This thesis focuses on intelligent information retrieval from large document collections using agentic retrieval architectures. The goal is to design and evaluate an intelligent search system similar to retrieval-augmented systems without generative components, emphasizing effectiveness and interpretability. The solution integrates LangChain and LangGraph with vector databases such as Weaviate and language models from Ollama and OpenAI. A custom multilingual benchmark dataset, with a specific focus on the Czech language, among others, supports the evaluation of retrieval quality by showing how graph structure and system design influence accuracy and efficiency, and provides guidelines for building robust multilingual retrieval systems. The results demonstrate that while hybrid retrieval and cross-encoder reranking significantly improve performance on multilingual datasets, highly complex agentic pipelines and deep search architectures do not automatically guarantee better retrieval quality compared to simpler baselines, often increasing latency without proportionate gains. Additionally, the effectiveness of end-to-end agentic workflows is shown to depend heavily on the underlying language model's capability to make robust and reliable intermediate decisions.