextual Retrieval (anthropic.com). Practical techniques for keeping retrieved pieces meaningful, reinforcing why structure and clarity in your sources matter.

What's Next

• Understanding Your Agent's Knowledge: Places and Resources -- how knowledge is organized
• Creating and Managing Places -- set up the containers for your knowledge
• Adding a Website Resource -- add and focus knowledge from your website