AI Assistant for the Czech School Inspectorate

The Kvalitní škola and EDU.CZ portals contain thousands of pages of methodologies, criteria and examples of good practice.
For the Czech School Inspectorate, we built an AI assistant that knows its way around the entire content of kvalitniskola.cz and edu.gov.cz, answers questions in plain Czech, and provides precise links to the source with every answer.

Client – Czech School Inspectorate (ČŠI)
Sector – Public administration, education
Solution – AI assistant for the Kvalitní škola and edu.gov.cz portals (RAG)
In operation since – September 2026
Technologies – Python, LangGraph, Qdrant, LiteLLM, Langfuse, Claude (Anthropic)

The Challenge

‍
The Czech School Inspectorate evaluates the quality of education against criteria published on the Kvalitní škola portal, together with methodologies and recommendations. The content is extensive and valuable, but when a school principal asks what specifically to do to improve in a particular area, they have to find the relevant passages and connect them on their own.

ČŠI was therefore looking for an assistant that would answer such questions directly and practically. A tool like this would also significantly streamline the work of the inspectors themselves. Previous experience with general-purpose AI tools made the success criteria clear:

Verifiability. Every answer must include a working link to a specific methodology or criterion. General-purpose chatbots make up links or leave them out entirely.
Depth. A short, generic answer won't help a principal. They need concrete steps grounded in the methodology.
Up-to-date content without manual work. The portals have no API and their content changes. Manually exporting documents and uploading them to an AI tool would not be sustainable.
Independence and low operating costs. A public institution doesn't want to be locked into a single model or cloud provider, and it needs predictable costs.

We built the assistant on a RAG (Retrieval-Augmented Generation) architecture.

The language model doesn't answer "from memory". Instead, it answers from passages that the system first retrieves from the portals' content. As a result, it answers only with what ČŠI and the Ministry of Education (MŠMT) have actually published, and it knows where every statement comes from.

Because the portals have no API, we developed scraping agents that regularly crawl kvalitniskola.cz and edu.gov.cz, detect new or changed content, and process only that. The text is cleaned and split into passages, and from the very beginning each passage keeps its source URL, title and section. When a page disappears, it disappears from the index too, so the links in answers stay valid. Digital PDFs are processed locally, and OCR is used only for scanned pages.

The user's query, reformulated based on the previous conversation where needed, goes through hybrid search. Semantic search captures the meaning even when the wording differs, while full-text search catches exact criterion names. A relevance model then re-ranks the candidates, and only the best passages are passed to the language model as context. The model composes a thorough answer in Czech and adds a list of the sources used below it, with links to the exact place in the portal.

The core of the solution is decoupled from specific technologies through narrow interfaces.

The language model, vector database, embeddings and individual data sources can be swapped by changing the configuration or a single adapter, without touching the logic.

All models are called through a unified gateway (LiteLLM), which enforces budget limits and tracks usage. The entire solution runs in containers and can be moved to ČŠI's own infrastructure, on-premise or in the cloud, at any time.

Instead of a traditional fixed-scope project with a set handover date, we can offer the assistant as a service that is paid only for the months the customer actually uses it. This also makes ongoing improvements more flexible and removes the need for in-house expertise: content updates and fine-tuning of answer structure, depth and citations happen continuously, based directly on feedback from real users. The service also covers operation, regular data updates, governance, monitoring, ticketing and monthly usage reporting.

‍

Search – Qdrant: hybrid search (semantic + full-text), relevance re-ranking
Orchestration – Python and LangGraph: an explicit state graph for each query
Language model – Anthropic's Claude model family via the LiteLLM gateway, with a guarantee that data is not used for training
Operations and cost control – Containers, Langfuse tracing, budget limits, audit log

The entire content of two portals in one assistant, updated automatically with no manual work.
Every answer comes with a verifiable source and links straight to the specific methodology or criterion.
Low operating costs: a single answer costs less than one Czech crown (just a few euro cents), with no permanently running GPUs.
From first design to production in 6 weeks.

From 15 minutes to 1 minute: the assistant has cut the time needed to find relevant sources for a given question from about 15 minutes to 1 minute, and it also proposes concrete solutions based on those sources. At the expected volume of around 500 queries per month, this saves approximately 117 hours of work. Usage will scale to an expected 5,000 queries per month, raising monthly savings.

Future phases may integrate the assistant directly into the ČŠI information system, allowing users (inspectors, school principals, teachers, representatives of school founders, etc.) to receive even more targeted answers based on publicly available parameters of their school (type, size, etc.). As more user groups are added, we expect usage to grow gradually to up to 5,000 queries per month. The architecture is ready for this: a new data source or channel is an extension, not a rebuild.

Facing a similar challenge?

We help companies and public institutions make their own knowledge accessible through AI, with a focus on verifiable answers, data security and vendor independence. Get in touch and we'll show you how it would work with your data.

AI Assistant for the Czech School Inspectorate

The Challenge

Our other work

Cookies