SaaS & Startups

AI support chatbot that answers from your own documents

Retrieval-augmented chat that answers from your policies and product docs — and admits when it does not know.

Industry: SaaS & Startups
Platform: AI

Most support chatbots fail the same way: they are either scripted trees that cannot answer anything unanticipated, or a raw language model that answers everything — including things that are not true. This is our reference architecture for the version that works. Answers are generated only from the business's own material — policies, product documentation, past tickets — with the source cited, a confidence threshold below which it declines, and an immediate handover path to a person. It is the build we deploy and adapt for clients wanting AI support that is safe to put in front of customers.

The challenge

The engineering difficulty in a support bot is not generating text — that part is solved. It is grounding: making sure every answer traces to a document the business actually wrote, and that the system says "I don't know" rather than inventing a refund policy. The second difficulty is operational. Documents change, and an assistant confidently quoting last year's pricing is worse than no assistant at all.

Our strategy

Retrieval first, generation second. The knowledge base is chunked, embedded and indexed; every question retrieves the most relevant passages, and the model is constrained to answer from them with the source attached. Anything below the confidence threshold is routed to a human rather than guessed. Re-indexing runs on document change, so the assistant's knowledge and the company's actual policies never drift apart.

The solution

A chat assistant embedded in a website, app or WhatsApp that answers from the organisation's own content, cites its source on every answer, escalates cleanly to a human with the full conversation attached, and logs every exchange for review. Administrators see what was asked, what was answered and which questions the assistant could not handle — which turns the bot into a running report on the gaps in your documentation.

Results

What changed

24/7

availability

Always answering

Routine questions handled at any hour, in the customer's own words.
100%

grounded answers

Nothing invented

Every reply is generated from retrieved source material and cites it.
1

click to a human

Escalation built in

Handover carries the whole conversation to the agent who takes over.

Key features

What we built

Answers from your own documents

Retrieval-augmented generation grounded in your policies, docs and past tickets.
Citations on every answer

Each reply names the source passage it came from, so it can be checked.
Knows when to stop

Below its confidence threshold it declines and hands over rather than inventing.
Human handover with context

Escalation carries the full conversation, so the customer never repeats themselves.
Multi-channel

The same assistant on the website, in an app, or over WhatsApp.
Gap reporting

Unanswered questions become a list of what your documentation is missing.

Frequently asked questions

Is this a client case study?

No — it is a capability showcase: a reference architecture from our AI practice that we deploy and adapt per client. There is no client name or live URL attached because it describes the build, not one customer's deployment.

How do you stop it inventing answers?

By construction. The model is only allowed to answer from passages retrieved out of your own documents, every answer carries its source, and anything below a confidence threshold is declined and escalated to a person instead of guessed.

What does it need from us to get started?

Whatever you already have in writing — policy documents, product pages, a support inbox export. The quality of the assistant is decided almost entirely by the quality of that material, which is why we audit it before quoting.

Can it run on WhatsApp?

Yes. The same grounded assistant is reachable over the web, inside an app, or through the WhatsApp Business API.