Home / Products / RAG system
Pixelschnitzel · Folio 03 · RAG

Your knowledge.
With sources.
Reliably queryable.

Contracts, manuals, tickets, emails, SharePoint, Confluence, the knowledge of your organisation is already there. It is just scattered everywhere. Our RAG system makes it accessible, with a dependable answer and a traceable source.

Answer With source citation
Permissions From the source system
Operation German cloud · On-premise
Book a RAG demo → How it works
Contracts SharePoint Manual Tickets Emails Confluence DMS RAG Pixelschnitzel
· Preface ·

Full-text search
finds words.
We find answers.

Search “Which rule applies to complaints over €1,000?” and you get PDFs. Type it into a standard chatbot and you get a made-up answer. Neither is usable in day-to-day business. A good RAG system does something different: before every answer, it retrieves the relevant documents from your own sources, hands them to a language model and lets it answer only on that basis, with a source citation. What is not in the sources is not claimed.

· Structure ·

Four steps.
From source
to answer.

i

Connect. Gather the sources.

SharePoint, Confluence, file servers, DMS, mail servers, ticketing systems. We collect the knowledge where it is created, without staff having to maintain it all over again.

SharePointConfluenceFile serverDMSOutlookJiraZammad
ii

Understand. Structure & vectorise.

Documents are split up semantically, enriched with metadata, permissions, validity period, source, and moved into a vector store. That is the foundation for good answers.

QdrantWeaviateEmbeddingsOCRMetadata
iii

Answer. Question → source → answer.

The question is matched against the vector store, relevant sources are selected and handed to a language model. It answers solely on the basis of those sources, and names them.

Llama 3.3Qwen 2.5Guardrails“Don’t know”Citations
iv

Govern. Permissions, audit, feedback.

Whoever asks only sees what they are allowed to see. Every request is audited. Staff rate answers, and from that, source selection and prompt strategy improve.

RBACAudit logFeedback loopMonitoring
· What’s different ·

Answers with a
citation,
not from the gut.

Every answer shows where it comes from. Hallucinations become noticeably rarer, and the system is allowed to say honestly: “I don’t know that.”

Contract § 4.3
// Source 1
Policy 12
// Source 2
Manual 7.1
// Source 3
Answer
→ based on S1 + S2
· Application ·

Six use cases
that take the load off right away.

i.

Sales. What did we agree on?

“What terms apply to customer X?”, answered in seconds, with a reference to the contract and the quote.

ii.

Support. Solution knowledge from tickets.

10,000 old tickets become a knowledge treasure, staff tap into it before doing their own research.

iii.

Compliance. Policies within reach.

Terms and conditions, data protection, compliance rules, answered with the source passage instead of a 50-page PDF.

iv.

Clinic. Guidelines & SOPs.

Hygiene regulations, training materials, guidelines, queryable on the ward, not in a binder.

v.

Administration. Case & procedural knowledge.

Procedural manuals, decrees, legacy cases, found faster, answered concretely, with a source reference.

vi.

Plant. Maintenance & engineering.

Engineering drawings, maintenance logs, instructions, answers at the machine, not in the filing cabinet.

Library with stacked books and files
“Knowledge that sleeps in a file
is knowledge that does not exist.”
· Frequently asked ·

Answers,
even before
you ask.

What sets RAG apart from a normal chatbot?
A normal chatbot answers from the model’s knowledge, and is therefore often wrong. Before every answer, RAG retrieves the relevant documents from your sources, hands them to the model and lets it answer only on that basis. The result: verifiable statements with a source.
What happens to permissions from SharePoint or DMS?
They are respected. Every request is checked against the permission model, and in the answer a member of staff only sees what they are also allowed to access in the source system.
Can the system also say “I don’t know”?
Yes, and that is a key feature. If the sources do not provide a dependable answer, the system says exactly that, rather than making something up.
Where does the system run?
Both are possible. The standard is operation in a controlled environment in Germany. On request, fully on-premise, see On-premise AI for business.
How long does it take to get started?
First dependable answers on a defined source (a manual, a Confluence section) typically after 2–4 weeks. Broader integration in 6–12 weeks, depending on the number of sources.
What does it cost?
Setup in the low five figures plus ongoing operation, depending on the volume of sources, model operation and load. A concrete figure after the first conversation.
How is the system kept up to date?
Sources are re-indexed regularly. Changed documents flow in automatically, superseded ones are removed. You define the frequency and depth together with us.
· First conversation ·

Give us
one source.
We’ll show you answers.

One defined source set, a manual, a Confluence area, contracts, is enough. In a single meeting you’ll see what the answers and sources would look like. Realistic, on your real content.

Book a RAG demo → Rather write to us