What does RAG stand for?

Retrieval-augmented generation. The assistant retrieves relevant passages from your documents first, then generates an answer using only those passages, with citations.

Will the assistant make things up?

The point of RAG is to prevent that. Answers are built from retrieved passages and carry citations to the source document, so anyone can check where a fact came from.

Can it read scanned PDFs and documents in Slovenian or Croatian?

Yes. Vandri's pipeline includes OCR for scanned files and handles regional-language documents, not only English.

Does it respect who is allowed to see what?

Yes. Retrieval is permission-aware, with role-based access and audit logs, so the assistant only surfaces what each user is allowed to access. It runs inside EU data residency.

RAG Over Your Company Documents \ Vandri

The problem RAG solves

A general chatbot knows a lot about the world and nothing about your company. Ask it about your return policy, your insurance product, or last year's tender documents, and it will either say it does not know or, worse, invent a confident answer.

Feeding all your documents into a prompt is not the fix either. There is too much of it, it changes constantly, and most of it is irrelevant to any single question.

RAG is the practical answer. Instead of hoping the model memorised your files, you let it search them at the moment of the question and answer from what it finds.

How RAG works, step by step

Ingest. Your documents are read and split into passages. This includes PDFs, scanned files (via OCR), SharePoint libraries, ERP exports and more.
Index. Each passage is turned into a vector and stored in a search index. Vandri uses Qdrant for this.
Retrieve. When someone asks a question, the system finds the passages most relevant to it.
Generate. The language model writes an answer using only those retrieved passages, and cites which document each fact came from.

The important part is step four. Because the answer is built from retrieved text and carries citations, the person reading it can click through and check the source. That is what makes it trustworthy enough to use at work.

Why grounding matters more than the model

A common mistake is to think the assistant is only as good as the language model behind it. In practice, for company data, the retrieval and grounding matter more than which model you pick.

A brilliant model with poor retrieval gives fluent, confident answers from the wrong documents. A solid model with excellent retrieval and citations gives answers your team can rely on. Getting the ingestion, chunking, OCR quality and permission handling right is most of the work, and it is where enterprise RAG projects usually succeed or fail.

Vandri's RAG stack uses advanced retrieval with citations and OCR, so answers come from your documents rather than from guesses.

Permissions are part of the product

If ten people can query the assistant, it must respect what each of them is allowed to see. A finance document should not surface for someone who could never open the file directly.

This is not a detail to add later. Permission-aware retrieval, role-based access and audit logs need to be built in from the start. Vandri handles this inside EU data residency, following GDPR and ISO 27001 principles. If your data is sensitive, that is the difference between a demo and something you can actually deploy. More on that in our guide to GDPR-compliant AI in the EU.

What this looks like in production

Two of our shipped systems show the pattern:

BlueBox, for Sava RE. A RAG assistant over more than 2,000 insurance documents, built directly into the agents' CRM. An agent asks a question and gets an answer with citations in seconds, without leaving the tool they already use.
AIDE, for the University of Maribor. An AI intranet over the faculty's SharePoint. Staff ask a question and it answers from the real documents, so people stop hunting through folders.

Both follow the same principle. The assistant does not replace the source of truth. It makes the source of truth answerable in plain language, with a link back to the original.

Where a RAG assistant fits

RAG is a strong fit when:

You have a large, growing body of documents that people need answers from.
The answers already exist in your files, they are just hard to find.
Different people should see different things.
You need to trust and check the answers, so citations matter.

It is a weaker fit when the knowledge is not written down anywhere, or when a question needs a real-time calculation rather than a document lookup. In those cases a different approach, or a combination, works better. We will tell you which.

How Vandri builds it

We start narrow. Pick one document set and one group of users, connect the source, and ship an assistant that answers well for that scope. Then we watch real questions, fix what retrieval misses, and expand. You get something useful in weeks, not a year-long project that never lands.

Want an assistant that knows your documents?

Tell us where your knowledge lives, whether that is SharePoint, an ERP, a shared drive or thousands of PDFs. We will tell you honestly whether RAG is the right tool and what the first version should cover. Book a call or see the RAG solution and our work.

RAG over your company data: an AI assistant that answers from your own documents