Build a bot with the admin UI

This is the path of least resistance to a working ScaiBot. You'll spend about fifteen minutes clicking through the admin UI, and at the end you'll have a deployed bot that answers from your own documents.

You don't need to be a developer. You do need:

A ScaiGrid admin login with permission to create bots.
One or two documents (PDF, Word, Markdown, plain text) you want the bot to answer from.
A page on your website where you can paste a <script> tag — that's the only HTML you'll touch.

If you'd rather drive the same flow from the API in curl or Python, see the advanced tutorials below. Functionally identical; just code instead of clicks.

1. Open the ScaiBot section#

Sign in at https://scaigrid.scailabs.ai/admin. In the left sidebar, expand ScaiBot and click Bots. You'll land on the bot list — empty on a new tenant, otherwise showing what your team has already built.

Click Create Bot in the top-right.

2. Fill in the basics#

The Create Bot dialog has four required fields and a few optional ones:

Field	What to put
Name	A slug for your bot, lowercase with hyphens. Example: `acme-support`. Used in URLs and APIs.
Display Name	The friendly name visitors see. Example: `Acme Support`.
Model	Pick one from the dropdown. `scailabs/poolnoodle-omni` is a strong default — fast, capable, multilingual.
Welcome Message	The first thing visitors see when they open the chat. Keep it short: "Hi! Ask me anything about our product."
Primary Color (optional)	The accent colour for buttons and the launcher bubble. Match your brand.

Click Create. The bot appears in the list with status Active.

3. Upload knowledge#

Click the bot's card to open its detail page. There are tabs along the top: Settings, Knowledge, Tone, Escalation, Playground, Embed, Conversations.

Click Knowledge. By default, every new bot has managed knowledge — meaning this bot has its own private set of documents you upload here.

Drag your document(s) onto the upload area, or click Upload and pick them. Supported types:

PDF — best for product handbooks, manuals, brochures.
DOCX — Word documents; headings are preserved as semantic boundaries.
Markdown — chapter headings (H1, H2, H3) become indexed sections.
HTML — useful for help-centre exports.
Plain text — chunked by paragraph and sentence.

Each document goes through a quick pipeline: text extraction → chunking → embedding → indexing. You'll see the status flip from Uploading → Indexing → Indexed next to each file. Most documents under a few hundred pages finish in well under a minute.

Tip. Quality of answers depends heavily on quality of source. A scanned PDF with no text layer is OCR'd (usually fine), but a corrupted-text PDF will produce muddled chunks. If your bot answers oddly, peek at one or two of its chunks via the document detail panel.

4. Configure the tone#

Click the Tone tab. This is the difference between a generic bot and one that sounds like your team.

The page has a handful of structured controls. You don't write a system prompt — ScaiBot composes one from these fields:

Persona name & description. Sam, a friendly support specialist at Acme. This gives the bot a voice and shapes how it refers to itself.
Formality. Casual / neutral / formal. Casual reads like a coworker; formal reads like a vendor proposal. Pick the one that matches your brand voice.
Verbosity. Terse / concise / verbose. Concise is the default and almost always right. Terse is good if you want one-liners; verbose explains in depth.
Empathy level. Low / medium / high. High is appropriate when visitors might be frustrated (returns, complaints, account issues). Low is appropriate for purely transactional bots (order status lookups).
Language. ISO code — en, de, fr, etc. The bot's primary response language. If a visitor writes in another language, the bot usually follows them unless you tick strict language.
Vocabulary. A list of words the bot prefers and won't paraphrase away. Useful for brand names (Acme, Widget, Acme Cloud) and product-specific terms.
In-scope topics / out-of-scope topics. Two short lists. Product features, pricing, account in-scope; legal advice, competitor comparisons out-of-scope. The bot deflects out-of-scope questions instead of trying to answer.
Custom instructions. A free-text field for anything the structured controls don't cover. Always offer to escalate if the user mentions billing. Keep it short — a paragraph at most.

There's a Preview button at the bottom. It uses your bot's actual knowledge to generate a sample response under the proposed tone. Try a typical visitor question, eyeball the result, tweak the sliders. When it sounds right, click Save.

Tip on tone. If you can't tell the difference between two tone settings, you probably don't need to choose one — leave defaults. The places where tone matters most are: empathy (visitors notice immediately), formality (sets the brand voice), and the custom instructions (where you write the things you actually want the bot to always do).

5. Set up basic escalation (optional, recommended)#

Click Escalation. A bot without escalation rules will try to answer everything; a bot with the right rules knows when to step aside.

Start with two rules:

Rule 1 — Explicit human request. Click Add Rule.

Name: Explicit human request
Trigger: Explicit request to speak to a human
Action: Send email to your support inbox.
Message to visitor: "Got it — let me get a human on this. One moment."

Click Save. This rule catches phrasings like "talk to a person", "real human", "agent please" — across many languages.

Rule 2 — Sustained frustration. Click Add Rule again.

Name: Frustrated visitor
Trigger: Sentiment with threshold Strongly negative over 3 messages.
Action: Send email.
Message: "Let me bring in someone who can help."

This is a safety net — if a visitor's recent messages show sustained frustration, the bot hands off without waiting to be asked.

You can add more later: keyword triggers (e.g. refund, chargeback), intent triggers (e.g. "buying-intent"), or a confidence trigger that fires when the bot is unsure.

6. Test in the playground#

Click Playground. This is a private chat with your bot — exactly what visitors will see, but only you can use it.

Try three kinds of questions:

A question covered by your knowledge. The bot should answer with a citation back to one of the documents you uploaded.
An off-topic question. What's the weather? The bot should deflect — pointing visitors back at the scope you set.
A frustration-test. This is the third time I've asked. Are you broken? The bot should hit your sentiment rule and hand off.

If any of these don't feel right, go back to Tone or Escalation and adjust. Then come back to the playground and try again. Iteration is fast — there's no deploy step.

7. Get the embed code#

Click Embed. The page shows a single <script> tag, pre-filled with your bot's id and a fresh embed token.

html
<script
  src="https://scaigrid.scailabs.ai/static/modules/scaibot/widget.js"
  data-bot-id="bot_..."
  data-token="sbt_..."
  async>
</script>

Click Copy. The token shown here is short-lived (an hour by default) — fine for a quick test, but for a live site you want your server to mint fresh tokens per visitor. That's what the advanced embed tutorial covers. For testing now, the static token is enough.

8. Paste on your site#

Open the page where you want the chat bubble to appear, paste the <script> tag just before the closing </body>, save, and reload. A small chat launcher appears in the bottom-right corner. Click it; chat with your bot.

If it doesn't appear:

Strict content-security policy? Your page may be blocking the script. Open your browser's developer console — there'll be a clear CSP error. See the embed-on-your-site tutorial for the exact CSP rules to add.
Bot id wrong? Double-check the data-bot-id matches what's on the bot's detail page.
Wrong host? The widget script must come from scaigrid.scailabs.ai.

9. Watch the conversations roll in#

Back in the admin UI, click Conversations. Every chat your bot has had is here — visitor identity (if known), full transcript, which knowledge chunks were retrieved for each answer, escalation events.

A few minutes after deploying, this is the most useful page in the product. You'll see what visitors actually ask, where the bot does well, and where it falls back. That feedback is what drives the next iteration on knowledge and tone.

What to do next#

Add more knowledge. A bot is only as good as its documents. Plan a quarterly refresh.
Refine escalation. Look at the conversations log — if visitors are saying things like "this isn't helpful, can I talk to someone", add the matching keyword rule.
Read the concepts. A short browse through Knowledge and RAG and Tone will pay back tenfold in better answers.
Go API. When you want server-side token issuance, programmatic bot management, or webhook-driven escalations, jump to the advanced tutorials — same product, full control.