Iterative GraphRAG Architecture Improves LLM Response Quality for Complex Queries
A recent research study introduces KGiRAG, an iterative GraphRAG architecture that leverages feedback to enhance the responses of large language models (LLMs) to intricate sensemaking questions. This method involves assessing the quality of responses to progressively refine the outputs, ultimately producing a robust and well-supported answer. When tested on the HotPotQA dataset, this iterative approach demonstrated superior semantic quality and relevance over a single-shot baseline. The study also tackles the limitations of LLMs, such as issues with hallucination and restrictions related to context size.
Key facts
- KGiRAG is an iterative GraphRAG approach for responding to sensemaking queries.
- It leverages response quality assessment to iteratively refine outputs.
- The approach addresses LLM hallucination and context size limitations.
- Evaluated using queries from the HotPotQA dataset.
- Iterative strategy yields higher semantic quality and improved relevance over single-shot baseline.
- Paper published on arXiv under Computer Science > Information Retrieval.
- arXiv ID: 2604.20859.
- Submitted in April 2025.
Entities
Institutions
- arXiv