AI knowledge system for national tourism platforms
An AI-powered support solution that organises complex documentation and helps different user groups access relevant information across tourism platforms.
Client and context
The Croatian National Tourist Board (Hrvatska turistička zajednica) operates national tourism platforms used by a wide range of audiences — from tourism boards and accommodation providers to businesses and administrative users. These platforms are supported by an extensive body of documentation: regulations, instructions, categorisation rules and operational guidance.
Challenge
The documentation was comprehensive but difficult to navigate. Different user groups needed different subsets of information, and finding a precise answer often meant reading through long documents or contacting support directly.
An AI-based solution had to answer accurately from controlled sources, respect the differences between user groups, and work reliably in multiple languages — without inventing information that is not in the documentation.
Our approach
We treated the documentation itself as the core of the system. Documents were processed, structured and classified by topic and target user group before any AI component was introduced. This produced a knowledge base in which every piece of content has a clear origin and audience.
On top of this we built a retrieval-augmented generation (RAG) architecture: the assistant retrieves the relevant, classified content first and formulates answers strictly from it. Responses are multilingual and the system records which sources were used, supporting review and continuous improvement.
What we delivered
- Document processing pipeline for structuring and updating source content
- User and content classification aligned with real audience groups
- RAG architecture with controlled, source-bound answering
- Multilingual response capability
- Analytics on questions, topics and answer coverage
- Deployment, monitoring and ongoing maintenance
Technical overview
Structured, classified document store with versioned updates and audience-aware retrieval.
Retrieval-first answering bound to controlled sources, with guardrails against unsupported claims.
Multilingual understanding and responses across the languages used by the platforms.
Deployment with monitoring, analytics and a maintenance process for document updates.
System visuals
Simplified system diagram. Conversation shown is illustrative.
Outcome
Different user groups can now find relevant, documented answers directly, in their own language, without navigating the full body of documentation. The classification layer keeps answers appropriate to each audience, and analytics show which topics users ask about most — informing both support and documentation improvements.
Ongoing work
The system is actively maintained. New and updated documents flow through the same processing and classification pipeline, and the answer quality is reviewed continuously against real user questions.