Agentic Graph RAG: Hybrid Explicit-Implicit Retrieval and Reasoning Over the MITRE ATLAS Knowledge Graph

Abstract

Combining large language models (LLMs) with structured knowledge graphs can be a potential way of addressing issues with reliability and explainability of question answering systems in particular domains. While LLMs have been shown to be effective at natural language understanding, they may still hallucinate or fail to perform well on structured tasks because of a lack of grounding. In contrast, although knowledge graphs explicitly represent entities and their relationships, accessing them through query languages such as Cypher requires specialized expertise. In this paper, we propose an Agentic Graph RetrievalAugmented Generation (RAG) framework that combines implicit and explicit retrieval within an agentic pipeline. Specifically, our framework includes (i) implicit retrieval, which relies on semantic similarity over vector embeddings to identify relevant information contextually, and (ii) explicit retrieval, involving natural language-to-Cypher query generation and its execution to obtain precise results from the knowledge graph. The explicit retrieval pipeline incorporates validation and correction loops to ensure schema-compliant Cypher query generation, and both retrieval strategies are integrated into a unified agentic workflow. Our framework operates on a semi-structured knowledge graph constructed from MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) YAML data and deployed in a Neo4j database. The proposed framework is evaluated using an LLM-as-a-Judge setup based on the correctness metric. Experimental results demonstrate that implicit retrieval achieves correctness scores between 0.51-0.58, while explicit retrieval improves performance to $0.73-0.91$. The proposed hybrid agentic framework achieves the best performance, reaching correctness scores of up to 0.98, demonstrating that integrating implicit semantic retrieval with explicit Cypher-based knowledge graph querying significantly improves answer correctness.