Advancing faithful KBQA services via prompt-driven knowledge graph-enhanced LLMs in cloud

Overview

Knowledge-based question answering systems that leverage large language models face significant constraints related to factual accuracy and hallucination. This work addresses limitations in current knowledge graph integration approaches by proposing a prompt-driven reasoning framework that treats knowledge graphs as structured relational entities rather than static repositories. The proposed system targets improvements in both simple and multi-hop reasoning tasks within cloud-based intelligent services.

Methods and approach

The PDR framework operates through two integrated phases. The Subgraph Retrieval phase employs a refined PageRank algorithm to align input queries with knowledge graph structure while incorporating document retrieval mechanisms to extend graph boundaries, generating subgraphs optimized for answer coverage and relevance. The Reasoning phase embeds retrieved subgraphs as task-specific reasoning cues within prompts, directing the LLM to generate chain-of-thought reasoning and candidate knowledge graph paths. A stepwise filtering mechanism subsequently evaluates semantic coherence and structural alignment with the knowledge graph to identify pertinent answers, enhancing interpretability through explicit reasoning trace preservation.

Key Findings

Experimental evaluation demonstrates that PDR achieves performance improvements over existing baselines across both simple and multi-hop reasoning tasks. The framework generates more accurate answers while maintaining improved interpretability through explicit reasoning pathways. The integration of refined subgraph retrieval with structured prompt engineering produces superior results compared to approaches that treat knowledge graphs as fixed repositories or rely on LLM capabilities without external knowledge integration.

Implications

The framework addresses critical reliability concerns in knowledge-based question answering by reducing hallucination through grounded reasoning and explicit knowledge graph alignment. The separation of subgraph retrieval from reasoning enables more efficient resource utilization in cloud environments while maintaining semantic and structural fidelity. This approach establishes a methodology for treating knowledge graphs as dynamic reasoning resources rather than static information stores, with potential applications across enterprise knowledge retrieval systems.