ReAct-Driven AI Agent for Document Analysis : An analysis of the ReAct approach and performance comparison with baseline LLMs

Sammanfattning

The rise of Large Language Models (LLMs) has transformed Natural Language Processing (NLP) tasks, enabling impressive capabilities in reasoning, summarization, and question answering. However, their application to realworld, multi-step document analysis in industrial contexts remains limited by a lack of interpretability, tool integration, and consistent performance. This thesis investigates whether the Reasoning and Acting (ReAct) paradigm can address these limitations by combining LLM reasoning with external tool usage in a structured loop. The core problem is enabling accurate, traceable, and robust extraction of structured information from documents like policies and procedures—tasks where direct prompting or retrievalaugmented generation often fail. To tackle this, a ReAct-style agent was implemented using LangGraph and GPT-4o, equipped with tools such as deep search and code execution. The agent was evaluated against baseline methods across tasks like information extraction, classification, and document question answering, using both internal (Volvo) and public datasets. Results show that the ReAct agent achieves higher accuracy on multi-step reasoning tasks, and significantly improves interpretability by exposing intermediate reasoning traces, with only moderate latency overhead. The project demonstrates that ReAct agents can support more auditable, modular, and capable Artificial Intelligence (AI) systems for real-world document analysis, enabling workflows that were previously difficult to automate reliably.