A RAG-Driven Conversational System for Citizen Information Requests at Ecuador's Tax Authority

This study presents the design and implementation of a Retrieval-Augmented Generation based intelligent chatbot for tax consultation services within the Ecuadorian Internal Revenue Service. The system integrates modern natural language processing technologies including BGE-M3 multilingual embeddings, Groq's Language Processing Unit infrastructure, and LangGraph workflow orchestration to provide accurate, contextually relevant responses to taxpayer queries. The implementation processes 22 official documents, generating 1,847 semantic chunks stored in a Qdrant vector database. The system architecture incorporates adaptive query reformulation, real-time response streaming, and hexagonal design patterns to ensure scalable and maintainable tax consultation services. Comprehensive performance evaluation across 58 representative tax queries demonstrates 91.4% overall success rate with average response times of 1.039 seconds. User acceptance evaluation with 20 participants (10 professional accountants, 10 general taxpayers) reveals differential satisfaction patterns, with general taxpayers reporting higher acceptance scores (9.0/10) compared to accounting professionals (7.0/10). These results validate the practical viability of RAG-based systems for specialized regulatory consultation in governmental contexts while identifying specific improvement opportunities for expert user populations.