GraphRAG aplicado a uma inteligência artificial conversacional para redução de alucinações

Abstract

Large Language Models (LLMs) have driven the development of conversational artificial intelligence systems capable of answering queries in natural language. However, these models still present limitations related to the generation of incorrect or unsupported information from the sources used, a phenomenon known as hallucination. Among the approaches proposed to mitigate this problem, Retrieval-Augmented Generation (RAG) stands out by incorporating external information during the response generation process. Nevertheless, traditional RAG presents limitations regarding the explicit representation of semantic relationships between entities, which may affect queries that require reasoning over interconnected information. In this context, this study aimed to develop and evaluate a conversational agent based on GraphRAG, comparing its performance with an agent based on traditional RAG regarding adherence to retrieved information and the occurrence of unsupported responses. The research is characterized as applied, exploratory in nature, and developed through an exploratory comparative study between two knowledge retrieval architectures. Two agents were developed using the Google Agent Development Kit (ADK), one based on RAG and another based on GraphRAG integrated with a knowledge graph stored in Neo4j. The evaluation was conducted through standardized questions, analyzing the fidelity of responses to the document content and the ability to recognize missing information. The results indicated that the GraphRAG-based agent provided greater contextual control and traceability.