Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Single Pass, Looping Through Time: KG-IRAG with Iterative Knowledge Retrieval

Ruiyi Yang, Hao Xue, Imran Razzak, Hakim Hacid, Flora D. Salim

TL;DR

KG-IRAG addresses the challenge of time-sensitive, multi-step knowledge reasoning by integrating knowledge graphs with an iterative retrieval loop guided by two cooperating LLM agents (planner and verifier). The framework anchors retrieval on explicit time points, expands evidence through lightweight, localized graph walks, and uses a verifiable stop rule to ensure temporal consistency before synthesizing an answer. Three new temporal QA datasets—weatherQA-Irish, weatherQA-Sydney, and trafficQA-TFNSW—demonstrate that iterative, time-anchored retrieval improves accuracy and reduces hallucinations compared to single-pass GraphRAG baselines and other RAG variants. The approach delivers both performance gains and efficiency, with demonstrated generalization to external temporal QA benchmarks like TimeQuestions, highlighting its practical impact for time-aware KGQA tasks.

Abstract

Graph Retrieval-Augmented Generation (GraphRAG) has proven highly effective in enhancing the performance of Large Language Models (LLMs) on tasks that require external knowledge. By leveraging Knowledge Graphs (KGs), GraphRAG improves information retrieval for complex reasoning tasks, providing more precise and comprehensive retrieval and generating more accurate responses to QAs. However, most RAG methods fall short in addressing multi-step reasoning, particularly when both information extraction and inference are necessary. To address this limitation, this paper presents Knowledge Graph-Based Iterative Retrieval-Augmented Generation (KG-IRAG), a novel framework that integrates KGs with iterative reasoning to improve LLMs' ability to handle queries involving temporal and logical dependencies. Through iterative retrieval steps, KG-IRAG incrementally gathers relevant data from external KGs, enabling step-by-step reasoning. The proposed approach is particularly suited for scenarios where reasoning is required alongside dynamic temporal data extraction, such as determining optimal travel times based on weather conditions or traffic patterns. Experimental results show that KG-IRAG improves accuracy in complex reasoning tasks by effectively integrating external knowledge with iterative, logic-based retrieval. Additionally, three new datasets: weatherQA-Irish, weatherQA-Sydney, and trafficQA-TFNSW, are formed to evaluate KG-IRAG's performance, demonstrating its potential beyond traditional RAG applications.

Beyond Single Pass, Looping Through Time: KG-IRAG with Iterative Knowledge Retrieval

TL;DR

KG-IRAG addresses the challenge of time-sensitive, multi-step knowledge reasoning by integrating knowledge graphs with an iterative retrieval loop guided by two cooperating LLM agents (planner and verifier). The framework anchors retrieval on explicit time points, expands evidence through lightweight, localized graph walks, and uses a verifiable stop rule to ensure temporal consistency before synthesizing an answer. Three new temporal QA datasets—weatherQA-Irish, weatherQA-Sydney, and trafficQA-TFNSW—demonstrate that iterative, time-anchored retrieval improves accuracy and reduces hallucinations compared to single-pass GraphRAG baselines and other RAG variants. The approach delivers both performance gains and efficiency, with demonstrated generalization to external temporal QA benchmarks like TimeQuestions, highlighting its practical impact for time-aware KGQA tasks.

Abstract

Graph Retrieval-Augmented Generation (GraphRAG) has proven highly effective in enhancing the performance of Large Language Models (LLMs) on tasks that require external knowledge. By leveraging Knowledge Graphs (KGs), GraphRAG improves information retrieval for complex reasoning tasks, providing more precise and comprehensive retrieval and generating more accurate responses to QAs. However, most RAG methods fall short in addressing multi-step reasoning, particularly when both information extraction and inference are necessary. To address this limitation, this paper presents Knowledge Graph-Based Iterative Retrieval-Augmented Generation (KG-IRAG), a novel framework that integrates KGs with iterative reasoning to improve LLMs' ability to handle queries involving temporal and logical dependencies. Through iterative retrieval steps, KG-IRAG incrementally gathers relevant data from external KGs, enabling step-by-step reasoning. The proposed approach is particularly suited for scenarios where reasoning is required alongside dynamic temporal data extraction, such as determining optimal travel times based on weather conditions or traffic patterns. Experimental results show that KG-IRAG improves accuracy in complex reasoning tasks by effectively integrating external knowledge with iterative, logic-based retrieval. Additionally, three new datasets: weatherQA-Irish, weatherQA-Sydney, and trafficQA-TFNSW, are formed to evaluate KG-IRAG's performance, demonstrating its potential beyond traditional RAG applications.

Paper Structure

This paper contains 55 sections, 2 theorems, 18 equations, 1 figure, 11 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Combination of Graphs with LLMs
  4. Retrieval-Augmented Generation (RAG)
  5. LLMs Temporal Reasoning
  6. Beyond Single Pass, Looping Through Time: KG-IRAG Framework
  7. Setup and Iterative Overview
  8. Temporal KG Construction
  9. Triplet Retrieval
  10. Lightweight implementation.
  11. Iterative Reasoning
  12. Answer Generation
  13. Robustness to noisy evidence.
  14. Why Not Single-Pass Retrieval?
  15. Case Study
  16. ...and 40 more sections

Key Result

Lemma 1

Any single-pass retriever that inspects a temporal radius $w$ (in the same unit as $d^\star$) can succeed only if $w \ge d^\star$. If a dataset contains instances with $d^\star > w$, the single-pass procedure necessarily misses the first feasible window unless it increases $w$.

Figures (1)

  • Figure 1: The KG-IRAG framework: LLM1 generates an initial retrieval plan and a reasoning prompt, guiding LLM2 through iterative retrievals or stopping to generate the final answer.

Theorems & Definitions (5)

  • Definition B.1: Nearest feasible offset
  • Lemma 1: Single-pass coverage condition
  • proof : Sketch
  • proposition 1: Context dilution
  • proof : Sketch