[Demo: all blocks] SAV Encyklopédia Beliana: an AI assistant grounded in encyclopedia articles
- Client
- Slovak Academy of Sciences (SAV)
- Industry
- Publishing & education
- Timeline
- MVP in 1 month (unconfirmed)
- Services
- AI & LLM solutions
/ The challenge
An authoritative encyclopedia in the age of AI assistants
Encyklopédia Beliana is Slovakia's national encyclopedia, published by the Slovak Academy of Sciences. General-purpose AI tools now answer questions on many of the topics it covers, and the project was designed to address the concerns this raises.
Answers without a source
General AI tools can answer questions on Beliana's topics without pointing readers to the encyclopedia.
Risk of misinterpretation
Without the original articles, generic tools may misread them or mix in less reliable sources.
Questions instead of searches
Many readers prefer to ask a specific question rather than search for and read a full article.
Control over the content
SAV wanted an official AI interface to its own content instead of leaving its use to third-party systems.
Slovak Academy of Sciences, Bratislava/ Our approach
Answers grounded in Beliana's own articles
We proposed a retrieval-augmented generation (RAG) assistant. For each question, the system first retrieves relevant Beliana articles, and the language model then writes the answer from them.
This ties answers to the encyclopedia's own content and lets readers see where each answer comes from, so they can check it against the original article.
Grounded in the encyclopedia
The assistant is designed to answer from the retrieved articles and to say when Beliana doesn't cover a question.
Sources with every answer
Responses include links to the articles they draw on, so readers can check the original text.
A self-hosted model
The GPT-OSS 20B language model runs on self-hosted infrastructure rather than through a third-party AI service.
/ How it works
From a reader's question to a source-linked answer
The assistant is embedded as a widget on the encyclopedia's website. Each question goes through the same four steps.
Question
A reader asks a question in Slovak in the widget.
Embedded widget
Retrieval
Relevant passages from Beliana's articles are retrieved from a ChromaDB vector database.
ChromaDB
Generation
A self-hosted GPT-OSS 20B model writes the answer based on the retrieved passages.
GPT-OSS 20B
Answer with sources
The reader receives the answer together with links to the source articles.
Source links
Readers can also rate answers with a thumbs up or down and leave a comment, which gives the team material for reviewing and improving the assistant.
Built with
- RAG
- GPT-OSS 20B
- ChromaDB
- MCP
/ The encyclopedia online
Where readers meet Beliana today
Screenshots of the public encyclopedia — the context the assistant was built for. Screenshots of the assistant itself are pending from SAV.
/ The outcome
What the delivered system enables
Questions in Slovak
Readers can ask about encyclopedia topics in natural Slovak and get an answer based on Beliana's articles.
A path to the full article
Links in each answer lead readers to the articles behind it, where they can read and check the original.
Access for other AI agents
Through the Model Context Protocol (MCP), other AI agents can query Beliana's content via a standardized interface.
/ Results (sample data)
The project at a glance
Sample figures for testing the layout: the first two come from internal notes and are not yet confirmed by SAV.
- to launch the MVP
- 1 month
- encyclopedia articles indexed
- ~1,000
- steps from question to cited answer
- 4
Readers can now ask Beliana a question directly — and every answer shows them where it came from.
Placeholder until SAV approves a quote from a named person
Photo credits: Encyclopaedia Beliana volumes: Pepo13, CC BY-SA 3.0, via Wikimedia Commons; SAV building: Kelovy, public domain, via Wikimedia Commons
