AI-Powered Document Intelligence with Retrieval-Augmented Generation

This thesis presents the design and development of an intelligent agent based on the Retrieval-Augmented Generation (RAG) architecture, integrating a Large Language Model (LLM) to support automated analysis and generation of official resolutions with a fixed structural format. The proposed system addresses the limitations of conventional language models in handling domain-specific, high-precision information by combining neural text generation with targeted knowledge retrieval from structured and unstructured sources. The work includes the implementation of a data ingestion pipeline for indexing resolutions, leveraging semantic embeddings to enable similarity-based retrieval. A carefully engineered prompting strategy, enriched with few-shot examples, guides the LLM in generating contextually accurate responses while preserving non-generative factual elements such as legal references, personal names, and financial data. The evaluation process covers both retrieval performance, measured through vector similarity metrics, and generation quality, assessed via domain-specific accuracy and completeness criteria. Results demonstrate that the RAG-enhanced LLM approach offers factual reliability and contextual relevance. The proposed architecture gives a scalable and adaptable framework for intelligent document processing, with potential applications in public administration, legal tech, and regulatory compliance automation.