Automating information extraction from Emails using Large Language Models

In the corporate world, managing large volumes of unstructured data, especially in email communications, significantly impacts efficiencyΜανουδακη Χρυσηις in information retrieval. In this diploma thesis, we design and implement a robust system for processing and extracting structured data from multilingual email threads within the logistics and transportation domain using Large Language Models (LLMs). The research methodology is structured around three phases: data preprocessing, information retrieval architecture, and agentic system integration. First, a preprocessing pipeline is developed to handle noisy data. This pipeline uses Regular Expressions and LLM-based cleansing to translate content to English, segment threads, and remove irrelevant metadata. Then, semantic deduplication is performed using vector databases. Concurrently, efficient inference strategies are examined to ensure low latency required by real-time applications. The second phase of the thesis involves investigating and comparing three Retrieval-Augmented Generation (RAG) workflows: a baseline RAG using vector databases; a GraphRAG approach that leverages knowledge graphs (KGs) to map complex entity relationships; and the LightRAG framework, which employs a dual-level retrieval paradigm. Finally, these components are integrated into a supervisor multi-agent architecture, which is accessible via a user-friendly chatbot interface. The supervisor orchestrates specialized agents that perform preprocessing, graph management, and complex query resolutions to fulfill natural language requests. Our results indicate that in most cases the proposed system was able to capture the core points of the email communications and generate meaningful replies to a range of queries. Overall, this thesis presents a scalable and reliable framework that combines knowledge graphs and agent reasoning to improve the accuracy and reliability of extracting complex logistics data from unstructured text.