By the end of this guide you'll have a working agent that answers from your own content, cites where each answer came from, and runs on your website. Ten minutes is a realistic estimate for a short document — a large PDF or a site crawl takes longer to process, but none of that time is yours to spend watching.
1. Before you start
You need three things, and only one of them takes any effort:
- A Chat.co account. Sign in or sign up at chat.co/auth/login.
- Something for the bot to know. A PDF, Word doc, spreadsheet or Markdown file — or just the URL of a site you want crawled.
- Access to your website's HTML, if you plan to embed the widget at the end. Not needed if you only want the hosted share link.
2. Create your bot
From the dashboard, open the bot selector at the top of the left sidebar and choose Create New Bot. You'll be offered two routes.
Manual Setup
You name the bot, set its chat limit and visibility, and configure everything yourself. This is the path the rest of this guide follows.
AI Builder
Describe what you want in plain language and Chat.co assembles a starting configuration — persona, prompt and suggested sources — that you then refine.
Choosing Manual Setup gives you a short form:
| Field | What it does |
|---|---|
| Bot Name | How the bot appears in your sidebar and dashboard. Visitors see the display name you set later under Appearance, not this one. |
| Chat Limit | Caps how many messages this bot may consume from your account allocation. Useful when you run several bots and want one of them ring-fenced. |
| Public / Private | Private bots are reachable only by you and collaborators you invite. A private bot cannot be embedded publicly — switch it to Public in Settings when you are ready to deploy. |
3. Add knowledge
Your bot exists but knows nothing. Open it, then look at the Knowledge Base group in the sidebar. There are five source types, and they are genuinely different things — not five doors to the same upload box.
Documents
File uploads. PDF, DOC, DOCX, XLS, XLSX, CSV and Markdown, up to ten files at a time.
Links
Crawl a website or a single page. You review the discovered pages and choose which ones to keep before anything is indexed.
Text
Paste raw text directly. Best for policies and snippets that live in someone's head rather than in a file.
Q&A
Explicit question/answer pairs. These take priority over retrieved passages, so use them to pin an exact wording.
Audio
Upload recordings and Chat.co transcribes them into searchable knowledge.
Uploading your first document
- 1
Open Knowledge Base → Documents
With your bot selected, click Documents under Knowledge Base in the sidebar.
- 2
Drop the file in
Drag files onto the dropzone or click to browse. Accepted types are
.pdf,.doc,.docx,.xls,.xlsx,.csvand.md. Ten files per batch; the per-file size ceiling comes from your plan. - 3
Wait for processing
Each file is extracted, chunked and indexed. Short documents finish in well under a minute; long scanned PDFs take longer because the text has to be recognised first. The row shows its state as it goes.
- 4
Confirm it landed
A completed row means the content is retrievable. If a row fails, the troubleshooting guide covers the usual causes — password-protected PDFs and image-only scans being the two most common.
4. Pick a model and write the prompt
Go to Tune Bot in the sidebar, then the Behavior view. Three controls matter here.
| Control | What to do with it |
|---|---|
| AI Model | Chat.co offers models from OpenAI, Anthropic, Google, xAI and several open-weight families. For a first bot, take a current mid-tier model — they are fast, cheap per message, and grounded answers depend far more on your knowledge base than on the model tier. |
| Instruction Prompt | The system prompt: who the bot is, what it refuses, what tone it uses. The single highest-leverage field on this page. There is an Enhance action that rewrites a rough draft into something more specific. |
| Creativity | Temperature. Keep it low for factual support bots — you want the model to stay close to the retrieved text rather than improvise around it. |
5. Customise the widget
Open Appearance in the sidebar. It's a guided builder with a live preview beside it, split into five steps:
| Step | Covers |
|---|---|
| Basics | Display name, welcome message, suggested questions |
| Brand Style | Avatar, colours, which side the launcher sits on, notification sounds |
| Lead Capture | Whether and when the chat asks visitors for contact details |
| Chat Features | Which models visitors may pick, whether sources and file uploads are shown, voice input |
| Branding | Footer branding controls |
Suggested questions are worth more attention than they look. An empty chat box is a hard prompt for a first-time visitor; three good starter questions are the difference between a bot that gets used and one that gets closed.
6. Test before you ship
Use Test Bot to talk to the agent exactly as a visitor would — same retrieval, same model, same prompt. Work through four kinds of question:
- 1
Questions you know are answered
Straight from the document you uploaded. These should come back correct and cited. If they don't, the problem is retrieval, not the model.
- 2
Questions phrased the way customers phrase them
Not the document's wording — your customers'. “Can I get my money back?” rather than “What is the refund policy?”
- 3
Questions with no answer in your content
The bot should decline cleanly. If it invents an answer, tighten the system prompt before you deploy — this is the failure that costs trust.
- 4
Check the citations, not just the text
Open a citation and confirm it points at the passage the claim came from. A correct answer with a wrong citation is a bug you want to find now.
7. Deploy
Open Deployment in the sidebar. Three delivery methods, plus a security tab.
Share Link
A hosted chat page you can send to anyone. No code, no website required.
Website Widget
The floating launcher. One script tag, customisable before you copy it.
Embedded Chat Box
An iframe you place inline in a page, sized to fit your layout.
The widget script
On the Website Widget tab you design the launcher first — position, colour, auto-open behaviour — and the snippet updates as you go. Copy it and paste it before your closing </body> tag:
<script defer src="https://chat.co/api/embed" data-bot-id="YOUR_BOT_ID"></script>
If you customised the launcher, the copied snippet also carries a window.ChatCoEmbed configuration block above the script tag. Paste both lines — the config has to come first.
8. Where to go next
You have a working agent. What you do in the first week after that determines whether it stays useful.
Knowledge upload playbook
Which source type for which content, and how to keep a growing knowledge base clean.
Analytics, ROI & knowledge gaps
Find the questions your bot failed to answer and turn them into content.
Human handoff
Give visitors a route to a person when the bot reaches its limit.
Channel integrations
Put the same agent in Slack, Teams, WhatsApp or your helpdesk.
