Technology

magyarkozlony.ai is an AI-powered search and analysis platform for the Hungarian legal corpus. Users ask questions in natural language; the system queries the indexed corpus (Magyar Közlöny, the Fundamental Law, with more to come) and returns answers with page-precise citations and the original page image.

Beyond chat, the system also generates finished files: PDF reports, Excel sheets, PowerPoint decks and charts — from a single question.

The platform is multilingual (HU/EN), fully GDPR-compliant, and all data stays in the European Union.

The system is NOT a single monolithic AI but a network of specialised agents working together. Each agent is optimised for a specific task and is constrained to its own permission scope.

A central orchestrator (LLM) interprets the user's request, decides which agents to invoke, and assembles the final answer. Complex questions may involve multiple agents collaborating.

Available agents: • Search agent — hybrid (vector + keyword) search over the indexed corpus. • PDF report agent — generates a structured, branded PDF from the matched content. • Excel agent — produces a formatted xlsx (deadlines, effective dates, amendment lists). • PowerPoint agent — generates finished slides with MK citations. • Chart agent — draws an interactive bar or line chart based on the question. • Email agent — sends the generated file with a single click.

The advantage of this architecture: the LLM only decides "what to do"; the actual execution (search, file generation, sending) runs in deterministic code. The AI cannot "hallucinate" data — sources always come from a real, verifiable place.

1. The user asks a question in natural language in the chat interface.

2. The orchestrator (AWS Bedrock — Anthropic Claude Sonnet model) interprets the request and selects the relevant agents.

3. Agents work in parallel: hybrid search across the vector database (LanceDB) and Postgres, file generation, chart rendering, and so on.

4. The orchestrator combines the agents' results and produces a natural-language answer with page-precise citations.

5. A source card accompanies every citation: gazette number, page number, and the original PDF page image at full size — verify accuracy in one click.

All data stays inside the European Union — every step of processing happens in Germany.

Hosting: Hetzner Online GmbH, data centre in Germany.

AI processing: AWS Bedrock, eu-central-1 region (Frankfurt). The AWS DPA contractually guarantees that user data is NOT used to train AI models and is NOT retained after processing.

Encryption: all data is stored with AES-256 encryption; network traffic uses TLS 1.3.

Isolation: user sessions are logically separated; data is accessible only to the relevant user.

Daily automated backups with geo-redundant storage.

AI model: Anthropic Claude Sonnet (on AWS Bedrock, EU region). The model only sees the question and the relevant context, never the full corpus.

Embedding: Cohere multilingual model for hybrid search — works well for both Hungarian and English.

Search: combined vector (semantic) and BM25 keyword search. LanceDB stores the indexed text chunks; Postgres stores metadata (gazette number, date, page).

Frontend: Next.js + React with a multilingual UI.

Ingest: new Magyar Közlöny issues are ingested automatically every day; new content becomes searchable within minutes of publication.

For technology questions and partnership inquiries: [email protected]

Developer and rights holder: Chain Advisory Kft. Joint operator: Erba 96 Kft.