home / use cases / chat, leads and customer service / support-chatbot-kb

A support chat that only knows your documents

A chat widget on your site answers support questions from your own knowledge base — manuals, warranty terms, setup guides — citing the source, in the visitor's language. When your documents don't contain the answer, it says so and opens an email thread with a human instead of guessing. All of it runs on the box in your office.

Watch the demo

starts when: A customer types a question into the chat on your site

The problem

Most support questions have already been answered — in the manual nobody opens, on the FAQ page nobody scrolls, in the warranty PDF nobody downloads. So they arrive as email instead, joining the 121 messages a day your team already faces, and “what’s the torque setting for the M8 model” waits a business day for an answer that was on page 14 all along.

Off-the-shelf support bots fail this in two directions at once. The scripted ones can’t reach page 14. The generative cloud ones can — and will also cheerfully invent a page 15, while your manuals and your customers’ questions stream through a vendor’s servers.

How your Aivell does it

Your documentation becomes the front line, with a colleague who has actually read it:

  1. The question arrives in the widget. A snippet on your site carries it to the box; the visitor just sees a chat that answers.
  2. The answer comes from your documents — only. The knowledge base holds your manuals, policies and guides, indexed onboard. A guardrail limits answers to the published folder, so the bot’s knowledge is exactly your documentation, nothing more.
  3. Sensitive topics carry their source. Warranty and returns answers name the document they came from, so a promise in chat is always a promise your policy already made.
  4. Gaps become escalations, not fiction. No matching source? The Aivell says so, collects the visitor’s email, and opens a thread to support@ with the transcript — Sara replies from her normal inbox.
  5. The knowledge base learns from its misses. Unanswered questions are collected, so each week you write the one page your customers actually needed.

The chore reads like instructions to a new support hire — tone, escalation, the topics never to guess at. The difference is that this hire has perfect recall of your manuals, answers at 11 PM in four languages, and physically cannot cite anything you didn’t publish.

The chore, as you'd write it.

No flowcharts, no code — a chore is just a message to your Aivell, in your own words. It shapes it into a solid, guarded procedure and follows it to the letter. This one:

Giulia Aivell · answer support chats ● active

answer from the knowledge base — manuals, warranty terms, setup guides

reply in the visitor's language; short answer first, detail if asked

warranty and returns questions always name the source document

nothing in the kb fits? say so honestly, take their email, open a thread to support@

an angry customer or an order complaint goes straight to Sara, transcript attached

never guess compatibility between models and spare parts — escalate those

outside office hours, say when a human replies — weekdays by 9

keep a list of unanswered questions so the kb grows a page a week

You
Got it — I'll start right away, and check with you whenever something needs a human.
Giulia Aivell

The tools it uses.

Connected with single-click passkeys — no copied tokens, no OAuth hell.

Chat widget fields support questions on your site
Knowledge Base your manuals and policies, indexed onboard
Email escalates to a human with the transcript

Sensible guardrails.

Every tool can be limited to exactly what this chore needs. For this one, you might set:

Knowledge Base Answers may draw only on the published-support folder. Draft docs and internal notes are invisible to this chore.
Email New mail goes only to support addresses on your own domain, always with the chat transcript attached.

Questions, answered.

How do I keep it from making things up? +

By construction, not by hoping. A guardrail restricts answers to your published support folder, and the chore's rule for gaps is explicit — admit it and escalate. A question your documents can't answer becomes an email to your team, not an invented paragraph.

What does it take to load our manuals in? +

Point the knowledge base at your documents — PDFs, office files, exported help pages — and they are indexed on the box. When a manual changes, the updated file replaces the old one; the chat starts answering from the new version, no retraining ritual involved.

Do customer support chats pass through an AI provider? +

No. The widget on your page talks to the box in your office, where the model, the index and your manuals all live. What your customers ask — and what your documentation says — stays inside your network, which is the whole point of running support on your own hardware.

How does a human take over a conversation? +

By email, which your team already lives in. When the chore escalates, support@ receives the transcript and the visitor's address; a person replies from the normal inbox. No agent console to staff, no separate support platform to log into.

Related.

What is an Aivell?

An Aivell is an on-premise AI colleague: a small box that plugs into your network, runs its own AI onboard and takes the repetitive work off your desk. Prompts, documents, data — nothing ever leaves your office. Unplug the internet: it keeps working.

You describe each task as a chore, in plain language. Your Aivell turns it into a solid, guarded procedure and runs it in the background — with guardrails on every tool, a complete audit trail, and approvals in your hand when you want them. One-time setup, fixed monthly fee. No tokens, no overages.

Watch the demo

onboard inference · air-gapped · fixed monthly fee