RAG / knowledge bases
A chatbot that answers from your documents — with sources.
RAG (retrieval-augmented generation) connects an AI model to your knowledge base: answers come from your documents and cite the source, instead of the model guessing. We build it on AWS — your data stays with you.
The problem
A model alone guesses. RAG answers from sources.
A public chatbot answers from what it "remembers" from training — it can be confident and wrong at the same time (a hallucination). For a company that is a risk: a wrong answer about a product, a procedure or a contract costs trust, sometimes money.
RAG reverses the order: it first retrieves the right passages from your knowledge base, then has the model answer strictly on that basis — with a link to the source. Your staff and customers can see where the answer comes from and verify it.
How we do it
From documents to a sourced answer.
The same pipeline we built for our own product — we transfer it into your organisation.
1. Order in the data (ETL)
We gather and clean the sources — documents, databases, pages, files — and version them in a single, controlled knowledge base. The hardest work is before the model, not inside it.
2. Knowledge base on AWS
We split documents into chunks, compute embeddings and index them in a vector knowledge base — Amazon Bedrock Knowledge Bases, the native RAG mechanism on AWS. The data stays in your cloud account (EU region); it does not travel to public models.
3. Strict RAG with citations
The model is instructed to answer only from the retrieved passages and to show the source. When the base has no answer, it returns "I don't know" instead of making one up.
4. Evaluations and monitoring
We measure answer quality and cost per query automatically. The base and the prompts change over time — we keep the quality up, not "set and forget".
Architecture
What a production RAG on AWS looks like
Every layer runs in your AWS account (EU region) — from data sources to monitoring.
- Data sources
- Chunking
- Embeddings
- Vector store
- LLM (Amazon Bedrock)
- Application
- Monitoring & evals
At every layer: per-role permissions, source citations and audit logs — what regulated firms require.
For regulated companies
A production knowledge system with governance — not a basic chatbot.
Finance, healthcare and manufacturing cannot afford a data leak or an "almost right" answer. The same mechanisms that keep quality up also provide control and an audit trail.
Data separation and EU region
The knowledge base runs in your AWS account and EU region. Data does not reach public models or their training.
Permissions and access
We scope access to sources per role and user — an answer uses only the documents the asker is allowed to see.
Audit logs
We log queries and the sources used — leaving a trail of who got which answer, when and on what basis.
Human-in-the-loop
Where the stakes are high, we add human approval before an answer is used.
Compliance: AI Act and GDPR
Citations, evaluations and the audit trail support AI Act and GDPR obligations; the risk classification is closed by an AI Act audit.
Need AI that also acts in your systems? See AI agents. Pair it with an AI Act audit and our full services.
Industries
Four industries we go deepest in.
In finance, manufacturing, healthcare and hospitality, we deploy AI where answers must come from controlled data and remain under human oversight. The common foundation is RAG or automation on AWS, source-level permissions, quality tests and governance matched to AI Act risk. Each industry still needs different data, integrations and accountability boundaries, so the dedicated industry pages explain the use cases, system limits and implementation priorities in detail.
Proof
We built this ourselves — in production.
RAG
versioned knowledge, source citations and controlled refusal
$0.0006
cost per AI scan in production
302,516
drug-interaction records in the production knowledge base
mojApteczka production data, June 2026. In healthcare an "almost right" answer is dangerous — which is why strict RAG, source citations and automated evaluations were a requirement, not a decoration.
See the full case studyGo deeper
RAG in practice — from our blog
Questions
Before you ask
No. We build the knowledge base in your own AWS account (Amazon Bedrock Knowledge Bases). The documents stay with you, under your control — which also matters for GDPR and trade secrets.
Strict RAG reduces hallucinations: the model is constrained to answer only from the retrieved sources and to cite them. When the base has no answer, the system returns "I don't know". On top of that, automated evaluations watch quality over time.
ChatGPT answers from general knowledge and does not know your documents. RAG connects the model to your knowledge base — answers are specific, sourced and limited to what you actually have.
Priced by scope — it depends on the number and state of the sources, integrations and query volume. We start with an audit that sets the scope on facts, not promises. No hidden stages.
First an audit (2–4 weeks), then a proof of concept on a real slice of your data — a working proof before you commit the full budget. The production timeline depends on the number and state of your sources and integrations; we set it after the audit, on facts, not on a promise of "in X weeks".
Free tool
Is your organisation ready for a bounded AI PoC?
Answer 24 questions about process, data, shadow AI, ownership, risk and governance. The result points to the next step without claiming automatic conformity.
Have documents your AI should answer from?
Let's start with an audit: we'll check whether your sources are RAG-ready and how fast we can ship a working system — without experimenting on your budget.