A chat assistant built on your site, your PDFs, your policies. It answers only from them, names the source under each reply, and hands off to a person when your material does not hold the answer.
The demo assistant below reads exactly six documents, nothing else. No source means no answer. One named source rides under every reply it gives, and every document and its vectors live in the EU.
This is not a mockup. Type a question and watch it answer only from the demo shop's six documents, name the file it used, and hand off to a person the moment its material runs out. Ask it to ship to the US and see it refuse instead of guess.
Most chat widgets improvise, because sounding helpful is easier than being right. On a business people trust, that is not a small bug. Here is the same customer question, answered twice.
“Yes, we offer same-day delivery on all orders. It will be there by tonight.”
“We do not offer same-day delivery. Standard is 2 to 4 working days, and orders before 2pm ship the same day. Want me to flag your order as urgent for the team?”
Same model underneath. The difference is the engineering around it.
Grounding is not a setting. It is retrieval-augmented generation, RAG for short: your material is indexed once, and each question walks the same fixed path before a word reaches the person asking.
From the widget, from Slack, from an email. It enters as plain text and nothing more.
The indexed passages of your own documents are ranked for the closest match. Nothing outside them is in scope.
The reply is composed from the retrieved passage only, in your tone, never from the model's general memory.
The document it drew from rides under the answer, traceable after the fact.
every_answer.log
“Standard delivery is 2 to 4 working days. Orders before 2pm ship the same day.”
Cited: Shipping-info.mdNo matching passage means no answer. The assistant declines and hands off instead of filling the gap.
The grounding never changes. What changes is the material it reads and who it answers.
Product, policy and how-to questions answered in your tone, around the clock. Off-topic goes to a person, never gets improvised. The one most businesses start with.
One assistant over your wiki, SOPs and policy documents. Your team gets the real answer in seconds, not a confident invention or a half-hour search.
A first layer that reads each incoming request and routes it to the right team or the right answer. Faster first response, hard cases stay with people.
Every worry here is earned, because most chatbots deserve it. This is how the build answers each one before it becomes your problem.
We match the tone to your actual website and documents, so it answers the way your best person on a good day would. Not in generic chatbot English.
It answers only from your sources, names each one, and stops when unsure instead of inventing. That behaviour is engineered in, not switched on.
You point us at the content and approve the result. The indexing, the tuning and the testing happen on our side.
We keep the index fresh, watch the answers, and re-tune as your content changes. It does not quietly drift while nobody is watching.
The same four that ship with our voice agent. Wired in, never a toggle, never an upsell. Each one is a filter a reply has to pass before it reaches the person asking.
Sensitive or off-topic requests go straight to a person. No loop, no dead end.
Below the confidence gate it says it does not know, instead of guessing.
No source means no answer. Out of scope gets a plain no, never a confident wrong reply.
It says it is an AI, names its source, and logs every answer it gives.
A grounded assistant is excellent at some jobs and a poor fit for others. Selling it for the wrong one is exactly the kind of slop we are against. So, plainly:
We index your material, tune the tone on your real copy, and test against your own questions before go-live. It runs on our servers in the EU, in Falkenstein. Your documents never train a public model, and a data processing agreement is part of the contract.
The exact grounded build drops into the places your customers and your team already are. One index behind all of them.
You point us at the content. Everything else, including everything after go-live, happens on our side.
A link to your site, your PDFs, price lists, a wiki export. Whatever already holds the answers. Nothing else ever enters the index, so nothing else can ever come out.
Your material is split into clean passages and indexed. The tone is tuned on your real copy, so it answers the way you write, not the way chatbots write.
We test it against your own real questions and agree what counts as correct, together, before it ever talks to a customer.
We keep the index fresh as your content changes, watch the answers, and re-tune the edges. Grounded on day one, still grounded in month six.
Not a settings toggle, and not a promise made after the fact. This is how the system is built.
The paperwork is public: read the data processing agreement and the sub-processor list before you ever write to us.
Retrieval-augmented generation. Instead of answering from a model’s general training, the assistant first retrieves the relevant passages from your own documents, then answers only from those. It names the source and declines when your material does not hold the answer.
An off-the-shelf widget guesses to sound helpful. A grounded build answers only from your real sources, carries the source with every reply, and stops when it is unsure. Each one is hand-engineered and ships with the four safeguards. No no-code glue, no third-party widget owning your data.
Your sources: website, PDFs, price lists, policies, a wiki export, whatever holds the answers. We build the retrieval pipeline, test it against real questions, and only then put it in front of people.
It stops. Below the confidence gate the assistant says it does not have a reliable answer and offers the next sensible step: a handoff, a callback, or a person. No hedging that still sounds like an answer.
EU only, in Falkenstein. Your documents and their embeddings never leave the EU and never train a public model. A data processing agreement is part of the contract.
Builds start at €690. The most-chosen Signature build is €1.290, and Bespoke starts at €2.490. Every build comes with a Care plan from €79 per month that keeps it indexed and tuned. Send a briefing and you get a fixed quote for exactly what fits, no hourly surprises.
Yes. A 60-day performance commitment, written into the contract: it answers correctly from your content, or your money back. The bar is at least 90 percent correct on a test set of your real questions, agreed in writing before you sign and verified together at go-live.
