The Challenge: The Technical Complexity of Managing Distributed Knowledge
Collaborating with the Foundation for the Development of the Education System (FRSE)—which serves as the National Agency for the Erasmus+ Programme in Poland—presented us with an ambitious goal: how do you effectively govern a vast, fragmented knowledge ecosystem that must remain accessible to thousands of beneficiaries in real time?
The core issue extended beyond raw data volume to its internal structure. The formal Programme Guide had to co-exist with dynamic news updates, while the system needed to flawlessly differentiate past application deadlines from upcoming calls for proposals. From an engineering standpoint, our mandate was to eliminate informational noise, achieve high answer accuracy (grounding), and build an architecture that scales effortlessly.
The Decision: Engineering an Intelligent Architecture
Rather than building a superficial wrapper on top of a large language model, we engineered a comprehensive architecture based on the Retrieval-Augmented Generation (RAG) framework. At Wise People, our goal was not merely to “deploy AI,” but to deliver a secure, predictable, and fully auditable knowledge engine.
- Hybrid Infrastructure & ETL: We built the core application on Laravel and Livewire to deliver a high-performance knowledge management panel. Data resides within a hybrid architecture: PostgreSQL serves as the relational database engine, while Qdrant operates as a high-performance vector database. To orchestrate ETL workflows, we integrated n8n to handle automated content extraction, document chunking, and vector indexing.
- Knowledge Hierarchy Engine: We established strict source weighting. The official Programme Guide acts as the primary ground truth, while news updates and event calendars serve as secondary context. This rule engine prevents contradictory guidance, consistently prioritizing statutory guidelines over temporary bulletins.
- Semantic Query Translation: We designed the system so that users do not need to master formal institutional jargon (such as specific Key Action numbers) to get accurate answers. Our semantic search layer maps natural language inquiries directly to exact documentation sections, lowering barriers for first-time applicants.
- Grounded LLM Integration: We leveraged secure, high-performing OpenAI models (GPT-4) governed by rigorous system prompts. We introduced chronological validation layers that verify date accuracy, explicitly separating outdated program announcements from active application windows.
The Outcome: A Scalable, Enterprise-Grade System
Through this custom implementation, FRSE secured a platform that goes far beyond typical chatbot capabilities:
- Operational Transparency: The custom administrative dashboard offers end-to-end visibility into user session quality, individual answer ratings, and built-in cost tracking metrics—a critical factor in enterprise AI deployments. The FRSE team maintains complete governance over model operating expenses.
- Reduced Team Workload: Beneficiaries receive concise, structured responses paired with direct source citations, noticeably decreasing repetitive manual support tickets.
- Communication Consistency: Smart contact routing logic (such as segmenting queries between KA1 and KA2 actions) directs users to the appropriate subject-matter specialist whenever human support is required.
The critical factor in this collaboration was understanding the client’s operational context and adapting our engineering approach to an organization with thousands of source documents and international visibility. It was an ambitious challenge, but the close collaboration of both teams delivered an outcome that directly benefits every student engaging with the new Erasmus+ AI assistant.
Robert Szarata, CEO Wise People
FAQ
Deploying artificial intelligence within a public institution that manages thousands of pages of regulations required far more than a simple LLM integration. Below, we address key questions about the technical behind-the-scenes of the FRSE project: from eliminating hallucinations via RAG architecture and structuring knowledge source hierarchies, to ensuring full cost control and operational transparency in an enterprise-grade system.
The Foundation for the Development of the Education System (FRSE) operates as the National Agency for the Erasmus+ Programme and the European Solidarity Corps in Poland. Managing substantial budgets and hundreds of institutional workflows annually, FRSE required an AI solution to offer 24/7 informational access to thousands of beneficiaries while freeing internal staff from repetitive inquiries so they could focus on substantive project support.
Off-the-shelf AI chatbots remain vulnerable to hallucinations (fabricating details). The Retrieval-Augmented Generation (RAG) pattern compels the language model to search exclusively within curated, validated documents provided by FRSE before generating an output. Every answer is grounded directly in official Erasmus+ regulations, ensuring factual and regulatory accuracy.
In enterprise environments, the core differentiator is never just the underlying model, but the governance and reliability of the surrounding knowledge ecosystem. In our experience, three pillars determine success:
Data Reliability (Grounding): The AI must operate strictly inside a secure sandbox of approved institutional assets. Without an enterprise RAG pipeline that enforces source citation, organizations lose control over their communication.
Architectural Transparency: Stakeholders require continuous visibility into where the model sources its answers and the operational costs incurred per request. Custom admin tooling turns AI from an opaque “black box” into a predictable business asset.
Contextual Usability: Effective systems decode user intent regardless of whether the user knows technical terminology. A mature AI deployment translates internal administrative structures into natural, user-centric answers without requiring domain expertise from the applicant.
