Data, RAG & backend
Thousands of pages of policies, contracts and documentation, and an answer in seconds, backed by a source citation. I'll build you a private RAG architecture your data never leaves and never trains anything.
Who it makes sense for
- Company knowledge sits in documents nobody can search quickly
- Compliance forbids sending internal data to public AI tools
- New hires take months to onboard because "only Frank from accounting knows that"
What you get
- A private vector database over your documents, policies, contracts, wikis, e-mails
- Every answer linked to its source document, full auditability
- Access control: whoever can't read a document won't get an answer from it either
- APIs with no data retention, or an open-source model entirely in your cloud
How it works
Documents are split into semantic blocks and stored in a vector database inside your infrastructure. A query finds the relevant passages and the language model composes the answer from them, it never answers "from its own head". If the answer isn't in the documents, the system says so plainly.
Where it actually helps
Support and helpdesk
An agent queries your own documentation instead of paging through three systems. The answer comes with a link to the policy it came from, so it can be quoted to the customer directly and nobody has to fill in the gaps.
Onboarding new people
A new joiner spends their first months asking colleagues about things that are written down but cannot be found. RAG takes that load off your experienced people without anything having to be rewritten into a new wiki.
Contracts and compliance
Questions like what notice period do we have with this supplier are answered today by searching PDFs. With a citation on every answer it becomes traceable and auditable, which in regulated industries is a condition rather than a bonus.
Why RAG and not a fine tuned model
The most common question on a first call. Fine tuning a model on company data sounds appealing, but for a knowledge base it loses on almost every count.
- Freshness. A changed policy just gets re uploaded with RAG. A fine tuned model has to be trained again.
- Traceability. RAG shows which document the answer came from. A fine tuned model answers from memory and evidences nothing.
- Access control. RAG can filter by the user's permissions. A model that absorbed the data into its weights absorbed it for everyone.
- Cost. The difference is usually an order of magnitude, both at deployment and at every change.
Fine tuning makes sense for style and answer format, not for facts. If we hit a case where it does, I will say so.
When I will talk you out of it
- When the documents do not exist. RAG can only answer from what is written. If the knowledge lives in people's heads, writing it down is the first job.
- When there are few documents. Up to roughly a hundred pages, decent full text search is usually cheaper and faster.
- When the documents contradict each other. Three versions of the same policy turn RAG into a generator of disputed answers. Clean up first, deploy second.
Parameters
FAQ
Does the model train on our data?+
No. RAG doesn't train the model, documents are searched only at query time. I use APIs with no data retention, or a model running entirely on your side.
Can it handle our industry terminology?+
Yes. Deployment includes an evaluation set built from your real questions, I measure answer quality on your domain terminology before handover and continuously after.
What does a RAG deployment cost?+
From CZK 25,000, delivered in 4 to 6 weeks. The price is driven mainly by the volume and state of the documents, because cleaning up the source material is routinely a third of the work, then by the number of sources and whether access control is needed. Add running costs for model calls. Full pricing is on a separate page.
What formats can the documents be in?+
PDF including scans, Word, Excel, wiki, Confluence, SharePoint, emails and plain text. Scans go through text recognition first. Format is rarely the obstacle; the obstacle is usually the same document existing in three versions in three places.
What is the difference between RAG and an AI agent?+
RAG is the layer that finds an answer in your documents and evidences it with a source. An AI agent uses that layer and can also act, opening a case, moving an appointment or writing something into a system. Almost every agent I build has RAG underneath. RAG on its own fits where the system only has to answer.
Related
- AI agents & LLM applications, when the system should act on the data, not just answer
- AI automation, reading documents and deciding inside a process
- RAG in practice, how to give a model company data without the whole internet seeing it
- Pricing, prices and lead times in one place
Bring three questions you currently answer by digging through documents by hand, and I'll show you how RAG answers them.
Write to me, I'll reply within hours.