Retrieval augmented generation makes a language model answer from your documents instead of from its training data. In Ireland it usually arrives with a second requirement attached: the documents stay inside the European Union.
Both are solvable, and they are more solvable together, because residency decides the architecture.
How Retrieval Augmented Generation Is Built Here
Chunking on the document’s own structure. Sections, clauses and headings, because fragments that end mid-sentence retrieve badly no matter what sits downstream.
Retrieval you can inspect. For any answer, which passages were retrieved and in what order, which is what makes a wrong answer diagnosable.
Citations that resolve. Every claim points to the passage behind it and somebody can open it.
Refusal when the documents are silent. An admission that the corpus does not cover it beats a confident invention, and it is the behaviour that decides whether people still trust the system in month three.
An evaluation suite. Real questions with agreed answers, scored before and after every change, and documented well enough to show an auditor.
A residency decision made on day one. Which components hold your text, where each one runs, and what crosses the network.
Where the Data Actually Goes
Three things touch your text: whatever stores the documents, whatever computes the embeddings, and the model that writes the answer.
All three get a named region, agreed at scoping and written into the architecture note. Where the requirement is strict, the answer is an open model on infrastructure you control, which is on-premise AI. Where it is looser, EU-region hosted services are enough.
That record is also the evidence base the EU AI Act rewards, and your compliance advisers can read it without translation.
The Baseline Comes First
Before any retrieval augmented generation is built, we measure what plain keyword search over the same corpus already achieves against your own questions.
That is the number to beat, and it is usually higher than people expect. Where the margin is thin, improving the search is the honest recommendation, and it arrives in week one at our cost.
Where the margin is wide, you have a business case carrying a figure rather than an impression, which is what a board will ask for.
What You Own at the End
The code, the prompts, the chunking configuration, the index settings and the evaluation set, in your repository from the first commit.
There is no subscription and no dashboard. You can re-run the evaluation yourself and get the same score we did, which is the point of handing it over rather than hosting it for you.
Tell us what your people keep asking, where the answer lives, and whether it may leave the building.
Related Services
Whether it works is AI evaluation, scored on whether the cited passage actually supports the claim.
Where the documents cannot leave your infrastructure at all, start at on-premise AI. Where the job is extracting structured fields rather than answering questions, that is document automation.