Frequently asked questions

Mignuti Chatbot is an embeddable AI chatbot for B2B websites. You upload documents (PDF, text, Markdown), add Q&A pairs or crawl a URL — Mignuti Chatbot trains a chatbot on that content. One line of code embeds the widget on any site (WordPress, Shopify, Webflow, custom HTML). Hosted in the EU (Germany), GDPR-ready out of the box.

Two audiences: digital agencies in DACH/BeNeLux who resell chatbots to their SMB clients (recurring 20% partner commission), and SMBs themselves — tax advisors, real-estate agents, hotels, e-commerce shops, consulting firms — who want a 24/7 first-line of customer support without an enterprise-sized invoice.

Start free with 100 messages and our badge. Then: Lite €9/mo (500 messages), Starter €29 (1,000), Standard €79 (5,000, custom branding, lead capture), Pro €199 (20,000, 5 bots, 10 team seats, badge can be switched off), Enterprise quoted individually — full white-label with your own domain. All prices net of VAT. Annual billing saves about 20 percent. Beyond your quota the bot stays online: each further message costs €0.03 to €0.015 depending on plan, and a hard cap protects you from a surprise bill.

Free is a 7-day trial with 100 messages, 3 documents, 1 project and the "Powered by Mignuti Chatbot" badge. Upgrade in one click to Lite, Starter, Standard or Pro anytime — or simply pause the project.

At checkout by card, PayPal, SEPA Direct Debit, Apple Pay or Google Pay. Stripe handles the payment — you enter card details there, never with us. Every plan can also be booked against an invoice, paid by bank transfer, no credit card needed: a short email to media@mignuti.com, we set up the account and send the invoice. The 7-day trial needs no payment details at all.

Yes — monthly or annual, cancel from the dashboard or without logging in at /en/cancel, effective at the end of the current billing period. No support call needed. Your data is downloadable as a GDPR export at any time.

Under 60 seconds for a Q&A-pair bot, under 10 seconds for a crawl preview of the first pages, under 10 minutes for a 25-MB document. The dashboard shows live progress while the document is parsed and indexed.

Documents as PDF, text, Markdown or JSON-FAQ (PDF up to 25 MB each), full website URLs (Mignuti Chatbot crawls several sub-pages with CSS-selector isolation), or single Q&A pairs typed in directly. Everything lives in one Knowledge-Base view. More office formats (Word/DOCX, RTF, ODT) are in the works.

Yes, from the Standard plan up. Connect the folder once, drop your files in — the bot reads new and changed files by itself. Delete a file in Dropbox and it disappears from the knowledge base too. PDF, TXT and Markdown up to 25 MB are read; syncing runs daily, weekly, monthly or only on demand. We only ever read: nothing is written, moved or deleted in your Dropbox, and we only see the connected app folder, not the rest of your storage.

Because uploading happens once and knowledge ages constantly. The price list changes, the manual gains a chapter — and the bot keeps telling the old version for months without anyone noticing. With a connected folder you maintain your files where your team maintains them anyway. For one-off documents, uploading is still the faster route.

Yes, from the Standard plan up. Connect your Google account and pick individual files in the familiar Google dialog — PDF, text, Markdown and Google Docs. Those files are all the bot reads; it sees nothing else in your Drive. Change a file and it is read again; remove it from the selection or delete it in Drive and it leaves the knowledge base.

Because we took the most frugal access Google offers: we only ever see the files you click yourself. Folder sync like Dropbox would require read access to your entire Drive — more convenient for us, but a mandate a chatbot does not need. Google Docs are pulled as plain text, which for a chatbot is actually better than a PDF.

Anything that lets you paste a <script> tag: WordPress, Shopify, Webflow, Wix, Squarespace, Framer, plain HTML, React, Vue, Astro, Next.js. The widget is a Web Component with full Shadow-DOM isolation, so it cannot break or be broken by host CSS.

Copy one line of HTML from the Embed page in your dashboard and paste it into your site's <body>. The widget loads asynchronously, weighs about 37 KB (gzip), and only fetches additional resources after the visitor opens the chat. See the docs for platform-specific tutorials.

Yes, two ways. In the crawl settings you exclude whole areas, for example your entire news path. In the page list you remove single pages: those go on a block list and are not picked up again on later refreshes. This is the usual case for press archives — the articles stay online and findable on your site, the bot just stops reading them. It picks sources by content similarity, not by date, so a 2023 release often matches a question just as well as your current page.

Yes. The Knowledge area has a JSON import with a downloadable template. Prepare your set, upload it, and export everything back to JSON at any time to edit and re-import. A second import creates no duplicates: existing questions are skipped. Every pair stays individually editable, switchable and deletable afterwards — unlike an uploaded document, which only exists as a whole.

Yes — type Q&A pairs directly into the dashboard. Best for opening hours, pricing, return policies, founder bio. Each pair is embedded instantly (no async worker), so you can test it in the live preview ~1 second after saving.

Database, file storage and embeddings run on our own server in Germany; Azure OpenAI (fallback) in Stockholm. Both EU. The primary LLM (Anthropic Claude) runs in the USA — safeguarded by the EU SCCs 2021/914, with personal data automatically redacted before transfer. Full architecture diagram on /trust.

Yes. We provide a Data Processing Addendum (Art. 28 GDPR), maintain a public subprocessor list with 30-day change notifications, support end-user data exports + deletions per Art. 17/20, and run PII redaction on inbound messages before they reach the LLM.

Article 50 disclosure ("you are talking to an AI") is shown in every chat session — unblockable, contractually fixed for all customers. The chatbot identifies itself as AI in both the welcome message and a persistent footer banner. We also offer a downloadable AI-Act compliance statement.

Hallucinations are limited by retrieval-only answering (Mignuti Chatbot never invents content beyond the sources), but they can happen. Our AGB cap our liability at 12 months of fees. We refuse high-risk verticals (medical advice, individual legal/tax advice) by default — every vertical template ships with content-policy keywords that route those queries to "please contact a human".

Yes, on Standard and Pro plans. Reply to media@mignuti.com with your standard form and we sign within one business day on weekdays.

Yes — out of the box. The bot detects the visitor's language and answers in 12 languages (en, de, es, fr, it, nl, pl, pt, tr, ar, hr, ru). Your knowledge-base can be in one language, the visitor can ask in another — Claude Haiku 4.5 + crosslingual embeddings handle the bridge.

Four layers. (1) The bot answers only from sections of your own sources. (2) A minimum relevance score (0.35 by default, adjustable per project) drops weak matches; if nothing clears the bar, your fallback text is served instead of a guess. (3) The system prompt forbids inventing content. (4) When the best match looks shaky, the answer is generated first, checked against the sources, and only sent if it holds up. You can see every relevance score per answer in the Message Inspector.

Yes. Below every answer, the widget shows numbered citation chips: "[1] FAQ.pdf · p. 14 · 87%". Clicking expands an excerpt of the source paragraph. Customers can verify any answer in 1 second — that is the trust-multiplier vs. faceless chatbots.

Through the Glossary tab in the Knowledge area. You add a term with its other spellings, for example "MAPS" with "Master of Social Work", plus an optional meaning. Timing is what makes it work: the expansion happens BEFORE the search, not at answer time. Otherwise semantic search finds nothing, because to a language model "Maps" is the English word for charts. The mapping works both ways: ask for the long form and the abbreviation is searched too. For longer lists there is a JSON import with a template. Abbreviations that appear on your pages are found automatically; the glossary is for the ones that appear nowhere in the text.

By default from your content — which is not always what you want. If a subpage lists one employee's direct line, the bot may quote it even though that person is not responsible. That is why the bot settings have a "Central contact details" field. Once it is filled, the bot uses only those details and never quotes phone numbers or e-mail addresses found on your pages. Names, departments and postal addresses stay usable, otherwise it could no longer answer "who looks after this?". Leave the field empty and nothing changes.

By default yes: the widget shows the three most relevant sources with a relevance score under each answer. The bot settings have a "Show sources to visitors" switch. Some operators want the traceability on display, others would rather not expose their site structure. Only sources that clear your configured minimum relevance are listed, so an answer that says "I do not know" carries no source list. The test chat always shows every match regardless, so you can trace what an answer is based on.

No, they are checked first. Suggestions are written in the context of the answer just given — but tapping one starts a NEW search with the new question, which can turn out differently. So two checks run before a suggestion is ever shown: anything your blocked terms would reject is not offered, and if there is no material on the topic in your knowledge base, the suggestion disappears. This costs no extra waiting time because the existing search index is used. Deliberately generous: it checks whether material exists, not whether the answer will end up detailed.

Document replaced — re-upload, the old embedding is auto-replaced. URL source — set the cadence to daily/weekly, Mignuti Chatbot re-crawls automatically. Q&A pair — edit in the dashboard, embedding regenerated in 1 second. No retraining, no downtime.

Yes. Configure a handover URL or webhook — when the visitor types "speak to a human" or the bot is uncertain, the conversation is forwarded via webhook (e.g. to Slack, an email tool or your CRM). Webhook signatures are HMAC-SHA256.

From Pro the "Powered by" badge can be switched off, but the widget still runs on our domain. Full white-label with your own domain (chat.your-agency.com) and your branding, where the customer never sees the word "Mignuti Chatbot", is available on the Enterprise plan.

20% of the net plan price of every paid subscription you bring in — for as long as the customer pays and the program runs. Settled quarterly by credit note, paid out from €50. Terms at /en/legal/partner-terms, apply on /en/agency-partner.

Yes — Pro ships with 5 projects. Every client is a project, each with its own bot config + embeddings. The dashboard shows health/cost/leads per client in one table, with bulk actions (CSV-export, open-in-tabs, pause-all). Need more than 5 clients? → Enterprise with a custom contract.

They can self-service migrate to their own Mignuti Chatbot account in one click. We do not lock customers into the agency. This is intentional — it makes you easier to sell because clients know they're never trapped.

On the Pro plan: monthly co-branded reports auto-mailed to your clients ("In April, your bot answered 1 247 conversations, 23 hot leads, 89% CSAT") with your logo. Plus a partner-asset-pack: case-study templates, pitch decks, security one-pager.