Technology

An AI you can check

AskEU uses retrieval-augmented generation (RAG): the AI first finds the relevant official rules, then writes the answer from them only. Every answer is checked automatically and the rules are reviewed by people. The language model can be swapped; what matters is that every answer can be traced to its source.

How the AI works

From your question to an answer with sources, in six steps

AskEU is not a general chatbot. The language model only reads and writes. It can only use rules from our checked library, every answer is verified, and legal experts review the library.

01

Understands your situation

Reads your question in any of four languages and picks out what decides your rights: where you are insured, where you are, your status and the care you need.

02

Finds the rules

Searches our library of checked official rules and keeps only those that apply to your country and status.

03

Writes the answer

Writes the answer using only the rules it found. It is not allowed to add anything from its own memory.

04

Checks every fact

Matches every statement to its source. Anything without a source is removed before you see it.

05

Admits what it doesn’t know

If no checked rule covers your case, it says so and tells you which official office to contact.

06

Experts keep it accurate

Legal experts check the library. Answers that users report are reviewed by people.

Architecture

The library of rules is built and reviewed in advance. When answering, the AI can only use what the search gives it.

RULE LIBRARY · BUILT IN ADVANCEANSWERING · IN REAL TIME extracts ↺ sources checked weekly → reviewed again Official sourcesEUR-Lex · NCPs · fundsPrepare and labelextracts + country + statusExpert reviewsocial-security lawRule librarysearchable + date checked QuestionEN · DE · PL · UAStatus profilecountry · insurance · statusSearchfiltered by country and statusWrite the answeronly from the extracts foundFact checkunsupported statements removedAnswer + sourcessource and date per claim

Evaluation

We measure accuracy instead of assuming it. The system has to reach these targets before public release.

MetricRelease targetHow it is measured
Correct answers≥ 90%100 real questions, answers graded by a legal expert
Critical errors0Any answer that could cost someone money or delay their care
Claims with a citation100%Automatic check on every answer
Saying “I don’t know” correctly≥ 95%Questions with no checked rule must get “I don’t know” plus a contact
Response time< 10 sTypical time from question to answer

AI governance

Always clear it’s AI

Users always know they are talking to an AI, as the EU AI Act requires.

People stay in control

Experts approve every country module. Any answer can be reported, and reports are reviewed.

Not tied to one model

We use a leading language model with data processed in the EU. We can switch models without rebuilding the product.

Every answer traceable

Each answer stores which rules and which source versions it used.

Safety limits

What the AI will never do

Diagnose an illness or recommend treatment
Answer without an official source
Guess when it doesn’t know
Ask for or keep your medical records
Promote insurers or clinics for money
Pretend to be a person
Technology
Retrieval-augmented generationHybrid search (BM25 + embeddings)Multilingual embeddingsPostgres + pgvectorCitation verifierPython · FastAPINext.jsEU hostingEvaluation harness

Transparency under the AI Act

Article 50 of the EU AI Act says people must be told when they are dealing with an AI system. AskEU says so before the first question and marks every answer as AI-generated, together with its sources.

Regulation (EU) 2024/1689, Art. 50 Responsible AI →

Be among the first to use it.

Early users help us make the answers better. Available in English, German, Polish and Ukrainian.