Knowledge base

FoxChat works without a KB. It works meaningfully better with one.

Your assistant crawls your site and learns from the live DOM, so even with zero docs you get a working AI agent. But the difference between "no KB" and "real KB" shows up in escalation rates, accuracy, and SEO.

FoxChat is white-label: you name the assistant whatever you like and present it as your own. We call ours the FoxChat AI Operator (you name it yourself) by default in our demos — you'll name yours. Throughout this page, "your assistant" means whatever you decide to call it.

Your assistant without a KB

Site-only mode

Your assistant crawls your public pages, extracts headings and content, and answers from what's there. This is the default after install — zero setup.

Works in 3 minutes after install
Answers questions visible on your site
Falls back to "I don't know — let me get a human" when uncertain
Higher escalation rate, more operator load
Doesn't surface refund policy, pricing nuance, or edge cases not on the public site

Your assistant with a KB

Grounded mode

You add structured Q&A pairs (or import from Zendesk / Intercom / Crisp / Help Scout / Freshdesk / CSV). Your assistant uses hybrid retrieval to find verbatim answers from your docs.

Resolution rate jumps 30-50% in our customer data
Pricing, refund, and policy answers are extracted verbatim — never invented
Confidence band stays in the green
Public help docs at getfoxchat.com/docs/<you> with built-in SEO
AI search engines and assistants can now cite your docs

Three ways to get a real KB into FoxChat

Option A

Import from your existing tool

Zendesk, Intercom, Crisp, Help Scout, Freshdesk, or CSV. One click. Categories, tags, and locales preserved.

Import from Zendesk

Option B

Build it in the KB Maker

The "Edit Knowledge Base" tab in your dashboard has an inline editor with categories and a preview pane that shows the article in both widget and public-docs styles.

Open the KB maker

Option C

Let your assistant generate it from your site

The auto-generate flow scans your site, identifies high-value pages, and proposes Q&A pairs. You approve in one click. Best of both worlds.

Auto-generate

What your assistant learns from

Your assistant draws on two pools of knowledge, and understanding the split is the key to getting good answers.

Your live pages, automatically. The moment the snippet is installed, your assistant reads your public site — homepage, pricing, feature pages, docs, FAQs — and answers from what is actually written there. You do not lift a finger for this, and it stays in sync as you update your pages. The catch: it can only answer what your pages say out loud. Anything you know but never wrote down — the real refund window, what counts as one seat, the workaround for a common setup snag — is invisible to it.

The knowledge base you add. This is where you write down the answers that are not (and should not be) plastered across your marketing pages. FAQs, policy details, onboarding gotchas, the "yes, but only if…" answers your support team repeats every week. Each entry is a question-and-answer pair your assistant retrieves and answers from verbatim. This pool is the difference between an assistant that handles the easy questions and one that handles the questions that actually drive people to contact you.

How to add and improve knowledge

Open the Edit Knowledge Base tab in your dashboard. Add an entry as a clear question and the answer you would want a visitor to receive. Group related entries with categories so they stay organized as the base grows, and use the preview pane to see exactly how the answer renders both inside the chat widget and on your public help docs before you publish.

Improving an existing base is mostly editing, not rewriting. When an answer is thin, expand it. When two entries overlap and confuse retrieval, merge them. When a policy changes, update the one entry that owns it — the new answer is live for every visitor on the next save, with no retraining and no waiting.

The fastest way to grow a base is to let the questions write it for you. Your assistant logs the questions it could not confidently answer; each one is a ready-made entry waiting to be filled in. Answer it once, and that gap is closed for everyone who asks next.

Write entries the way visitors actually ask

Your assistant matches a visitor's question against your entries by meaning, not by exact keyword. That means the words you choose decide whether the right answer gets found. Write each entry in the visitor's language — the plain, slightly messy way a real person types — not in internal or marketing phrasing.

Harder to retrieve

"Our platform employs a tiered subscription model designed to scale with organizational requirements."

Easy to retrieve

"Wondering how much it costs? There's a free 14-day trial, then paid plans — want me to show you the pricing page?"

The second version repeats the words people actually type — "how much," "cost," "free trial" — so it gets found when they ask. A few practical habits: title each entry as the question a visitor would type; keep one entry to one topic; spell out the words people search with, including the casual ones; and link to the relevant page rather than baking a price or policy detail into the text, so you only have to update it in one place.

How to spot gaps

A gap is any question your assistant could not answer with confidence, or had to hand to a human. FoxChat surfaces these for you, so you are not guessing where the holes are.

✓Watch the unanswered questions. Each one is a missing entry. Answer the ones that repeat first — they are the questions costing you the most.
✓Watch escalations. A rising handoff rate on a topic means the base is thin there. Fill it in and the handoffs drop.
✓Watch low-confidence answers. If your assistant answers but hedges, the entry exists but is too thin or worded too differently from how visitors ask. Expand it or rephrase it in their words.
✓Close the loop weekly. Fifteen minutes a week answering the top unanswered questions compounds fast. Within a month most of the easy escalations are gone.