Find the rule and verify the evidence
The Czech Road and Motorway Directorate, ŘSD, works with technical regulations, internal directives and guidelines for construction preparation, asset management and public procurement. Employees may not know a document’s exact title or terminology. They need an answer to a specific question and a way to check its source.
For ŘSD, we deployed an AI assistant for internal regulations, built on DOC BRO. Semantic search connects everyday language to relevant document passages. Answers include references that lead users to supporting sources.
How the solution fits together
The application and database of indexed document passages run within ŘSD’s infrastructure. Azure provides language models and source-file storage. User access integrates with Microsoft Entra ID and the internal network environment.
The diagram represents the documented hybrid deployment: application and database on premises, models and source files in Azure. Arrows show functional relationships, excluding the planned future migration to Azure.
Natural-language questions with source references
Employees open the assistant in their internal environment and ask a question. The system retrieves relevant passages and composes an answer linked to supporting documents. Follow-up questions can continue within the same conversation context.
An employee looking for remote-working rules, for example, can ask about the conditions without first finding the directive’s title. They can then open the document behind the response. Source references are part of the workflow: they distinguish the assistant’s wording from the original regulation.
Search also accounts for document validity at a specified date. This matters when reviewing historical cases governed by earlier rules. OCR makes scanned materials without an original text layer available for retrieval.
Documents need a path into search
The assistant depends on the contents of its knowledge base. Integration therefore covers the path from the source directive system through Azure Blob Storage to indexing in DOC BRO. Synchronisation transfers documents, and the application prepares their passages and metadata for search.
PostgreSQL with pgvector stores the data needed for semantic retrieval. Azure OpenAI supports query processing and answer generation over retrieved context. The user interface uses React, application logic uses Python, and operational components run in Docker containers.
Access control and visibility into usage
Entra ID authentication uses existing identities and groups. Access passes through the internal firewall and reverse proxy layer. The database is accessible from the application environment rather than being exposed directly to users.
Alongside chat, the solution provides knowledge-base administration and usage reporting. Administrators can inspect conversation counts, token consumption and response feedback. Users can flag an unhelpful answer and add a comment to help identify the problem.
An administration screen from the supplied documentation. Figures are a captured application state, not current ŘSD operational results.
Feedback reports can be exported for further analysis. Administrators can also change the default model for new conversations without restarting the application. The operations team gains tools to track quality and consumption alongside the assistant itself.
Technologies and our role
| Area | Technologies | What we deliver |
|---|---|---|
| Working with regulations | DOC BRO, OCR, semantic search | Referenced answers and document validity handling |
| Application | Python, React, Docker | User interface, application logic and administration |
| Knowledge base | PostgreSQL, pgvector, Azure Blob Storage | Synchronisation, indexing and retrieval |
| AI services | Azure OpenAI | Query processing and answers over retrieved context |
| Access | Entra ID, Traefik, oauth2-proxy | Integration with ŘSD identity and infrastructure |
What the partnership delivers
Employees can search by meaning and verify answers against original regulations. ŘSD gained a solution adapted to its environment, including document, access and operational management. Questions, supporting documents and feedback form a traceable workflow for working with internal rules.
“Through its work with Cloudfield, the Road and Motorway Directorate successfully deployed an AI assistant for finding information in internal regulations. Built on DOC BRO, the solution enables fast and semantically accurate document search, provides relevant answers in natural language and makes it easy to check how rules have been interpreted.”
— Ing. Josef Šejnoha, Head of Strategy and Communication Department, Ředitelství silnic a dálnic, s.p.