RAG / Knowledge base
AI assistant over your company data
An assistant that answers from your documents — and shows, for every answer, which document it came from. For companies where the answer can be found, it just takes too long.
RAG (retrieval-augmented generation) is an approach where the language model answers not from what it learned in training, but from your documents. The question first retrieves the relevant passages from your documentation, and the model answers only from those — with a link to the source.
That solves the main problem with ordinary chatbots: when the model has no basis, it has nothing to invent. Typical deployments: an internal assistant over procedures, contracts, technical documentation or price lists. Scope and price depend on volume and on how orderly the documents are — real projects start at €1,200.
When it makes sense
A RAG assistant does not solve a lack of information. It solves the situation where the information exists but finding it takes longer than it should:
Internal procedures
Employees keep asking the same things — procedures, exceptions, who approves what. The answer is in a policy nobody reads end to end.
Technical documentation
Manuals, specifications, datasheets. Support and sales both need a fast and accurate answer.
Contracts and terms
Questions like “what did we promise this client” are currently answered by opening PDFs.
Onboarding
A new person spends the first month asking about things that are written down somewhere — they just don't know where.
If your questions are not about documents but about customers and products, you are probably looking for an AI chatbot instead. If something also has to happen as a result of the answer, that is an AI agent.
How it works
- The documents are processed. Files are read and split into meaningful pieces — not mechanically by character count, but along the structure of the document.
- An index is built. Each piece gets a vector representation, which allows searching by meaning rather than by exact wording.
- The question retrieves the basis. When a question is asked, the most relevant passages from your documents are found.
- The model answers from what was found. The language model receives the question together with the retrieved passages and answers only from those.
- The answer shows its source. Every answer shows which document and which part it came from, so it can be verified.
That last step is essential in a company setting. An answer without a source cannot be verified, and what cannot be verified is something people stop trusting after the first mistake.
I go through it in detail — how documents are split into parts, what the vector database does, how to choose the model and where the limits are — in the article AI assistant over company data: how it works and what it needs.
What decides the quality
In RAG projects the choice of model or tool is not what decides. This is:
- Document quality. Current, unambiguous, non-duplicated material. If three versions of the same policy are in circulation, the assistant will answer confusingly — exactly like a person would.
- How the text is split. Bad splitting is the most common cause of bad answers — the answer falls between two pieces and neither makes sense alone.
- Scope. A narrowly defined area works far better than “all company documents”. That is why it starts with one subject area.
- Permissions and sensitivity. Who is allowed to learn what has to be settled. The assistant must not show the salary policy to someone with no access to it.
- Updating. Documents change. Without a process for refreshing the index, the assistant will be answering from old versions within six months.
What RAG does not solve
It does not replace order in your documents. If nobody knows which version applies, the assistant will not fix that — it will only make it visible. Which, incidentally, is often the most useful output of the first week.
It does not guarantee correctness. The model can be wrong even with the right passage retrieved. That is why the answer always shows its source and, for anything important, is meant to be verified rather than taken on trust.
It is not a legal or tax opinion. The assistant finds what is in your documents. It does not judge whether that is correct or lawful.
It has a running cost. Every question means a retrieval and a model call. At high volume that has to be budgeted up front.
Where your documents go
This is the most common and most legitimate question in RAG projects, so I address it in the proposal:
- The index can run on your side. The retrieval part can be operated inside your infrastructure — the documents never leave it.
- The model is a separate decision. With an external model (Claude, GPT), the retrieved passages are sent to the provider. You get that written out precisely.
- The scope can be limited. Sensitive categories of document can be excluded from the index entirely.
- Access rights. The assistant can respect who has access to which documents.
Nexflow does not provide legal advice. Where personal data or sensitive documentation is processed, I recommend an assessment with your DPO before deployment — and I will gladly supply a technical description of where the data actually travels.
What it costs
- from €1,200An assistant over one defined subject area — a bounded set of documents, one channel, answers with sources.
- individualSeveral areas, access rights, running the index inside your infrastructure, connections to internal systems.
RAG is the one service where I refuse to quote a “from €390” range. An honest deployment includes preparing the documents, testing answers on real questions and settling access — that cannot be done in a few hours. On top of the deployment, budget for running the index and the model API.
I recommend starting with a pilot over one subject area. It will reveal both the quality of your material and the real usefulness — and it is far cheaper than finding that out on a company-wide project.
Frequently asked questions
What is RAG?
RAG (retrieval-augmented generation) is an approach where a language model answers based on documents retrieved for the question, instead of answering from its training data. That is how it knows your internal information and can show where the answer came from.
How is it different from an ordinary chatbot?
An ordinary chatbot answers from what you put into its instructions, or from the model's training. A RAG assistant searches your documentation for every question and answers from the retrieved passages — so it handles a volume that would never fit into instructions, and it can say where the answer came from.
What documents can it handle?
Ordinary company formats — documents, PDFs, spreadsheets, internal pages. What matters is not the format but whether they are current and unambiguous. Scanned documents without a text layer have to be processed first.
Can the assistant invent an answer?
The risk drops substantially because the model answers only from retrieved passages and shows the source. It cannot be eliminated entirely, so the assistant is deployed as a way to find information quickly, not as the final authority for decisions with consequences.
Do our documents have to go to the cloud?
The retrieval part can be operated inside your infrastructure. When an external model is used, the retrieved passages are sent to the provider — the proposal describes precisely what is sent and what can be excluded.
How much does it cost?
An assistant over one defined subject area starts at €1,200. Larger solutions with access rights and in-house operation are quoted individually. On top of that, budget for running the index and the model API by number of questions.
How is it kept up to date?
A refresh process is part of the deployment — when a document changes, the index is updated. Without that, the assistant would eventually answer from old versions, which is worse than not having it.
Where should we start if we are unsure?
With a pilot over one subject area — internal procedures, say, or the documentation for a single product. It shows both the quality of your material and the real usefulness, without investing in a company-wide solution.
Sources
- Anthropic — official documentation — working with context and citing sources
- n8n — official documentation — nodes for working with vector databases
Related articles
Related work
Related services
Have an area where the answer exists but takes too long to find?
Tell me which documents it involves and who is asking. I will tell you whether it suits RAG, and what would need fixing in the material first.
Get a free consultation →