Teaching your agent (Knowledge)

The single biggest predictor of a useful agent is what it knows about your business. Out of the box your agent is fluent, polite, and knows general world facts — but it doesn't know your service area, your prices, your warranty terms, or what your booking process actually looks like. This chapter shows you the four ways to teach it, when to use each, and how to keep its answers accurate over time.

Plain English summary: You're feeding the agent short, focused "knowledge items" — like FAQ answers, policy snippets, or paragraphs from a brochure. The agent stores these in a smart index and pulls the most relevant ones into each answer. The more you add, the better the answer.

Where to find the knowledge tools

Where What it's for
AI Agents → Agents → [agent] → "What the agent knows" tab The agent's primary knowledge file. Add, edit, activate, pause, and delete items here.
AI Agents → Knowledge documents Upload PDFs (brochures, price sheets, policies). We extract the text — including OCR for scanned documents — and turn each section into individual knowledge items you can review before they go live.
AI Agents → Knowledge gaps A live list of questions visitors actually asked that the agent couldn't answer well. One-click "Add as knowledge" turns each gap into a new answer in the agent's file.
Agent's "Websites" tab → Run crawl (Growth+) Point the agent at your website. We fetch your pages, extract the meaningful content, and offer each page as a knowledge-item draft for you to review.

Method 1 — Type or paste an answer directly

Best for: FAQ-style answers, policy snippets, short canned responses.

  1. Open the agent's edit page.
  2. Click the What the agent knows tab.
  3. Click Add knowledge (top right).
  4. Fill in:
    • Title — short, e.g. "Cancellation policy" or "Service area". This is how you'll find it later.
    • Topic — pick from the list (pricing, scheduling, policies, etc.) or leave blank. Helps you keep things tidy.
    • One-line summary (optional) — a one-sentence version of the answer. Helpful when the body is long.
    • What the agent should know — the actual answer, in plain English. Imagine you're writing a quick note to a new employee who'll be answering this question. Short paragraphs work better than long blocks of text.
    • Make available to the agent — leave on Live to let the agent use it immediately, or pick Draft to save without activating yet.
  5. Click Create.

Rule of thumb: Each knowledge item should answer one question in 2–8 sentences. If you find yourself writing more than that, split it into multiple items — the agent retrieves them separately, so smaller, focused items get pulled more accurately.

Method 2 — Upload a PDF document

Best for: brochures, price lists, employee handbooks, policy documents, intake packets — anything you already have as a PDF.

  1. Go to AI Agents → Knowledge documents.
  2. Click Upload document.
  3. Pick the agent the document is for and choose the PDF file.
  4. Click Upload.

We do the rest in the background:

  • Text extraction — we read every page and turn it into searchable text.
  • OCR fallback — if the PDF is a scan (image-only), we run optical character recognition to extract the words. Look for the OCRed count on the row.
  • Table detection — price lists and structured tables are detected separately so they don't get jumbled into paragraphs.
  • Section splitting — long documents are broken into individual knowledge items so each section can be retrieved on its own.

When processing is done, the row shows Status: Ready with an Items: N count.

Click the document row to review every extracted item. Each one is created as a draft — the agent will not use it until you activate it. You can:

  • Activate all draft items (top-right button) — turn them all live at once after a quick skim.
  • Discard all draft items — if the extraction came out garbled (rare; usually only on heavily-formatted PDFs), you can clear them and try a different file.
  • Edit individual items — tweak the title or content before activating one at a time.

Why do they start as drafts? Because PDF extraction is fast but not perfect. Reviewing the drafts protects you from the agent confidently quoting a stale price or out-of-context paragraph.

What kinds of PDFs work best?

Works great Works OK Best to avoid
Text-based PDFs (exported from Word, Google Docs) Scanned documents — OCR handles them but expect minor typos Highly-designed brochures with text inside images
Price sheets, FAQs, policy docs, service descriptions Multi-column newsletters Forms with checkboxes and signature lines
Onboarding packets, employee handbooks Documents in low-resolution scans Photos of whiteboards or business cards

Method 3 — Crawl your website (Growth+)

Best for: getting the agent up to speed on everything your site already explains — services, about us, FAQ pages, blog posts, etc.

  1. Open the agent's edit page.
  2. Click the Websites tab.
  3. If you haven't added a site yet, add one (just your homepage URL).
  4. Click Run crawl next to the site.

We'll fetch up to 20 pages on Starter, 100 on Growth, and 500 on Pro in one crawl, respecting your robots.txt and skipping obvious junk (login pages, cart pages, etc.).

When the crawl finishes you'll see a notification. Open the crawl results to see every page we found — each one is a draft knowledge item with the page's URL, title, and the meaningful text we extracted.

For each page you can:

  • Activate — turn that page into a live knowledge item the agent can quote.
  • Discard — skip it (useful for paginated archive pages, "All blog posts" indexes, etc.).
  • Merge — combine related pages into one item (useful when one product spans several pages).

You don't have to activate everything. Start with your top 10–20 pages (Home, About, Services, Pricing, FAQ, Contact). Add more later.

Re-crawling: When your site changes meaningfully, run another crawl. We diff against the last crawl so you only see what's new or updated — you won't have to re-review pages that haven't changed.

Method 4 — Fill knowledge gaps automatically

Best for: catching the questions you didn't know visitors were asking.

The agent quietly tracks every question it gets asked. When it gives a weak answer — or no answer — that question is logged as a knowledge gap. You'll see them under AI Agents → Knowledge gaps.

Each row shows:

  • The question (verbatim from the visitor).
  • How many times it's been asked.
  • Which agent was asked, and on which page.
  • An AI draft column — for many gaps, we'll have a suggested answer ready for you to review.

For each gap you can:

  1. Click Generate draft — we use the agent's existing knowledge + general world knowledge to draft a candidate answer. The draft is not added to the agent's knowledge until you save it.
  2. Click Use suggested answer to accept the draft (often with a small edit) and add it as a new live knowledge item.
  3. Click Add as knowledge to write your own answer instead.
  4. Click Ignore if the question is out of scope (e.g., "What's the weather?") — we'll stop showing it.
  5. Click Reopen if you change your mind about an ignored or resolved gap.

Make this a weekly habit. Five minutes in the Knowledge gaps page is one of the highest-leverage things you can do — every gap you close is an answer the agent will give correctly from now on.

The Knowledge gaps menu item shows a badge with the number of open gaps so you don't forget.

Knowledge item states (cheat sheet)

State What it means Where to set
Live The agent can use this item right now. Most items live here. Set during create or via the row's edit form.
Draft Saved but not visible to the agent yet. Use when you want to write something, walk away, and come back. Same as above.
Paused Was live, now temporarily hidden. Use when an item is out of date but you want to fix it later instead of deleting. Same as above.
Processing We're indexing the item so the agent can search for it. Usually takes seconds. Automatic.
Indexing failed Something went wrong during indexing. Click Refresh content on the row to retry. Automatic — click to retry.

You can filter by state at the top of the What the agent knows table.

Tips that actually move the needle

  1. Write like you're briefing a new hire. Short sentences. Concrete examples. Avoid marketing copy — the agent doesn't need to be sold on you, it needs to be informed.
  2. One question per item. "What are your hours?" and "Where are you located?" are two items, not one.
  3. Cover the boring stuff first. Hours, location, service area, contact methods, pricing brackets, cancellation policy, refund policy. Visitors ask these constantly.
  4. Update price knowledge whenever pricing changes. A stale price item is worse than no price item.
  5. Date-stamp anything time-sensitive. Add the date inside the answer (e.g., "As of June 2026...") so the agent knows whether to lean on it.
  6. Check Knowledge gaps weekly. Five minutes; biggest accuracy win you can get.

Privacy & where your knowledge is stored

  • Knowledge items are stored on your tenant only — no other tenant or customer can see them.
  • The text is sent to our embedding provider (Upstash Vector + the LLM provider) so the agent can search by meaning, not just keywords.
  • You can delete any item at any time. Deletion removes it from the agent immediately and from the search index within minutes.
  • We never sell your knowledge data or use it to train shared models.

For the full data flow, see the engineering data model. For deletion guarantees, email privacy@coffield.io.

Next steps

Last updated
Need more help? Email support@coffield.io

Your experience on this site will be improved by allowing cookies.