The Temira handbook

Getting started

From zero to a live assistant in four moves. First-timers get a guided Setup wizard — a five-step walkthrough (profile → knowledge → branding → AI engine → launch) with a progress bar that you can finish or skip and return to.

1 · Create your workspace

Sign up with your business name and type (products / services / both). You get a 14-day trial of everything and a portal at /portal. Your assistant lives at a link like https://your-platform/t/your-business/

2 · Teach it

In Portal → Knowledge, add sources — 16+ formats across four families. Each becomes searchable knowledge with its origin tracked, and the assistant cites it in answers. See Teaching your assistant for the full list.

3 · Brand it & pick the brain

Appearance: assistant name, logo, accent colour, welcome line, starter questions, answer language. Voice & persona: writing style, tone and speaking voice. AI engine: choose a provider and model — Claude Fable 5 is the default recommendation, and failover is automatic. See Appearance, Voice & persona and The AI engine.

4 · Go live

Paste one line before </body> on any website:

Options: data-voice="1" opens in voice mode; data-label="Ask us" changes the launcher text; data-copilot="1" lets the assistant read the host page and spotlight buttons during walkthroughs (for in-app help); data-open="0" suppresses the proactive pop-up on a specific page. No website? Share the direct link on WhatsApp, Instagram or your email signature — or connect a messaging channel so the assistant answers customers right inside their app (see Answer on every channel).

Proactive pop-up

In Appearance → Proactive pop-up, switch the teaser on, choose a delay and (optionally) your own message. A small branded bubble greets each visitor after the delay — once per visitor per day. Clicking it opens the chat; the ✕ dismisses it quietly.

Embedding in your own app

The same one-liner works inside a web app for your staff or signed-in customers. Add data-copilot="1" and the assistant can see the current screen (title, headings, visible buttons, on-screen errors) and spotlight the exact element to click during step-by-step help. For native mobile apps, load your assistant's direct link in a WebView — it's an installable, standalone-capable page.

Teaching your assistant

In Portal → Knowledge, add sources — 16+ formats across four families. Each becomes searchable knowledge with its origin tracked, and the assistant cites it in answers:

• Web & feeds — a single page (paste a link), a whole-site crawl or a sitemap.xml (each reads up to your plan's page limit — 50 on Free, up to 10,000 on Enterprise; big sites are read in slices with live progress), or an RSS/Atom feed that keeps pulling new posts as you publish.
• Documents — PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), CSV, plain text, Markdown (.md) and HTML files. Drag-and-drop or browse. A scanned or image-only PDF — one that's really a photo of a page with no selectable text — is read automatically by AI vision (OCR, up to 8 pages), so a PDF that is really a picture still becomes searchable text.
• Images — screenshots, scanned price lists, flyers: read by AI vision. Drop many at once and they upload in parallel as one grouped image set (up to 60).
• Audio & video — voice notes and recordings, transcribed (configure a speech provider under Voice & persona). Paste any YouTube link — watch?v=, youtu.be, a Short, an embed link or just the video id, with or without https:// — see below.
• Q&A & FAQ — type facts directly, or paste Q:/A: pairs and each becomes its own verified answer. These outrank every other source. Short is fine — “Q: What time do you open? A: 7am, every day.” is a perfectly good source.

Processing runs in the background; the knowledge base rebuilds itself (including a local semantic index — meaning-level search with zero external calls) when the queue drains. As each source is read it's split along its own headings rather than at blind character counts, so a passage stays on one topic; tables are pulled out as structured rows instead of collapsing into a wall of numbers; and every passage is enriched with keywords across languages, so a question asked in one language can still surface content written in another.

Adding a YouTube video

Paste the link and the thumbnail appears as soon as we recognise it, so a wrong paste is caught before you press Add. We take the video's title, channel and description, and its transcript when YouTube gives us one.

To add the spoken content, paste the transcript — it always works and takes about ten seconds. On YouTube, open the video's description (…more under the title), scroll to its end → Show transcript → click the first line, Shift-click the last, copy. Timestamps are fine — they're stripped automatically. Then paste it under Knowledge → YouTube, or press ＋ Transcript on a video you already added. You can also upload the audio or video file and have it transcribed instead.

Managing what it knows

The Knowledge base panel shows live counts (passages, sources, pinned answers), whether the semantic index is on, and when it last built — with a Rebuild now button. The Sources table lists every source with its status, passage count and date, and a Retry (on error) or Delete control. Test what the assistant will find lets you type a question and see the exact passages it would retrieve — see also the retrieval & quality tools.

Beyond the basics, each source carries its own controls:

• Health score — Temira quietly asks each source a handful of questions drawn from its own content and grades how well it can answer them, giving every source an answerability rating. A low score points you straight at a source that reads fine to a human but is thin to a machine — and the questions it failed are shown right there.
• “What was read” — open a website source to see its coverage: every page that was captured, and every one that wasn't, each with the reason (excluded by robots.txt, over the plan limit, a redirect, an error). Nothing is a black box.
• Boost — nudge one source up or down in the ranking when you want, say, your official policy page to win ties over an old blog post.
• Cross-source de-duplication — when the same passage turns up in two sources (a PDF and the web page it came from), Temira collapses the duplicate, so answers don't repeat themselves and one topic doesn't crowd out the rest.

Answer-only sources, and rolling back

Not everything you teach the assistant should be downloadable. An uploaded file is private by default: its text answers customers, but the file itself is never handed back or linked — so an internal price sheet can inform a quote without ever leaving your workspace. You can set this per source with a lock, or turn on a whole-workspace privacy setting so every source is answer-only unless you say otherwise.

And before any big change — a bulk delete, a full re-crawl, a privacy flip — Temira can take a snapshot of your knowledge base. If a change goes wrong, one-click rollback restores the last good state, so experimenting is never a one-way door.

It answers from your knowledge — or it says it doesn't know

Your assistant is grounded: it answers only from the knowledge you gave it, the results of actions it actually ran, and what the customer told it in the conversation. Its general knowledge of the world is not a source — it will not describe what your business does based on your industry, your name, or what similar companies usually offer.

This is enforced in the platform, not just requested of the model. Before any answer is written, Temira checks whether your knowledge base genuinely covers the question. If it doesn't, the AI is never asked to answer — the customer gets an honest "I don't have that" together with an offer to talk to a person or open a support ticket. A brand-new assistant with an empty knowledge base cannot invent a description of your business, because there is no path by which it could.

Big websites, page limits & freshness

Website and sitemap crawls read your site's own published sitemap first (so no page is missed), then follow links for anything unlisted — up to your plan's page limit: 50 pages on Free, 300 on Starter, 1,500 on Growth, 5,000 on Scale, 10,000 on Enterprise (the Add-knowledge form shows your current limit). Big or slow sites are read in slices: the source's detail line shows live progress ("Reading your website — 212 pages so far…") and the crawl continues where it left off until the site is fully read or your limit is reached. If your site is bigger than your plan reads, the source says exactly how many pages were left — nothing is ever silently skipped — with a one-click upgrade to read it all.

On paid plans your website sources are re-read automatically so answers stay current — monthly on Starter, weekly on Growth, daily on Scale — with no action needed from you.

A re-read never leaves you worse off. While a source is being refreshed the assistant keeps answering from the last good version, so there's no window where it goes quiet. If a re-read comes back with far less content than before — a sign your site was mid-deploy, or a page template broke — Temira holds the shrunken result for your review instead of publishing it, and you either accept it or retry. And if many pages simply vanish between reads, you get a change alert by email so you can check whether that was intentional.

We read your site at your server's pace

Temira crawls politely — a couple of pages at a time with a pause between each. If your server pushes back (many hosts have anti-flood rules that reject rapid visitors), Temira slows down automatically, waits, and comes back for that page later instead of giving up on it, then speeds back up once your server is comfortable. It also honours your robots.txt, and tells you how many pages that excluded so you can change it if you'd rather they were included. The practical promise: the only thing that leaves a page unread is your plan's page limit — never a refused request.

You can also scope a crawl to one section of a site — point it at example.com/help and it stays under that path instead of wandering the whole domain, which keeps a support centre's knowledge clear of the marketing site. And single-page-app sites that once produced a storm of near-identical placeholder pages (every route returning the same empty loading shell before JavaScript fills it in) are now detected and skipped, so your passage count reflects real content, not scaffolding.

Watching a crawl, and pausing at your plan limit

While a site is being read, the Knowledge page shows a live progress card: which stage it's in, pages captured, links still queued, any pages waiting to be retried, and how much of your plan's page allowance is used — with the title of the page it just read, so you can see it moving.

If your site is bigger than your plan reads, the crawl pauses instead of failing. Everything captured is saved and your assistant answers from it right away, and the card tells you exactly how many pages are still waiting. After you upgrade, press Continue crawling and it resumes from precisely where it stopped — nothing already read is fetched again, and you never delete and re-add the source.

When a page tries to give your assistant orders

Anything read from the open web is treated as untrusted. Someone who can get text onto a page you crawl — a product review, a forum post, a blog comment — could otherwise write straight into your assistant's instructions. Temira neutralises the parts a computer would obey (command lines, control tags) as the content is read, and keeps the rest exactly as written.

Text that merely reads like an instruction is reported, not deleted: the source shows a ⚠ N flagged pill and the detail view names what was found, so you can look at the page and decide. A legitimate page is allowed to discuss the topic — these very docs do. Measured on a real 1,262-passage knowledge base, this changed 0 passages: the protection is invisible to genuine documentation.

More detail, including what it does and does not protect against, is under Privacy, data & security.

Appearance & branding

In Portal → Appearance, shape how the widget looks — with a live preview beside every change:

• Identity — assistant name, accent colour (a primary and a darker shade are derived for you), a logo you upload (png/jpg/svg/webp), a welcome banner and a spoken greeting (use the {name} token).
• Proactive pop-up — the teaser bubble toggle, its delay and your own message.
• Answer language — Auto-detect or one of 17 fixed languages (see Languages).
• Starter questions — up to six one-tap prompts that greet a new visitor.
• Embed — the copy-ready one-line script and your assistant's direct link.
• Show the knowledge line under the chat — the line reading “Your Business · 425 knowledge passages · Answers come from the business's own published information”. Off by default: most customers only want the answer, and a passage count means nothing to them — with it off the assistant reads as part of your own product. Turn it on when you want the grounding claim visible.

On paid plans you can remove the “Powered by Temira” badge; Scale and Enterprise add white-label (and, on Enterprise, your own domain).

Voice & persona

In Portal → Voice & persona, give the assistant a personality and, if you want, a real speaking voice.

Persona

• Writing style — a preset (Professional, Friendly, Playful, Formal, Expert, Caring, or Custom), plus a free-text tone in your own words.
• Answer length — Brief, Balanced or Detailed. Formality — Casual, Neutral or Formal.
• Emoji — allow tasteful emoji, or keep it clean.
• House-style rules — a free-text box for the things it must always (or never) say.
• Speaking rate — a slider from 0.6× to 1.4×, plus “speak brief acknowledgements” and an optional wake word (“Hey {name}”).

Bring your own speech engines

Voice is powered by your own (or the built-in) speech providers, configured here and proxied server-side so keys never reach a browser:

• Text-to-speech — an OpenAI-compatible endpoint, key, model and voice, with ▶ Hear a test sentence.
• Speech-to-text — a Whisper-compatible endpoint, key and model.
• Realtime voice — for full-duplex spoken conversations: a session URL, key, model and voice.

A live Voice setup checklist at the top of the page confirms both halves are in place — the assistant must be able to speak (text-to-speech) and hear (speech-to-text). It updates as you type and flags common slips (for example a api.openapi.com endpoint that should read api.openai.com), so a visitor never gets a bot that can talk but can't listen.

The AI engine

In Portal → AI engine, choose which model answers — and how it's metered.

• Providers — ten engines are supported: Anthropic (Claude), OpenAI, Google Gemini, DeepSeek, Groq, Mistral, xAI Grok, GitHub Models, OpenRouter and a local Ollama tier. Each row takes your API key (masked), a model dropdown, ▶ Test (shows latency and pass/fail) and Use to make it active. Paste your own key for unlimited answers billed by the vendor, or leave it blank and use the built-in engine metered by your plan.
• Behaviour — Automatic failover (if one engine is down mid-conversation, the next in line answers, and the real engine that answered is recorded), use the built-in engine when you have no key, an answer cache (identical questions served free), and a daily AI-answer cap (0 = unlimited).
• Usage (last 90 days) — a day-by-day table of provider, model, answers (with the built-in count) and tokens, so cost is never a surprise.
• Local model tier — point at an Ollama-compatible endpoint and Run certification; once a model passes a graded ≥90% exam against your knowledge, the assistant can serve from it first.

How it picks the evidence

Retrieval is tuned to hand the model varied, on-point evidence rather than five near-identical snippets: a diversity pass trades a little raw similarity for breadth, so an answer draws on the different things your knowledge actually says. Search is typo- and meaning-tolerant — the local semantic index matches on meaning, so a misspelt or loosely-worded question still finds the right passage.

If a question still comes back ungrounded, Temira doesn't give up on the first try: it rephrases the question into keywords and retries once before saying it doesn't know. An answer found on that second pass is marked cautious so you know it took the rescue path — and if there's genuinely nothing to stand on, it still refuses rather than inventing.

Semantic search uses the built-in local index by default. If you'd rather, you can bring your own embeddings — point at an OpenAI-compatible key — but it's off by default and never required.

What the assistant can do

Toggle these under Portal → Actions & catalog. Every action is validated server-side and rate-limited; a failed action is reported honestly.

• Price quotes — exact arithmetic from your portal price list. It refuses to quote if the list is empty.
• Support tickets — asks for the visitor's email, opens a ticket with a reference like APX-T1042, alerts your team.
• Callbacks — captures name + phone as a lead.
• Bookings — appointment requests with preferred time (service businesses).
• Order capture — items + quantities recorded with a reference; priced automatically when the items match your catalog.
• In-chat payment — priced orders come back with a secure hosted payment link from your own gateway — 14 built in worldwide (Paystack, Flutterwave, Stripe, PayPal and more) plus a Custom option (see Payments).
• Store search & store orders — with a connected store (Shopify, WooCommerce, BigCommerce, Magento or a custom API): live product search and real orders with your store's own checkout link.
• Order tracking — "where is my order APX-O87?" answered from your workspace or live from your connected store.
• Account & record lookups — optional, read-only: connect any of the 41 integrations (CRM, ERP, helpdesk, calendar, spreadsheet…) and the bot answers "where is my invoice INV-1042?", "is my subscription active?", "what's my ticket status?" from live data — consulting every connected system at once.

The Alerts section here also sets where each lead/ticket/booking is sent — an email address and/or a webhook (POST JSON to your CRM or Zapier).

Payments in the chat

In Portal → Actions → Payments in the chat, connect your own gateway and paste its key(s). Fourteen gateways are built in, spanning Africa, Asia, the Americas and Europe:

• Africa — Paystack, Flutterwave, Monnify, CinetPay.
• Global cards & wallets — Stripe, PayPal, Square, Razorpay, Xendit, Tap, Mercado Pago, Mollie.
• Crypto — Coinbase Commerce.
• Custom — wire any other provider that issues a hosted payment link over a REST API (this is how, e.g., Espees connects until a first-class option exists). Existing Paystack/Stripe workspaces keep working unchanged.

From then on, any order with a confirmed total (computed from your price list or your connected store — never by the AI) returns a hosted payment link the assistant shares in the conversation. Customers pay on the gateway's own secure page; settlement goes directly to your gateway account; the order flips to paid automatically on confirmation. For instant server-side confirmation, add the webhook shown in the portal to your gateway dashboard. You can also generate or copy a payment link for any order manually from the Inbox.

Connected systems (connectors)

In Portal → Actions → Connected systems, plug the assistant into what you already run, so it answers your customers from live data — "where's my order / booking / ticket / invoice?", "is my subscription active?", "when's my appointment?". Pick a system, fill a couple of fields, and press ▶ Test connection before saving.

41 ready-made integrations

The Type dropdown offers 41 named integrations across 10 categories — each a one-fill recipe. Selecting one shows exactly which fields to enter, a short description, a docs link and what the assistant gains:

• Commerce & inventory — Shopify, WooCommerce, BigCommerce, Square, Magento 2.
• Booking & scheduling — Calendly, Acuity, Cal.com.
• CRM & customers — HubSpot, Pipedrive, Salesforce, Zoho CRM, Stripe, Intercom, ActiveCampaign, Brevo.
• ERP & finance — ERPRevolution (first-class: its own signed API-v2 connector, the picker's first row and default), ERPNext, Odoo, QuickBooks Online, Xero, Sage Accounting, Zoho Books.
• Support & helpdesk — Zendesk, Freshdesk, Zoho Desk, Gorgias, Front.
• Payments & transactions — Paystack (customers), Flutterwave (transactions).
• Data & spreadsheets — Airtable, Notion, Google Sheets.
• Project & productivity — Asana, Trello, ClickUp.
• Social & media — Facebook Page, Instagram Business, YouTube, X (Twitter) — read your own posts / media / videos.
• Microsoft & Google — Microsoft 365 (Graph: Outlook, Bookings, SharePoint, OneDrive…), Google Calendar, Google Contacts, Google Drive.

Each recipe pre-sets the authentication — API key, bearer, basic, or a full OAuth 2 login (refresh-token or client-credentials) for the enterprise systems — so you only supply your own values. Every preset is a starting point you confirm with ▶ Test connection against your live account. OAuth systems (Salesforce, Zoho, QuickBooks, Xero, Sage, Microsoft 365, Google) need a one-time app registration on the vendor's side; the recipe handles everything after.

Or build your own

For anything not on the list, the original building blocks are still there:

• Webhook — one endpoint on your side receives {"op": "search | order_status | create_order | lookup", …} and your code decides everything. Optional bearer secret.
• Custom REST API — fits any documented API: describe each operation's method, path and body (with {query}/{ref} placeholders) and pick the auth — bearer, API-key header, basic, query key, or session login: provide a credential and the platform signs in the way a person would (posts to your login URL, extracts the token or session cookie), then calls the API on the visitor's behalf. Every named integration above is a ready-made recipe over this same engine.

Connect step-by-step — every gateway & system

Voice, vision & tours in the chat

What a visitor experiences in the widget goes well beyond typing:

• Voice mode — hands-free, two ways. Turn-based voice uses the browser's speech recognition and either the browser voice or your premium TTS; optional wake-word listening lets a visitor just start talking. Realtime voice is a full-duplex WebRTC speech-to-speech conversation with barge-in. Open straight into voice with data-voice="1".
• Vision — visitors can attach, drag-drop or paste an image (a photo of a product, a receipt, a screenshot) and the assistant reads it. It also accepts PDFs and text files inline.
• Screenshot tours & guided walkthroughs — the assistant can show, not just tell: step-by-step flows with a progress bar, a screenshot gallery with lightbox zoom, and — inside your own app with data-copilot="1" — live spotlighting of the exact button to click.
• Answer enhancers — Go deeper re-answers exhaustively; Cross-checked runs a multi-agent “council” (planner → researchers → composer → verifier) for high-stakes questions; follow-up chips keep the conversation moving.
• Feedback — a thumbs up/down on any answer, and a one-time satisfaction rating after the conversation warms up. Both feed your Learning & Intelligence tools.
• Memory & install — the assistant keeps a rolling case file for returning visitors, and the chat page is an installable PWA with an offline service worker.

Answer on every channel

Your assistant isn't only the widget on your website. From Portal → Channels it can answer your customers on four messaging apps too — WhatsApp, Instagram DMs, Facebook Messenger and Telegram — using the same knowledge base, brand voice and actions it uses on your site.

Which channels your plan includes: Telegram is on every plan, including Free; WhatsApp joins from Starter; Instagram and Messenger from Growth. Channels outside your plan appear as locked cards with an upgrade button — and anything you connected earlier keeps working and stays editable whatever your plan does later: a downgrade never disconnects a channel you already set up. (Trials run with every Growth feature, channels included.)

Pick a mode per channel

• AI + human recommended — the assistant handles everyday questions and hands off to a real person when the customer asks for one, sounds frustrated, or asks something it isn't confident about.
• AI only — the assistant answers every message itself.
• Human only — the assistant doesn't reply; your team is simply notified so a person can pick it up.

Where you answer a handoff

When a conversation comes to a human, you reply from the same Live chat page you already use for website chats — or from your phone via the Telegram / WhatsApp relay — and your reply goes straight back to the customer inside their own app. Live takeover works exactly as it does on the website.

What it costs

Answering customers is free. Instagram and Messenger have no per-message fee, and WhatsApp replies are free within the 24-hour customer-service window — i.e. whenever you're replying to someone who messaged you first, which covers everyday support and sales. The only paid case is optional WhatsApp marketing broadcasts, which you send only if you choose to.

To message a customer outside the 24-hour window (order updates, reminders), WhatsApp requires a pre-approved message template. Create and track templates from Portal → Channels → WhatsApp → Message templates — Temira submits them to Meta for approval and shows each template's status.

Connecting a channel

• Telegram — works instantly: create a bot with @BotFather, paste the token, and it's live, no approval needed.
• WhatsApp, Instagram & Messenger — connect through Meta with one-click Connect with Meta (Embedded Signup) where the platform operator has set up the Meta app: it detects your WhatsApp Business Account and phone number, wires up webhooks and registers the number automatically — no IDs to copy. Or connect by pasting your own access token. Because these run on Meta's platform, they need Meta's one-time business verification / app review before they go live — a quick setup you do once.
• Connecting manually? Two optional fields make the connection stronger: your WhatsApp Business Account ID (WABA) unlocks the message-template manager above, and your Meta app's App secret lets Temira verify that every incoming webhook genuinely comes from Meta (signature checking) — recommended.

Getting found & sharing

An assistant that answers 24/7 only helps where customers can actually reach it. The portal's Grow section puts your link everywhere that matters, one page per job: Your link (the short link, QR and reach score), Link hub (your link-in-bio page), Get the word out (every placement with ready-to-paste wording and a tick when it's done) and For your website (the embed snippet).

Thirteen places are listed, and twelve of them need no approval from anyone — you can do them today, in an afternoon. The thirteenth is answering natively inside WhatsApp, Instagram and Messenger, which needs Meta's one-time review (see Answer on every channel). It's listed honestly rather than hidden.

Your link

• Short link — temira.net/yourname, short enough to say out loud.
• Every link is tagged — the copy on each surface carries its own ?src= tag, so your dashboard can tell you which placement actually brings people in. Paste them as they are.
• QR code — download and put it on anything physical.
• Preview card — paste your link into WhatsApp, LinkedIn or X and it appears as a branded card with your name, tagline and colour instead of a blank rectangle.
• Copy in ten languages — English, Nigerian Pidgin, Yoruba, Igbo, Hausa, Kiswahili, French, Portuguese, Arabic and Spanish.

The thirteen places

Grouped as Owned (your website, Google Business Profile, email signature, invoices and receipts), WhatsApp (greeting message, away message, profile website field), Instagram (bio link, instant reply), Facebook (Page button), Offline (printed QR) and Native (Telegram, which connects instantly — and the Meta channels).

Link hub — for the one link Instagram, TikTok and WhatsApp let you have

Switch on the hub and your short link opens a page you build, instead of going straight to chat. Chat always leads and can't be removed — that is the point: every other link-in-bio page lists links, yours answers the question.

Blocks you can add, in any order: any link (with a thumbnail and a highlight) · WhatsApp, call, text, email and directions (filled in from your settings) · a row of social icons · section titles, announcements and banner images · a YouTube or Vimeo video that only loads when it's pressed · a product from your catalogue with its price · a capture form that lands in your Inbox · your top questions answered from your own knowledge · a live Open now / opens tomorrow at 08:00 from your business hours · an ask box they can type into right there · Save our contact, which puts you in their phone book carrying your assistant's link · a countdown · a payment link.

Make it yours: eight themes, five button styles, three corner styles, four typefaces, one-per-row or two-across, your own accent colour and a cover image. Schedule a block to appear on a date and retire on another, so a promotion publishes and pulls itself down without you being awake for either.

See what people actually press. Page views, taps, the share of visitors who tapped something, and a ranking of every block. That is the number that tells you whether your bio link is working — and it's the one thing link-in-bio tools charge for.

It's rendered on the server, so it loads on a bad connection, inside the Instagram and WhatsApp in-app browsers, and with JavaScript off. Search engines get proper business information — name, phone, address, opening hours, socials — so you can appear as a rich result rather than a blue link.

Put it in your website or app

Share → Put it in your website or app writes the snippet for you and gives you a test page to try it on before you touch your real site. Five ways to place it:

• Floating button — the classic chat bubble in the corner.
• My own button — no bubble at all. Your existing Help button opens the assistant. This is the one for an app that already has its own support button:
  <script src="…/embed.js" data-launcher="none" data-trigger="#help-button" defer></script>
• Inside a panel — renders in a container you choose, for a help drawer or support tab.
• Side drawer and Full screen — opened by your own button.

You can also drive it from your own code — Temira.open(), Temira.ask("Where is my order?"), Temira.setContext({plan:"pro"}) and events for open, close and lead. See Developer API.

Cards worth posting

Your own results as an image you can post — questions answered, how much was covered by your own knowledge, and answers written without a source: zero — carrying a link anyone can check. There's a report card (1200×630, for a post, an email or your site) and a Status card (1080×1920, for WhatsApp Status or Stories).

They stay hidden until your assistant has actually answered enough questions to have something worth showing. Sharing “0 questions answered” would do you no favours, so the page says so plainly rather than handing you an empty card.

Print pack

Eight printable assets at real trim sizes with 3 mm bleed — window decal, table tent, shelf talker, delivery-box sticker, staff badge card, vehicle decal, roll-up banner panel and business-card back. Each opens a print dialog; choose Save as PDF for a printer or print it yourself. The delivery-box sticker is the quiet favourite — it arrives in the customer's home and keeps answering questions long after the sale.

Where your conversations come from

Because every share link is tagged, Portal → Dashboard shows a breakdown of which placement each conversation came from over the last 30 days. Put the link in five places, see which two are doing the work.

While you were closed

A Night Shift summary on your dashboard counts what happened outside your opening hours — questions answered, people who left their details, and answers written without a source — with a running lifetime total. It is the plainest measure of what the assistant is worth: the business that was open when you weren't.

Reply on WhatsApp

When someone leaves their details overnight, open Portal → Inbox and tap Reply on WhatsApp. Your own WhatsApp opens, addressed to that customer, with a draft written from your knowledge base that you can edit before sending. No approval, no per-message fee, and it comes from the number they already know.

Public directory

You can list your business on temira.net/discover, a public page of businesses customers can simply ask a question. It is opt-in and off by default — we don't list a business publicly without being asked.

Once you're listed, people can search the directory, and you appear on the industry page and the location page for your business — /discover/in/bakery, /discover/near/lagos — which are the pages people actually search for. You don't have to switch the link hub on to be listed.

Your own domain

On request you can point a subdomain you own — chat.yourbrand.com — at your assistant with a CNAME, so the whole experience sits on your brand.

Languages

The assistant detects the visitor's language and replies in it automatically — spoken answers included — or you can lock a single language under Appearance. Seventeen are supported: English, French, Spanish, Portuguese, German, Italian, Dutch, Russian, Turkish, Hindi, Chinese, Arabic, Swahili, Hausa, Yoruba, Igbo and isiZulu.

Your command center

Portal → Dashboard is a live, self-refreshing view of your assistant at work — tiles count up as new events land, and figures flash as they change.

• KPI tiles — conversations handled, questions answered, leads captured, deflection rate, satisfaction and more. Click a tile for a plain-English explainer of exactly how it's computed, and drill through to the underlying records.
• Conversation activity — a 14-day chart you can switch between Questions, Chats and Leads, with hover detail.
• This month's AI usage — a meter of built-in answers used against your plan's allowance (the day-by-day, per-engine breakdown lives under AI engine).
• Top questions and Knowledge gaps — what visitors ask most, and what your content couldn't answer (with a jump straight to teach these).

Every data table across the portal filters, paginates (10–100 rows) and exports the filtered view to CSV.

Inbox

Portal → Inbox is everything the assistant captured, in five tabs:

• Leads — name, company, contact (email/phone), note and time.
• Tickets — reference, subject, who from, a status pill, and Close / Reopen.
• Bookings — reference, who, service, preferred time, status, Mark done / Reopen.
• Orders — reference (and any store reference), who, items, total, a status dropdown (new → confirmed → paid → processing → shipped → delivered / cancelled) and a Payment column: “paid ✓”, Copy pay link, or generate a payment link (set the total, send the link).
• Ratings — the star rating and comment a visitor left, with the time.

Conversations

Portal → Conversations lists the real ones — anything with at least one message, newest first. Pick a conversation to read the transcript, tap 🧠 AI summary for a one-click recap, or 🗑 Delete to erase it permanently (see Data & security for retention controls).

“Opened the assistant without asking anything”

A conversation is created the moment someone opens your assistant — before they type. Most of those people read the welcome screen and leave, and listing them next to real transcripts buried the ones that matter. They are counted separately and sit behind one button at the bottom of the list. Nothing is deleted or hidden from you; it's ordered so the transcripts come first. That count is worth watching on its own — a lot of opens and few questions usually means your welcome screen isn't inviting the question.

Answers survive a customer walking away

Someone who closes the tab, locks their phone or switches app while the answer is still arriving used to leave you a question with no reply beside it. The server now stores the answer in that situation, so what you read is what they were actually told.

Learning & Intelligence

Two tabs turn everyday conversations into a better assistant.

Learning

• Teach your assistant — every real question your content couldn't answer, listed with an answer box. Write the answer (or have the AI draft it) and it becomes a pinned, authoritative answer.
• Answering itself — a 30-day stacked chart of instant (cache/pinned) vs AI-generated answers, plus a summary of knowledge on file and a live feed of what it has learned.
• What your assistant saved you — conversations it handled end to end without a human, and (once you enter what one support contact costs you) the money that avoided.
• Your public trust report — a page you can publish at /t/<your-slug>/trust and link from your own website: answers written from your published knowledge, times the assistant declined rather than guessed, and answers written without a source — which is structurally zero, because with nothing relevant retrieved the AI is never called at all. Counts come from a log written at the moment of each decision, and the page shows totals only — no question, answer or visitor detail. It is off until you switch it on, comes with an embeddable badge, and stays silent about percentages until there are enough questions for one to mean anything.

Intelligence & the Quality Lab

• Self-test — grade the assistant against your own knowledge with an AI judge. Turn on nightly automatic self-test and show a “✓ Self-tested” trust badge on the widget when it fully passes.
• Daily digest — an optional email summarising questions, leads and unanswered questions (needs a verified alert email).
• Weekly knowledge-gaps email — separate from the daily digest: a once-a-week roundup of the questions your assistant couldn't answer, so filling the gaps becomes a regular habit rather than a fire drill.
• Groundedness audit — audit the last ten answers for drift, or review every answer a visitor thumbed down.
• Model arena — type one question and compare every configured model's answer and latency side by side.
• Improvement queue — unanswered questions clustered by meaning and ranked by volume × failure, so one pinned fix resolves a whole cluster. Each source's health score feeds in too, so a weak source surfaces here alongside the very questions it's failing.
• Visitor journeys — AI-maintained case files for returning visitors, so context carries across sessions.

Live human takeover & your team

When a visitor asks for a person — or you simply want to jump in — a human can take the wheel and the AI stands down until they leave. Growth+

Live chat

• Join a queued conversation (the badge in your portal lights up), reply, and End or 🔀 Transfer it to a teammate mid-chat.
• 📞 Call the visitor by voice over WebRTC — with a ring state, a live timer and an automatic recording saved when you hang up. The visitor gets an in-chat accept card, and their browser tab flashes “📞 … is calling” so the ring lands even if they're reading another tab. If nobody has answered after 45 seconds the console says so — your typed messages still reach them either way.
• 🎫 Make ticket mid-chat turns the conversation into a tracked Service-desk ticket without leaving the room: the customer sees a note with the ticket reference right in the conversation, the full transcript travels with the ticket, and you stay in the chat to wrap up properly.
• 🔔 Desktop alerts — browser notifications (with a sound fallback) when a visitor needs a human.
• Ending well — when a chat ends, the visitor sees a warm goodbye in your business's name (never internal bookkeeping), the AI resumes seamlessly, and they're invited to rate the human handoff (CSAT).

Answer from your phone

Enable the phone relay to take live chats on the go: connect a Telegram bot (token from @BotFather + your chat ID) or the WhatsApp Cloud API (access token, phone number ID, your number and a webhook verify token). Send a test message to confirm it before you rely on it.

Agent desk

Portal → Agent desk runs your human team alongside the AI, with seats metered by plan:

• Team — add teammates (name, email, skills), set each one 🟢 Online / 🟡 Away / ⚪ Offline, and enable, disable or remove them. The seat meter shows used vs available. Each teammate can also carry a client-facing alias — the name and title customers see in the chat (“Ada from Support”) — so your team shows a consistent front without exposing real names or internal handles.
• Performance — over 7/30/90 days: chats, average first-reply and resolve times, CSAT, a quality score, how much each taught, and reopens.
• Training — capture teammates' live answers for review, then Approve & pin the good ones so the AI learns them.
• Routing — “when the topic mentions … send to …” skill rules with priorities and a fallback.
• Leaderboard — an opt-in, quality-weighted points board with medals and badges.

Service desk

Portal → Service desk is the whole support-ticket lifecycle, built natively into Temira — not a bolt-on you buy separately. It's the only desk where the AI, the channels and the ticket are one system: nothing gets re-explained, every reply reaches the customer where they already are, and every resolution makes the assistant a little smarter. It's the support desk designed to put itself out of business.

How a ticket gets opened

• The AI opens it — when the assistant can't resolve something on the widget or any channel, its create-ticket action files a ticket automatically.
• One-click from a live chat — during a handoff, hit Make ticket and the new ticket is pre-filled with the full transcript and the customer's identity, so the customer never repeats themselves. The customer is told in the chat that the ticket exists (with its reference), and you stay in the conversation to finish up.
• From a customer channel — a message on WhatsApp, Instagram, Messenger or Telegram can become a ticket, tied to that same conversation.
• By hand — open one yourself in the portal.

What every ticket carries

Each ticket gets a public reference (e.g. APE-T1043) and travels a clear lifecycle — open → pending (you're working it) → waiting (on the customer) → resolved → closed, and it can be reopened at any time. On the ticket you'll find:

• Assignee — the teammate who owns it.
• Priority & SLA clock — low / normal / high / urgent, each with an elapsed-time target (urgent 4h, high 8h, normal 24h, low 72h) shown as on-track, due-soon or overdue.
• Category & tags — for filtering and routing.
• AI one-line summary — the gist at a glance.
• Threaded conversation — customer-visible replies and internal private notes, side by side.
• CSAT — a 1–5 satisfaction rating captured on resolution.

Replies reach the customer where they started

A public reply is delivered back on the same channel the ticket came in on — WhatsApp, Telegram, Instagram or Facebook Messenger — or by email if you've configured email. As with Temira's channels, answering customers on those messaging apps carries no per-message fee.

Where the AI helps (and where it never oversteps)

• Auto-triage on creation — writes the one-line summary, adds up to four tags and suggests a priority.
• AI draft reply — grounds a suggested reply in your knowledge base and the ticket thread for a human to review and send. It never sends on its own and never invents policy.
• Turns resolutions into knowledge — when you resolve a ticket, Temira condenses the thread into one suggested knowledge-base answer, filed as pending in the existing Agent desk → Training review queue. Approve it in one click and it becomes a permanent answer the bot uses — so the same question is deflected next time and ticket volume trends down.

CSAT & reopening

On resolve, the customer is invited to rate the ticket 1–5 on their own channel. A bare 1–5 reply is captured as the rating; a real follow-up message after resolution simply reopens the ticket so nothing falls through the cracks.

What you can see

The desk surfaces open tickets, overdue, average first-response time, average resolution time, CSAT average, and “taught the AI” — how many knowledge-base drafts came out of tickets. Open ticket load also shows up on the dashboard's Open work tile.

Beyond the basics, the desk gives agents a Customer history (360°) panel — every other ticket, order, booking and lead tied to the same person, so returning customers never re-explain themselves — plus Insert from knowledge, AI draft reply and Reply & resolve. SLA & business hours let you set resolve-by targets per priority and (optionally) count the clock only during your open hours; tickets auto-assign by topic, a daily reminder flags anything overdue or going stale, and an Analytics view charts volume, SLA attainment, category/priority mix, CSAT and how many answers your tickets taught the AI. Your customers can track and rate their own ticket at /t/<your-workspace>/ticket with just their reference and email.

Team & access (staff logins & roles)

Give your staff their own logins and scope exactly what each can do. In Portal → Team & access, invite live agents, support staff, content editors or analysts — each signs in at the same portal with their own email and password and sees only what their role allows. It's true role-based access control, enforced on the server (the menu just hides what a role can't use — the server refuses it either way).

Roles

Pick a ready-made role or hand-pick permissions:

• Administrator — full access (a manager; still not the owner, and can be removed).
• Support lead — conversations, live chat, inbox, service desk, dashboard, knowledge, intelligence and the team.
• Support agent — conversations, live chat, inbox, service desk and dashboard.
• Content editor — knowledge, appearance, voice, intelligence and dashboard.
• Analyst — the dashboard, read-only.
• Custom — any mix of the 14 permissions (dashboard, conversations, live chat, inbox, service desk, knowledge, intelligence, appearance, voice, AI engine, actions/payments/connectors, team, billing, business settings).

How access is protected

• Every request is checked against the signed-in person's role — including writes, so a staffer can't reach a restricted area by crafting the call. New areas are owner-only until explicitly opened to a permission.
• Secrets stay hidden — payment, connector and channel keys, and provider API keys, are stripped from the portal for anyone whose role doesn't manage them.
• No escalation — a manager can't grant a permission they don't hold themselves, or edit someone with more access than they have.
• Instant revocation — set a teammate to Disabled or remove them and they lose access to the workspace right away.

Plan, billing & usage

Portal → Plan & billing shows your current plan with a usage ring (built-in answers used this month), your trial / active / free status, days left and renewal date. Start free, pay a flat monthly fee, and pay yearly to get 12 months for the price of 10. The 14-day trial gives you Growth-level features with no card, and you can cancel anytime — the workspace falls back to Free with your data intact.

	Free	Starter	Growth	Scale	Enterprise
Built-in AI answers / mo	50	500	2,500	10,000	Unlimited
Unlimited with your own key	✓	✓	✓	✓	✓
Pages read per website	50	300	1,500	5,000	10,000
Channels	Telegram	+ WhatsApp	+ Instagram & Messenger	All	All
Automatic site re-read	—	Monthly	Weekly	Daily	Daily
Knowledge sources	25	200	1,000	Unlimited	Unlimited
Actions, leads inbox & analytics	✓	✓	✓	✓	✓
Voice conversations	—	✓	✓	✓	✓
Live human takeover	—	—	✓	✓	✓
Remove “Powered by” badge	—	—	✓	✓	✓
White-label / own domain	—	—	—	White-label	+ own domain

See Pricing for current prices. When you upgrade, Choose how to pay: card in USD via Stripe, card / bank / USSD in Naira via Paystack, or a manual bank transfer (we show the reference and amount). Your payment history is listed with reference, plan term, amount, gateway, status and date.

Settings & privacy

Portal → Settings holds your business profile and data controls.

• Business profile — name, tagline, “what do you do?”, business type, currency, industry, country, region, and your website / contact / pricing pages. These sharpen the assistant's answers and pick the right actions for your kind of business.
• Privacy & data — toggle transcript storage, the lead-capture card and voice conversations, and set a data-retention window in days.
• Contact support — send Temira a message (subject + body) straight from the portal.

Sign-in security & email verification

When you sign up, your workspace is created and you're signed in immediately — no waiting. We also email a one-time confirmation link to the address you used. Until you click it, a banner sits at the top of your portal with a Resend link button.

What works before you verify: everything about building and testing your assistant — knowledge, appearance, AI engine, quotes, tickets, the live widget. What waits for verification: the two features that email real people or move money —

• Lead-alert emails (the notify email under Actions) — so alerts can't be aimed at an inbox you don't control.
• Payment gateway keys (Paystack / Stripe under Actions) — so payment links and receipts are tied to a confirmed owner.

If the link never arrives, check spam, press Resend link (throttled to once a minute), or ask support to confirm the address for you. When you type any password on Temira — sign-up, sign-in or a secret key — the 👁 eye icon reveals it so you can check it before submitting.

Two-factor authentication (2FA)

Under Settings → Account security you can switch on app-based two-factor authentication. Turn it on, add the account to an authenticator app (Google Authenticator, Authy or 1Password) by scanning the link or typing the setup key, then enter the 6-digit code it shows to confirm. From then on, every sign-in asks for the current code — so a leaked password alone can't get into your workspace. To turn it off, enter a current code (or your password). Codes are standard TOTP (RFC 6238); the secret never leaves the server after setup.

Developer API

Everything the portal does is an API underneath, and you can call it programmatically. Send an Authorization: Bearer <token> header with your portal session token to reach grounded chat, business actions, the live-agent channel and your analytics. A full reference lives at /api-docs — always available to the workspace owner, and publishable to your team from Settings when you're ready. The read-only account-lookup action can also drive an API you define, so the assistant answers “where is my invoice?” from your own live data.

Controlling the assistant from your own page

Loading embed.js puts Temira on window — nothing else to install.

Temira.open()                      // open it
Temira.open({ask:"Where is my order?"})
Temira.ask("How do I add a user?") // open and ask
Temira.close() / Temira.toggle()
Temira.isOpen()                    // true / false
Temira.setContext({plan:"pro", orderId:"A-2291"})
Temira.setContext({audience:"customer"})  // lock the audience — hides the
                                   // Visitor / Customer / Buying picker cards
Temira.on("ready", fn)             // ready · open · close · ask · lead
Temira.destroy()                   // remove it entirely

Events also fire on document as temira:open, temira:lead and so on. Calls made before the script has loaded are queued and replayed — but only if this queue stub is on the page first. Paste it above your own scripts; it covers open, ask and setContext:

<script>window.Temira=window.Temira||{q:[],open:function(){this.q.push(["open"].concat([].slice.call(arguments)))},ask:function(){this.q.push(["ask"].concat([].slice.call(arguments)))},setContext:function(){this.q.push(["setContext"].concat([].slice.call(arguments)))}};</script>

And any element with data-temira works with no JavaScript of your own:

<button data-temira>Help</button>
<button data-temira="toggle">Support</button>
<a data-temira data-temira-ask="How do refunds work?">Refunds</a>

Triggers are bound by delegation, so a button your app renders later still works. A note on setContext: it personalises how the assistant answers — their plan, their order, their role. It is deliberately not treated as knowledge, so the grounding guarantee still requires a real source before anything factual is answered, and a personalised answer is never re-served to a different visitor from the cache. The reserved key audience is the exception: it steers the UI, locking who the assistant is talking to (visitor · customer · buyer, or your own audience keys) — an in-app Help button is customer-support by definition, so it shouldn't ask "I am a…". Same effect at load time with data-audience="customer" on the script tag.

Wiring setContext into your app

Think of it as a sticky note you hand the assistant about the person who is chatting: "this is Adaeze, an accountant, on the Enterprise plan, in the Lagos branch." The assistant uses it to speak to them properly — it never treats it as a fact about your business, and it never shows one user's note to another. Call it once the user is known (right after login is the natural place). Keys merge: call it again any time to add or overwrite a key.

// after the user signs in to your app
Temira.setContext({
  audience: "customer",       // reserved — locks the picker to customer-support
  name:     "Adaeze",         // a first name is plenty
  role:     "Accountant",
  plan:     "Enterprise",
  branch:   "Lagos Office"
});

// later, when something changes
Temira.setContext({ screen: "Register Bank Accounts" });

The rules, all of them:

• Up to 20 keys per context; extra keys are ignored.
• Every value becomes plain text, trimmed to 200 characters; key names to 40.
• Values must be strings, numbers or booleans — nested objects and arrays are dropped, so flatten first (orderTotal:"₦45,000", not order:{total:…}).
• It lasts for the page visit only — set it again after a reload. It is safe to call before the panel is open (it is delivered when the chat loads) and safe to call before the script finishes loading if you use the queue stub above.
• It reaches the assistant with the visitor's next question — there is no round-trip to wait for.
• Never put secrets in it — no passwords, session tokens or card numbers. It is conversational context for the AI, not a secure or authenticated channel, and it does not log the user in to anything.
• It cannot be used to feed the assistant answers: facts still come only from your knowledge base. If you want the assistant to know something, add it under Knowledge instead.

If you also want the assistant to see which screen the user is on automatically — including on-screen error messages — add data-copilot="1" to the script tag instead of hand-feeding a screen key; that is exactly what copilot mode is for.

Refer & earn

Every workspace has a referral link on the Refer & earn page (share it by WhatsApp, email or X). A business that signs up through it gets a discount on its first payment, and you earn account credit each time one of your referrals upgrades to a paid plan. The page tracks, in real time, how many businesses you've referred, how many converted, and the credit you've earned — with a full list of referred workspaces and credit events. The reward amount and signup discount are set by the platform (and shown live wherever they appear, including your portal and the public signup page).

Privacy, data & security

Each workspace is fully isolated — its own database, files and keys. Transcript storage can be turned off (counts only), retention is configurable, and any conversation can be erased permanently from the portal. Provider and gateway keys are stored server-side, shown masked, and never reach a visitor's browser. Passwords are stored only as salted PBKDF2 hashes (200,000 iterations) — never in plain text — and you can add app-based two-factor authentication for an extra step at every sign-in.

When your assistant fetches something on your behalf — a website you asked it to read, a connected-system lookup, a lead webhook — that request is fenced to the public internet: it can't be pointed at private or internal network addresses (your own servers, cloud metadata endpoints, localhost), so a crafted link can't turn a fetch into a way to reach infrastructure it was never meant to touch. And knowledge you upload can be kept answer-only — its text informs answers while the file itself stays private — with a per-source lock or a whole-workspace setting; see Answer-only sources under Teaching your assistant.

Sensitive account and platform actions — sign-ins (including failures), setting changes, payments, verifications — are written to an append-only activity log with a timestamp and source IP, so there's always an answer to "who did what, when".

Content your assistant reads cannot instruct it

Everything ingested from the web is untrusted, so the defence sits at three levels rather than relying on the AI to be sensible:

• What could be executed is neutralised as it is read. Command lines and control tags that Temira's own code would otherwise act on are made inert before they ever reach the assistant. This is deterministic — it does not depend on a model behaving well.
• The evidence is sealed. Knowledge passages are handed to the assistant inside a frame carrying a random reference generated per request, so a page cannot imitate the boundary and pretend to be your instructions.
• The assistant is told the difference. Only your settings and the visitor's own messages can direct behaviour; a passage never can, however it is phrased or whoever it claims to be from — and it never runs an action because a document asked it to.

What this does not do, stated plainly: it cannot stop plain persuasion. A page that simply lies — "the prices below are out of date" — contains nothing any code can detect. It is flagged for you and the assistant is told excerpts cannot instruct it, but wrong content in a knowledge base is ultimately wrong content. That is why the flag exists and why you see it.

Step-by-step instructions for deleting conversations, closing an account, or disconnecting Meta channels and revoking their access are on the Data Deletion page; the legal detail lives in the Privacy Policy and Security pages.