00 / RAG system · AI knowledge assistant
Your team asks in plain sentences. The RAG system searches manuals, contracts and tickets, answers from them and shows where it says so. If it finds no evidence, it tells you.
Intro call €0Cost estimate in four questionsBased in Duisburg, working across Germany
01 / What it does
A language model on its own answers from whatever it learned at some point. It has never seen your manuals.
So the knowledge assistant searches your documents first and only then answers. That turns manuals, contracts and tickets into an internal AI search across your company knowledge.
Under every answer you see which document and which section it comes from. One click and you read the passage yourself before you rely on it.
You decide who may ask about what. Access rights apply during the search itself. If you work for several clients, each tenant sees only its own documents.
When a document changes, the assistant reads it again. The next answer comes from the new version, and nobody has to maintain a list.
How data gets in ↓02 / References
Zeasy RAG is our store for code and documentation from all our projects. Before a coding agent here writes a building block from scratch, it asks there whether one already exists. We build the same thing for your documents.
In-house product · Developer tool
Zeasy RAG · Knowledge base for code
Our own knowledge base across all our repos and documentation. Coding agents query it before they write code, so they find what already exists and why it was built that way.
When a known vulnerability shows up in a library, we see which projects are affected.
Result Around 270,000 sections of code and documentation, as of October 7, 2026. Every answer names the project and the file.
Client project · Under NDA
AI ticket routing · AI classification
A language model rebuilt as a classifier and fine-tuned on many years of tickets. It writes no text. It only returns category, department and urgency.
Trained on cleaned, anonymized tickets from many years. The client runs the system in-house. Name and model stay under NDA.
Result The opposite approach: here the model learned a behavior instead of looking things up. It runs in-house at the client.
03 / No source
A language model writes a fluent sentence even when it doesn’t know the answer. For a question about notice periods or service intervals, that is more dangerous than no answer at all.
So we build the assistant to say only what is written in a section it found. What it can’t back up, it doesn’t claim.
Then it answers from exactly those sections and attaches them as sources.
Then it says it found nothing on that in the documents, and it doesn’t guess. Those questions can be collected. They show you which document is missing.
Then for you it doesn’t exist. The assistant doesn’t even reveal that there is a document on the question you aren’t allowed to open.
04 / Connecting data
The assistant is only as good as what it finds. So most of the work goes into connecting the right sources and keeping them current.
Manuals, contracts, tickets, code. We start with the documents your team asks about most often and add more once the first ones hold up.
Every document is divided into sections and stored so the search finds the right one, even when your question uses different words than the manual.
When a source changes, it gets read in again. Our own store is up to date shortly after every change.
Whether the assistant runs on a server we operate for you or in-house at your company, we decide together. The AI ticket routing, for example, runs at the client’s own site.
View hosting ↗05 / Scope
Both fall under AI and automation for us. Which one you need depends on whether a person asks a question or a workflow should keep moving.
A person asks a question and gets an answer with its source. The knowledge stays in the documents, the assistant looks it up. It can run as its own page or inside a web app you already use.
View web apps ↗A workflow should keep going without manual work: sorting tickets, filing emails, reading details from documents. There the model decides and doesn’t write an answer for people.
View AI automation ↗06 / What it costs
We bill by effort. How many hours it takes depends on the number and type of sources and on how fine-grained the access rights need to be. The calculator on the contact page gives you a first cost estimate in four questions. We plan the scope with you.
On top come running costs for operations, depending on where the assistant runs. If it runs with us under maintenance, changes of up to one hour of work are included.
What it’s about, how big, how demanding, how urgent. No obligation and no contact details. If you send an inquiry afterwards, you don’t type anything twice.
Get a cost estimateEvery payment comes after a sign-off. You can step out between any two milestones.
07 / How it works
A knowledge assistant doesn’t have to know everything on day one. It has to answer well the questions your team asks every day.
Which questions come up again and again, and where are the answers written down? If something off the shelf will do, we tell you here. Costs €0.
We come from Duisburg to your company, wherever you are in Germany. We watch who looks up what, where the documents live and who may read them today. That turns into a specification and a quote.
We connect the most important documents. Your team tests with real questions and checks every answer against its source.
More sources, access rights, connections to your tools. We build in one-week sprints, each ending with a demo.
The sources stay connected and current. The source code lives in your repository, even while we keep looking after it.
08 / Questions
09 / More
The intro call costs nothing, and neither does the cost estimate after it. If what you need can be bought off the shelf, we tell you its name and send no invoice.
Start a projectWe usually reply the same business day