KG-RAG: Knowledge-Graph-Augmented Retrieval

Standard retrieval-augmented generation retrieves flat text chunks by embedding similarity alone, which gives large language models little structured relational context. This can limit factual accuracy on multi-hop questions that require facts from several connected entities.

The work was conducted at the Natural Language Learning Lab, Jeonbuk National University, from February 2025 to February 2026 under Prof. Hyun-je Song. The research connected structured graph context with neural embedding spaces while examining retrieval quality and computational efficiency.

I designed and implemented the knowledge-graph retrieval pipeline and multi-hop reasoning mechanism as an undergraduate researcher. The wider research direction and resulting manuscript are lab work under Prof. Hyun-je Song.

• Structured plus semantic context: graph retrieval preserved relations that embedding similarity alone can miss.
• Multi-hop traversal: connected evidence was assembled for questions requiring more than one factual link.
• Ablation studies: embedding integration techniques were isolated to examine accuracy and computational-efficiency trade-offs.

The research implementation used PyTorch and LangChain to integrate graph-derived context with embedding-based retrieval and pass connected evidence into the generation stage.

Systematic ablation studies compared embedding integration approaches with attention to accuracy and computational-efficiency trade-offs.

The implemented pipeline supports multi-hop retrieval over a knowledge graph for grounded question-answering research.

Relational coverage and embedding quality need separate evaluation because either can become the retrieval bottleneck. Ablating the integration stages made those trade-offs easier to isolate.