Parsování a příprava dokumentů pro Retrieval Augmented Generation (RAG)

Abstract

This bachelor thesis compares modern document parsers from the perspective of their impact on Retrieval-Augmented Generation systems. The theoretical part summarizes RAG architecture, the requirementsforhigh-qualityparsing, andtheroleofMarkdowninpreservingthesemanticstructure of documents. The practical part describes an experimental testbed built on top of LlamaIndex, a unified intermediate representation for multiple parsers, and an evaluation methodology based on the LLM-as-a-Judge approach. The final analysis compares selected parsers with respect to correctness, faithfulness, retrieval quality, processing speed, and operational cost, and formulates recommendations for different document-processing scenarios.