Introduction
Chat Data Cookbook
Build an AI agent that answers from your own content, then put it on your website and messaging apps.
Create your AI agent
Add your files, website and Q&A, then train your AI agent on them.
Improve its answers
Set the base prompt, model and temperature so answers stay accurate.
Put it on your website
Allowlist your domain, copy the embed code and paste it. Guides for Shopify, WordPress, Wix and more.
Connect messaging apps
Start with WhatsApp. Messenger, Instagram, Telegram, LINE, Slack and Discord guides are in the sidebar.
Quick start
Follow these six steps in order. Most people have a working AI agent on their website within the first hour.
1. Create your account
Sign up with an email address and password, or with your Google account. Details: Create Chat Data Account.
2. Create your AI agent
Log in, open Agents and click New Chatbot. 
Pick one of four backends. Most people pick Train on your data, the default:
| Backend | Shown in the app as | What it does |
|---|---|---|
custom-data-upload | Train on your data | Trains the AI agent on your own data (the default) |
custom-model | Custom backend | Uses your own LLM backend endpoint |
medical-chat-human | Medical Chat · Human | Uses our human medical model |
medical-chat-vet | Medical Chat · Vet | Uses our veterinary medical model |
Not sure which one? See Train With Your Data and Using Medical Chat Models.
3. Add sources and wait for training
Add your files, website pages, text or Q&A under Sources. See Train With Your Data for every source type.
Wait for training to finish
Training usually takes about 5 minutes. If you test too soon, the AI agent hasn't learned all your data yet and its answers may be incomplete.
Structured data such as Q&A gets more weight than unstructured text. If the two disagree, the AI agent follows the structured data. See Structured Data V.S. Unstructured Data.
4. Configure the AI agent
Open Settings > Model and set these three things:
| Setting | What it does | Recommendation |
|---|---|---|
| Base prompt | Sets the tone and rules for every reply | Set this first. Keep the logic clear and short. With custom-data-upload, write CONTEXT INFORMATION to refer to your uploaded data. |
| Model | The AI model that writes the answers | Stronger models follow complex prompts better and hallucinate less, but are slower and cost more. See Basic Configuration for the options. |
| Temperature | How predictable or creative replies are | Keep it at 0 when answers must stick closely to your data |
Base prompt or training data? Put rules that apply to every conversation in the base prompt. Put answers to specific questions in your training data.
| Put it in | Use it for | Example |
|---|---|---|
| Base prompt | Rules for every conversation, whatever the user asks | "At the end of every conversation, ask about the user's medical history." / "Handle sensitive information carefully and stay polite." |
| Training data | Answers to specific questions | Q: "Can I order a blood test?" A: "Yes, you can order a blood test through our online portal. Here's how…" |
5. Test and fix answers
- Ask your AI agent questions in the Playground. To check a bad answer from your chat logs, paste the same question.
- Open the sources of the answer to see which data it used.
- If your data doesn't cover the question, or the right information was lost when your data was split into chunks, add a Q&A with the correct answer.
Step-by-step guide: Debug And Optimization.
| Common cause of poor answers | Fix |
|---|---|
The base prompt says something like given info instead of CONTEXT INFORMATION | Use CONTEXT INFORMATION. Without it, the AI agent ignores the context information we send from your data. |
| Too much long, unstructured text that contradicts itself | Add structured data such as Q&A. See Structured Data V.S. Unstructured Data. |
| Testing before training finished | Wait about 5 minutes, then test again. |
| The answer is nearly right but skips a condition in the prompt, or summarizes your Q&A instead of using it | Choose a stronger model in Settings > Model, especially for complex prompts. See Basic Configuration. |
6. Put it on your website
Allowlist your domain, copy the embed code and paste it on your site. Start with Embed Chatbots On Websites. It also links to the guides for Shopify, WordPress, Wix, Squarespace, Webflow, Framer, Bubble and Weebly.
Other places your AI agent can work
| Where | Guide |
|---|---|
| WhatsApp Integration | |
| Slack | Slack App Guide |
| Discord | Discord Integration Guide |
| Shopify store | Add Chatbots to Shopify Stores |
| WooCommerce product updates in real time | WooCommerce Guide |
More channels (Messenger, Instagram, Telegram, LINE and others) are listed in the sidebar.
What is Chat Data?
Watch this 3-minute video for a quick overview.
Have 10 minutes? This video walks through all the features.
How live-chat handoff works
Most businesses use the AI agent as a 24/7 support rep on their website. It answers most questions on its own (say 80%). For the rest, such as questions that need the very latest information or refunds that need special permission, a teammate takes over through live chat.
- The customer chats with the AI agent on your website.
- To reach a person, the customer clicks the live chat escalation button.
- Your teammate (the human agent) gets an email with a unique link for that conversation. The link opens the chat without logging in to Chat Data.
- Customer still in the widget: the teammate and the customer chat in real time.
- Customer has left: they get an email with a link to the new message. This works if they left contact details in the lead form. When they reply, your teammate gets an email.
All emails and links can be white-labeled with your own branding. Teammates can also reply from your Chat Data admin panel. Setup: Live Chat Escalation.
What you can do with Chat Data
| Area | Features |
|---|---|
| Automate tasks | AI Actions: web search, buttons, lead collection, sending files and images, and calling your own API endpoints in chat. Workflows built from natural-language prompts and triggered as AI actions. Action Items for appointments, article links and lead forms. Generate PDF documents from the conversation and a PDF template. |
| Talk with customers | Live chat escalation on websites and other platforms, with the AI deciding when to hand off. SMS notifications for escalations (import your number or buy SMS credits). Dynamic suggested questions based on the conversation and your data. AI re-engagement follow-ups on WhatsApp, Messenger and Instagram. |
| Voice | Real-time voice chat with ElevenLabs voices (519 voices across 152 locales). Customized voice replies that sound human. |
| Channels | Chat bubble and AI search bar widgets for websites. WhatsApp through the official API or WhatsApp Web (no business account needed, instant setup). Discord, Slack, Telegram, Messenger, Instagram, WordPress, Shopify, WooCommerce, Zendesk, Zapier, Make and PrestaShop. Image and file chat on the website widget, WhatsApp, Messenger and Instagram, for both the AI agent and live chat. |
| Data and models | Crawl websites protected by Cloudflare or IP rate limits through proxies. Power the agent with your own backend endpoint. Nested chatbots that link several agents into one. |
| Your brand and your users | White-labeling under your own domain, company name, logo and support email, with SSO login so users manage agents without registering with Chat Data. Logged-in user experience: personalize answers from user profiles, test results and history through our SDK. |
| Billing and analytics | Usage metering by Stripe to charge your customers. Public metrics dashboard links. |
| Privacy | Opt out of chat history access. Built-in PII anonymization: personal information is replaced with placeholders before it is sent to the AI model, then restored in the reply. This protects what the model sees; it does not anonymize stored conversation records (see HIPAA Compliance). |
Getting help
- Setup help: Schedule a meeting with the Chat Data team.
- Priority support: Professional plan customers can get our WhatsApp number for priority support. See VIP Support.
- Enterprise plan: larger, more powerful agents, isolated servers and database instances, and a management site with your company logo. Email [email protected].
- Docs feedback: if something here is missing or wrong, email [email protected].
Frequently Asked Questions
What can I learn from the Chat Data Cookbook homepage?
The homepage introduces Chat Data as a platform for building AI chatbots with your own data, links to the main documentation tracks for creation, training, advanced features, and integrations, and summarizes the broader product workflow from automated responses to human handoff.
What major capabilities does Chat Data highlight on the homepage?
The homepage highlights features such as AI-powered workflows, live chat escalation, real-time voice mode, logged-in user personalization, white-labeling with SSO, SMS notifications, multimodal inputs, AI actions, dynamic suggested questions, Stripe usage metering, PDF generation, custom backends, nested chatbots, privacy controls, and multi-platform integrations.
How does the Chat Data support workflow work according to the homepage?
The homepage explains that customers first chat with the AI agent on your site, the bot resolves most requests automatically, and edge cases can be escalated to a human agent through live chat. Human agents receive a whitelabeled handling link by email, can reply in real time, and customers can also be notified by email if they have already left the widget.
What setup guidance does the homepage give for improving chatbot accuracy?
The homepage recommends using the base prompt for universal instructions and training data for query-specific knowledge, then testing conversations, checking response sources, and updating structured Q&A when the existing data is incomplete or poorly segmented.