Scholarly record

FROM MODEL SELECTION TO SYSTEM DESIGN: ARCHITECTURAL FOUNDATIONS FOR AI AGENTS IN KNOWLEDGE-DRIVEN SYSTEMS

Kristina Dineva, Tatiana Atanasova, Ivaylo Keremidarski

Abstract

Developing effective artificial intelligence agents requires more than selecting an appropriate language model. Agent performance is fundamentally constrained by the quality, structure, and accessibility of the underlying knowledge, making curated, continuously updated knowledge bases a core architectural component. While large language models, reasoning-oriented models, and research-focused models provide distinct capabilities for language understanding, reasoning, and summarisation, their practical effectiveness depends on the system architecture in which they are embedded. In domains such as public administration, education, healthcare, and institutional services, AI agents must meet strict requirements for accuracy, transparency, contextual awareness, and reliability. These requirements cannot be met through model selection alone. This paper presents a knowledge-driven architectural framework for AI agents operating in knowledge-intensive environments. The proposed design integrates retrieval-augmented generation, structured knowledge bases, orchestration layers, external tool interfaces, and monitoring components into a modular, multi-layer system. The architecture supports data ingestion, preprocessing, indexing, knowledge representation, retrieval, reasoning, response generation, validation, and governance. The framework emphasises modular design to reduce hallucinations, improve traceability, ensure access to up-to-date information, and align outputs with domain-specific requirements and evaluation benchmarks. Particular attention is given to fragmented and heterogeneous data sources and to their transformation into consistent and actionable responses. The results highlight that the effectiveness of AI agents depends on the alignment between task complexity, knowledge structure, system architecture, and regulatory constraints.

Publication details

Back to publication list